WO2012014430A1 - 移動体検出装置および移動体検出方法 - Google Patents
移動体検出装置および移動体検出方法 Download PDFInfo
- Publication number
- WO2012014430A1 WO2012014430A1 PCT/JP2011/004173 JP2011004173W WO2012014430A1 WO 2012014430 A1 WO2012014430 A1 WO 2012014430A1 JP 2011004173 W JP2011004173 W JP 2011004173W WO 2012014430 A1 WO2012014430 A1 WO 2012014430A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- distance
- moving
- movement
- stationary
- calculation unit
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
- G06V20/58—Recognition of moving objects or obstacles, e.g. vehicles or pedestrians; Recognition of traffic objects, e.g. traffic signs, traffic lights or roads
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/215—Motion-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30241—Trajectory
Definitions
- the present invention relates to a technique for detecting a moving body in an image based on motion information in a moving image composed of a plurality of images. Especially for objects that move while the shape of the contour changes like a person, how much each movement trajectory seems to be a movement trajectory calculated from the area of the stationary object instead of the area of the moving object.
- the present invention relates to a moving object detection apparatus that detects an area of a moving object by using an index representing the distance between moving trajectories.
- a region extraction technique for detecting a moving body by extracting a moving body region in an image from an image including a moving body image (hereinafter simply referred to as a “moving body”) has been widely performed. Yes.
- the technology to extract the person's area is the focus control in the digital video camera and the digital still camera, the image quality improvement processing, the safe driving support system of the car, or the collision avoidance with the person in the robot. It is a basic technology that is commonly used for control or alarm.
- the extracted moving body region candidates are evaluated for the degree of similarity with a mobile body model prepared in advance.
- a mobile body model prepared in advance Is extracted as a moving object region.
- a region of a moving body that moves while being deformed such as a walking person
- there is a method of using a moving body model that takes deformation into account when extracting a region of a moving body that moves while being deformed, such as a walking person, there is a method of using a moving body model that takes deformation into account.
- a silhouette image of a moving object is extracted from a plurality of images as a moving object region candidate.
- a method of evaluating a similarity between a model related to deformation of a moving body parameterized in advance and an extracted silhouette image to estimate a region having a high similarity and a model parameter is disclosed. Accordingly, since a parameterized model can be applied to a person who moves while the shape changes periodically, it is possible to extract a region of a moving body.
- a typical method of the latter is to divide the image into a plurality of small areas, extract feature quantities based on the luminance values of the pixels in each small area, and then calculate the similarity of the feature quantities between the plurality of small areas.
- the conventional region extraction technique is used to correctly move a moving object when, for example, a scene in which a plurality of persons walk, the shape of the moving object changes significantly due to changes in posture, size, etc. There is a problem that it cannot be extracted.
- a mobile object is correctly detected by, for example, extracting one mobile object as a plurality of mobile objects by mistake, or extracting a region where a mobile object to be extracted does not exist as a mobile object. There is a problem that it cannot be done.
- An object of the present invention is to provide a moving body detection apparatus capable of detecting a moving body.
- a mobile object detection device includes a mobile object region from a plurality of movement trajectories each corresponding to each region in a moving image.
- a mobile object region from a plurality of movement trajectories each corresponding to each region in a moving image.
- a stationary index calculation unit that calculates a stationary index that represents the stationary object likeness of the moving track
- a distance calculating unit that calculates a distance that represents a similarity between the moving tracks, a stationary index of the moving track, and the movement track Based on the distance, the conversion is such that the ratio of the distance between the movement trajectory of any stationary object and the movement trajectory of any mobile object to the distance between the movement trajectories of any stationary object is larger than before the conversion.
- a region detection unit that detects a moving body region corresponding to the moving track of the moving body by separating the moving track of the stationary object and the moving track of the moving body based on the distance between the moving tracks. .
- the above conversion process is performed based on the stationary index of the movement trajectory. For this reason, it becomes easy to separate the moving locus of the stationary object and the moving locus of the moving body. Therefore, a moving body can be correctly detected even in an image that includes a moving body such as a person that moves while changing its shape and is captured by a moving camera.
- the present invention can be realized not only as a mobile object detection apparatus including such a characteristic processing unit, but also as a mobile object detection method using a characteristic processing unit included in the mobile object detection device as a step. Can be realized. It can also be realized as a program that causes a computer to execute the characteristic steps included in the moving object detection method. Needless to say, such a program can be distributed through a computer-readable recording medium such as a CD-ROM (Compact Disc-Read Only Memory) or a communication network such as the Internet.
- a computer-readable recording medium such as a CD-ROM (Compact Disc-Read Only Memory) or a communication network such as the Internet.
- FIG. 1 is a diagram showing a configuration of a moving object detection apparatus according to an embodiment of the present invention.
- FIG. 2 is a diagram illustrating a hardware configuration of a moving object detection apparatus configured by a computer.
- FIG. 3 is a flowchart showing an operation procedure of the moving object detection apparatus.
- FIG. 4 is a diagram illustrating an example of a shooting situation.
- FIG. 5A is a diagram illustrating an example of a picture of the first frame constituting the input moving image.
- FIG. 5B is a diagram illustrating an example of a picture of a frame between the first frame and the T frame constituting the input moving image.
- FIG. 5C is a diagram illustrating an example of a T-th frame picture constituting the input moving image.
- FIG. 5A is a diagram illustrating an example of a picture of the first frame constituting the input moving image.
- FIG. 5B is a diagram illustrating an example of a picture of a frame between the first frame and the T frame constituting the input
- FIG. 6A is a diagram illustrating an example of a motion vector.
- FIG. 6B is a diagram illustrating an example of a movement trajectory.
- FIG. 7 is a diagram illustrating a configuration of a stationary index calculation unit in the mobile object detection device.
- FIG. 8 is a diagram showing the classification of camera geometric constraints.
- FIG. 9A is a diagram illustrating epipolar constraint.
- FIG. 9B is a diagram for explaining the homography constraint.
- FIG. 10 is a diagram for explaining the structure matching constraint.
- FIG. 11A is a diagram illustrating an example of the movement trajectory in a picture and the size of a corresponding stationary index.
- FIG. 11B is a diagram illustrating a relationship between the size of the stationary index and the thickness of the arrow in FIG. 11A.
- FIG. 12 is a diagram illustrating a configuration of a distance calculation unit in the mobile object detection device.
- FIG. 13A is a diagram illustrating an example of a plurality of movement trajectories on a picture.
- FIG. 13B is a diagram illustrating the data distribution between the movement trajectories and the Euclidean distance.
- FIG. 13C is a diagram illustrating the data distribution between the movement trajectories and the geodetic distance.
- FIG. 14 is a diagram illustrating the Euclidean distance and the geodetic distance between the movement trajectories.
- FIG. 15 is a diagram illustrating a configuration of a subclass classification unit in the mobile object detection device.
- FIG. 13A is a diagram illustrating an example of a plurality of movement trajectories on a picture.
- FIG. 13B is a diagram illustrating the data distribution between the movement trajectories and the Euclidean distance.
- FIG. 13C is a diagram illustrating the data distribution
- FIG. 16A is a diagram illustrating an example of a plurality of movement trajectories on a picture.
- FIG. 16B is a conceptual diagram of a high-dimensional space made up of movement trajectories.
- FIG. 17 is a diagram illustrating a configuration of the weighted distance calculation unit.
- FIG. 18 is a conceptual diagram showing geodesic distance between subclasses.
- FIG. 19A is a conceptual diagram showing an example of a moving locus belonging to each of a moving object and a background superimposed on a scene image.
- FIG. 19B is a diagram showing, in a high-dimensional space, the result of class classification performed by the subclass classification unit on the movement trajectory existing in the scene shown in FIG. 19A.
- FIG. 19C is a diagram illustrating a result of the stationary index adding unit determining whether the stationary trajectory is stationary or moving with respect to each movement trajectory.
- FIG. 19D is a diagram illustrating a result of evaluation of stationary or moving in units of subclasses.
- FIG. 19E is a diagram illustrating a weighting rule.
- FIG. 19F is a diagram showing the weighted geodetic distance calculated by the weighting rule on a high-dimensional space.
- FIG. 20A is a diagram illustrating a movement trajectory and a subclass in a high-dimensional space.
- FIG. 20B is a diagram illustrating a classification result of subclasses.
- FIG. 20C is a diagram illustrating a display example of the integration result of subclasses.
- FIG. 20D is a diagram illustrating a display example of the integration result of subclasses.
- FIG. 21 is a diagram illustrating a configuration of the moving object detection device according to the first modification of the embodiment.
- FIG. 22 is a diagram illustrating a configuration of a stationary index calculation unit in the mobile object detection device.
- FIG. 23 is a diagram illustrating a detailed configuration of the weighted distance calculation unit.
- FIG. 24 is a diagram illustrating an example of an image created by the stationary index image creating unit of the stationary index calculating unit.
- FIG. 25A is a diagram illustrating a display example of a region division result.
- FIG. 25B is a diagram illustrating a display example of a region division result.
- FIG. 26A is a diagram illustrating a configuration of the moving object detection device according to the second modification of the exemplary embodiment.
- FIG. 26B is a diagram illustrating a configuration of a subclass classification unit.
- FIG. 27A is a diagram illustrating a configuration of a moving object detection device according to Modification 3 of the embodiment.
- FIG. 27B is a diagram illustrating a configuration of a subclass classification unit.
- FIG. 28A is a diagram illustrating a data distribution of a movement trajectory before dimension compression in a multidimensional space.
- FIG. 28B is a diagram illustrating a space obtained by dimensionally compressing the multidimensional space of the movement locus illustrated in FIG. 28A.
- FIG. 28C is a diagram illustrating a result of applying the clustering process to the dimensionally compressed data.
- FIG. 28D is a diagram illustrating an example in which clustering processing is performed on data of a moving person in a nonlinear space.
- FIG. 29A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 4 of the embodiment.
- FIG. 29B is a diagram illustrating a configuration of a subclass classification unit.
- FIG. 30A is a diagram illustrating an example of a movement trajectory.
- FIG. 30B is a diagram illustrating a multidimensional space including a movement locus.
- FIG. 30C is a diagram illustrating a clustering result when the threshold is larger than the maximum value of the Euclidean distance.
- FIG. 30D is a diagram showing the clustering result with respect to a certain threshold R 1.
- FIG. 30E is a diagram showing a clustering result for a threshold value R 2 that is smaller than the threshold value R 1 .
- FIG. 31A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 5 of the embodiment.
- FIG. 31B is a diagram illustrating a configuration of a subclass classification unit.
- FIG. 31C is a diagram illustrating another configuration of the subclass classification unit.
- FIG. 31D is a diagram showing still another configuration of the subclass classification unit.
- FIG. 32A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 6 of the embodiment.
- FIG. 32B is a diagram illustrating a configuration of a stationary index calculation unit.
- FIG. 33A is a diagram illustrating a configuration of the moving object detection device according to the modification example 7 of the embodiment.
- FIG. 33B is a diagram illustrating a configuration of a distance calculation unit.
- FIG. 33C is a diagram illustrating a configuration of a weighted distance calculation unit.
- FIG. 33D is a diagram illustrating a configuration of a weighted distance calculation unit.
- FIG. 34 is a conceptual diagram showing the Euclidean distance between subclasses.
- FIG. 35A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 8 of the embodiment.
- FIG. 35B is a diagram illustrating a configuration of a stationary index calculation unit and a weighted distance calculation unit.
- FIG. 36A is a conceptual diagram showing the distribution of subclasses in a high-dimensional space when the camera motion is large.
- FIG. 36B is a conceptual diagram illustrating the distribution of subclasses in a high-dimensional space when the camera motion is small.
- FIG. 36C is a diagram showing a weighting rule for the subclass observation value distance.
- FIG. 36D is a conceptual diagram illustrating a distribution of subclasses in a high-dimensional space after weighting the measured distance between subclasses when the camera motion is large.
- FIG. 36A is a conceptual diagram showing the distribution of subclasses in a high-dimensional space when the camera motion is large.
- FIG. 36B is a conceptual diagram illustrating the distribution of subclasses in a high-dimensional space when the camera motion is small.
- FIG. 36C is a diagram showing a weighting rule for the subclass
- FIG. 36E is a conceptual diagram showing the distribution of subclasses in the high-dimensional space after weighting the measured distance between subclasses when the camera motion is small.
- FIG. 37 is a diagram illustrating the configuration of the moving object detection device according to the ninth modification of the embodiment.
- FIG. 38 is a diagram illustrating another configuration of the moving object detection device according to the ninth modification of the embodiment.
- FIG. 39A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 10 of the embodiment.
- FIG. 39B is a diagram illustrating a configuration of a weighted distance calculation unit.
- FIG. 40 is a diagram illustrating a configuration of a moving object detection device including components essential to the present invention.
- One embodiment of the present invention is a moving object detection device that detects a moving object region from a plurality of movement loci each corresponding to each region in a moving image, and between two or more pictures constituting the moving image. For each of a plurality of movement trajectories that are movement trajectories of a block made up of one or more pixels constituting the picture in FIG.
- a stationary index calculation unit that calculates a stationary index representing the stationary object likeness of the movement trajectory, and the movement Based on the distance calculation unit that calculates the distance between the trajectories and the stationary index of the moving trajectory and the distance between the moving trajectories, the movement of the arbitrary stationary object with respect to the distance between the moving trajectories of any stationary object Conversion processing is performed such that the ratio of the distance between the trajectory and the moving trajectory of an arbitrary moving body is larger than that before the conversion, and based on the distance between the moving trajectories, the moving trajectory of the stationary object and the moving body Movement trajectory By separating the, and a region detecting section that detects a moving object region corresponding to the moving locus of the moving object.
- the area detection unit is configured such that a ratio of a geodetic distance between a moving trajectory of an arbitrary stationary object and a moving trajectory of an arbitrary moving object to a geodetic distance between the moving trajectories of an arbitrary stationary object is larger than that before the conversion.
- the moving object region corresponding to the moving locus of the moving object is detected by separating the moving locus of the stationary object and the moving locus of the moving object based on the geodetic distance between the moving locus. Therefore, the geodetic distance is a distance of a route reaching from one of the two movement loci to the other using a movement locus other than the two movement loci as a relay point.
- the above conversion process is performed based on the stationary index of the movement trajectory. For this reason, it becomes easy to separate the moving locus of the stationary object and the moving locus of the moving body. Therefore, a moving body can be correctly detected even in an image that includes a moving body such as a person that moves while changing its shape and is captured by a moving camera.
- the stationary index calculation unit estimates a geometric constraint that is established when each moving track is a moving track of a stationary object from the plurality of moving tracks, and uses the degree of satisfaction of the estimated geometric constraint as the stationary index. calculate.
- the stationary index calculation unit estimates one of epipolar constraint, homography constraint, triple linear constraint, and structural coincidence constraint from a plurality of movement trajectories, and determines the degree of satisfying the estimated geometric constraint. Calculated as the stationary index.
- the area detection unit calculates a geodetic distance between the movement trajectories based on the distance between the movement trajectories, and before weighting the calculated geodetic distances based on a stationary index of the movement trajectory. In comparison with, the weighting is performed so that the ratio of the geodetic distance between the moving trajectory of any stationary object and the moving trajectory of any moving object to the geodetic distance between the moving trajectories of any stationary object is increased.
- the moving trajectories in which the weighted geodetic distance is equal to or greater than a predetermined threshold are set to different classes. By classifying, a region of a stationary object and a region of a moving object on each picture may be included.
- an evaluation value indicating the stationary object likeness of the moving locus is By using the weighting to the geodetic distance between the movement trajectories, it is possible to reliably detect the moving body in the image.
- the distance calculation unit calculates a geodetic distance representing a similarity between moving trajectories
- the weighted distance calculation unit calculates the geodetic distance calculated by the distance calculation unit based on the stationary index. Therefore, the weight is set so that the ratio of the geodetic distance between the moving trajectory of any stationary object and the moving trajectory of any moving object to the geodetic distance between the moving trajectories of any stationary object is larger than before weighting. Thus, the weighted geodesic distance is calculated.
- the weighted geodetic distance is calculated by applying a weight based on the stationary index to the geodetic distance between the movement trajectories.
- the distance calculation unit calculates a linear distance representing the similarity between the movement trajectories
- the weighted distance calculation unit calculates the linear distance calculated by the distance calculation unit based on the stationary index.
- the weighting should be such that the ratio of the linear distance between the movement trajectory of any stationary object and the movement trajectory of any moving object to the linear distance between the movement trajectories of any stationary object is larger than before weighting.
- a weighted linear distance may be calculated, and the weighted geodetic distance may be calculated from the calculated weighted linear distance.
- the area detection unit is configured to classify the weighted distance calculation unit that calculates the geodetic distance between the movement trajectories based on the distance between the movement trajectories, and the moving area of the stationary object and the moving area of the moving object in different classes.
- a weighted threshold obtained by applying a weight based on the stationary index to a threshold of the geodetic distance used for determining whether to classify, and the geodetic distance calculated by the weighted distance calculating unit
- a region dividing unit that divides the region of the stationary object and the region of the moving object on each picture may be included.
- the threshold is weighted instead of calculating the weighted distance by weighting the distance. For this reason, it is possible to reduce the calculation time necessary for calculating the weighted distance and the memory capacity for holding the weighted distance.
- the area dividing unit sets a weight for the threshold of the geodetic distance when the stationary indices of the two movement trajectories are values representing “stationary object” and “stationary object”, respectively. It is larger than the weight for the geodesic distance threshold when the stationary index is a value representing “moving object” and “stationary object”, respectively.
- the above-described moving body detection device further includes a subclass classification unit that classifies the plurality of movement trajectories into a plurality of subclasses that are subsets of similar movement trajectories
- the weighted distance calculation unit includes: A geodesic distance between the weighted subclasses based on the stationary index is calculated based on a stationary index of the movement trajectory, a distance between the moving trajectories, and a classification result of the subclass, and the region dividing unit includes the weighted Based on the geodesic distance between the subclasses calculated by the distance calculation unit, the stationary object subclass and the moving object subclass are separated, thereby dividing the stationary object region and the moving object region on each picture. Good.
- each subclass is classified as either a stationary object area or a moving object area.
- Each subclass is a subset of similar movement trajectories. For this reason, the area
- the distance calculation unit calculates a geodetic distance representing a similarity between the movement trajectories
- the weighted distance calculation unit calculates a movement trajectory included in the subclass with respect to a representative value of the geodetic distance between the subclasses.
- a geodetic distance between the subclasses is calculated by applying a weight based on the representative value of the stationary index.
- the distance calculation unit calculates a linear distance representing the similarity between the movement trajectories, and the weighted distance calculation unit determines whether the movement trajectory included in the subclass is stationary with respect to the representative value of the linear distance between the subclasses.
- a weight based on the representative value of the index may be applied, and the geodetic distance between the subclasses may be calculated based on the representative value of the linear distance between the weighted subclasses.
- the geodesic distance between the subclasses is calculated after applying a weight to the representative value of the linear distance between the subclasses. For this reason, the derivation of the geodetic distance between the subclasses reflecting the likelihood of a stationary object can be easily realized.
- the weighted distance calculation unit when the representative value of the stationary index of the movement trajectory included in the two subclasses is a value representing “moving object” and “stationary object”, respectively, between the two subclasses.
- the weight of the geodesic distance between the subclasses may be set to a value larger than 1.
- the weighted distance calculation unit when the representative value of the stationary index of the movement trajectory included in the two subclasses is a value representing “stationary object” and “stationary object”, respectively, between the two subclasses.
- the weight of the geodesic distance between the subclasses may be set to a value less than 1.
- the weighted distance calculation unit when the representative value of the stationary index of the movement trajectory included in the two subclasses is a value representing “moving object” and “moving object”, respectively, between the two subclasses.
- the weight of the geodetic distance between the subclasses may be set to 1.
- the subclass classification unit may classify each movement trajectory into one of the plurality of subclasses based on the similarity between the movement trajectories.
- the subclass classification unit may classify each movement trajectory into one of the plurality of subclasses based on the similarity of luminance between blocks belonging to each movement trajectory.
- the subclass classification unit performs dimension compression of a second distance calculation unit that calculates a geodetic distance between the movement trajectories and a geodetic distance between the movement trajectories calculated by the second distance calculation unit. And a clustering unit that classifies each movement trajectory into one of the plurality of subclasses based on a geodetic distance between the movement trajectories.
- clustering can be performed stably in a non-linear space without requiring enormous calculations, and in particular, it is possible to stably perform clustering even for pictures including persons who move while their shapes change.
- the subclass classification unit selects, for each of the plurality of movement trajectories, a distance that is equal to or less than the predetermined distance threshold among a plurality of distances from the movement trajectory to another movement trajectory, and a distance that is not selected.
- the geodetic distance is calculated by obtaining the shortest path from the moving trajectory to another moving trajectory, and the geodesic distance between the moving trajectories becomes a finite value.
- the movement trajectory may be classified into one of the plurality of subclasses by classifying the group of the same into the same subclass.
- the moving trajectories can be classified into subclasses based on the discontinuous points.
- the stationary index calculation unit estimates the geometric constraint from a movement trajectory for estimating a geometric constraint established with respect to the movement trajectory of the stationary object, and based on the estimated geometric constraint, the distance calculation unit You may calculate the stationary parameter
- the geometric constraint is estimated from the movement trajectory for estimating the geometric constraint. For this reason, since a more stable and accurate geometric constraint can be obtained, a moving body can be detected more stably and correctly.
- the above-described moving body detection device further includes a camera motion acquisition unit that acquires motion information of a camera that captures the moving image, and the weighted distance calculation unit is based on the magnitude of the motion of the camera.
- the weight of the stationary index at the time of weighting is changed.
- the weighted distance calculation unit increases the weighting of the distance between the moving trajectory of the stationary object and the moving trajectory of the moving object as the movement of the camera increases.
- the above-described moving body detection device further includes a camera motion acquisition unit that acquires motion information of a camera that captures the moving image
- the weighted distance calculation unit is a representative value of the stationary indices of the two subclasses. Is a value representing “moving object” and “stationary object”, respectively, and weights the geodetic distance between the two subclasses, and the weight when the movement of the camera is equal to or greater than a predetermined threshold is WB, If the weight when the camera motion is smaller than the predetermined threshold is WS, the relationship of WB> WS> 1 may be satisfied.
- the above-described moving body detection device further includes a camera motion acquisition unit that acquires motion information of a camera that captures the moving image, and the weighted distance calculation unit includes representative values of the stationary indices of the two subclasses. Is a value representing “stationary object” and “stationary object”, respectively, and weights the geodetic distance between the two subclasses, and the weight when the motion of the camera is equal to or greater than a predetermined threshold is WB, When the weight when the camera motion is smaller than the predetermined threshold is WS, the relationship of WB ⁇ WS ⁇ 1 may be satisfied.
- the camera motion acquisition unit acquires motion information of the camera from an operation control signal for the camera.
- the camera motion acquisition unit may acquire motion information of the camera from a vehicle-mounted sensor.
- the weighted distance calculation unit further compares the stationary index of the moving trajectory with a stationary index threshold to determine the moving trajectory having a stationary index equal to or less than the stationary index threshold as a stationary object moving trajectory. Then, the movement trajectory having a stationary index larger than the stationary index threshold value may be determined as the movement trajectory of the moving body.
- the weighted distance calculation unit includes a threshold value input unit that receives a stationary index threshold value, and compares the stationary index value of the movement trajectory with the stationary index threshold value received by the threshold value input unit, thereby obtaining the stationary index threshold value.
- the movement trajectory having the following stationary index is determined as a movement trajectory of a stationary object
- the movement trajectory having a stationary index larger than the stationary index threshold is determined as a movement trajectory of a moving object
- the region detection unit further includes The detected moving body region may be displayed on the display unit.
- the stillness index calculation unit may further display each block constituting the picture on the display unit in a display mode according to the calculated value of the stillness index.
- the stationary index threshold depends largely on the distribution of the stationary index value. Therefore, if the value of the stationary index and its distribution on the image can be confirmed, the stationary index threshold can be adjusted without performing the region division process to the end. Thereby, the optimal stationary index threshold value can be determined earlier.
- the moving body detection device of the present invention is realized not only by configuring each processing unit by hardware, but also as a moving body detection method having the process executed by each processing unit as a step,
- An area of an object having a motion in a moving image realized as a program for causing a computer to execute the steps included in the moving object detection method, or realized as a computer-readable recording medium such as a CD-ROM storing the program
- FIG. 1 is a diagram illustrating a configuration of a moving object detection apparatus according to the present embodiment.
- the moving body detection apparatus 100 includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 105, a weighted distance calculation unit 106, and An area dividing unit 107 is provided.
- the moving object detection apparatus 100 is an apparatus that detects a moving object in a moving image by performing region division that specifies all or a part of the moving object region in the moving image.
- the moving body detection apparatus 100 is an apparatus that detects a region corresponding to a moving body from a plurality of movement trajectories respectively corresponding to a plurality of regions in a moving image.
- the moving body detection apparatus 100 receives a moving image captured by the camera 110 as an input, detects a moving body region in the moving image, generates a picture based on the detection result, and outputs the picture.
- the display 120 displays a picture output from the moving object detection apparatus 100.
- a picture is also referred to as an image.
- the image input unit 101 is a processing unit that receives input of a plurality of pictures constituting a moving image, and is, for example, a camera or a communication interface connected to the camera.
- the movement trajectory calculation unit 102 is a processing unit that obtains a plurality of corresponding points between a plurality of pictures based on a plurality of pictures received by the image input unit 101, and outputs the corresponding points as a plurality of movement trajectories. That is, the movement trajectory calculation unit 102 performs, for each block composed of one or more pixels constituting the picture received by the image input unit 101, between the two temporally adjacent pictures constituting the moving image. A block movement is detected, and the detected movement is connected to the plurality of pictures to calculate a plurality of movement trajectories. Corresponding points between pictures may be obtained for each pixel of the picture, or one corresponding point may be obtained for a plurality of adjacent pixels (blocks) in the picture.
- a certain corresponding point is obtained for each pixel or one for a plurality of pixels.
- the corresponding point of another picture corresponding to the pixel i of a certain picture and the corresponding point of another picture corresponding to the block i of the certain picture are both referred to as a movement locus of the pixel i.
- the stationary index calculation unit 103 is a processing unit that applies a geometric constraint to each of the movement trajectories obtained by the movement trajectory calculation unit 102 and obtains a stationary index that represents the stationary object likeness of each movement trajectory. That is, the still index calculation unit 103 performs, for each of a plurality of movement trajectories that are trajectories of a block composed of one or more pixels constituting a picture between two or more pictures constituting a moving image, the movement trajectory. A stationary index representing the likelihood of a stationary object is calculated. The geometric constraint used for the stationary index will be described later in detail.
- the “distance” between two movement trajectories in this specification is not only the distance between two points in the corresponding two-dimensional image space, but also the arithmetic distance between multidimensional data, as will be described later. Including. In general, distance and similarity have a contradictory relationship. That is, when the distance between two data is small, the similarity is high. Conversely, when the distance between two data is large, the similarity is low.
- Linear distance refers to a distance obtained only between two data
- geometric distance As will be described later, this indicates a distance obtained by passing through a point other than the two data.
- linear distance in this specification is a wider concept distance including the “linear distance” that is generally used, that is, geometrical positions such as position, velocity, acceleration, etc. in picture coordinates between moving trajectories. It is a distance that serves as an index representing the degree of similarity.
- Euclidean distance is used for the following description as a representative of “linear distance”. A detailed example of the distance will be described later in the description of the distance calculation unit 104.
- stationary and moving mean an object whose position in the world coordinate system does not change in time with respect to the world coordinate system fixed to the earth (ground) unless otherwise specified. Is assumed to be a stationary object and a temporally changing object is assumed to be a moving object. However, the criteria for “still” and “moving” are not limited to the above. For example, when photographing the inside of a moving vehicle with a camera installed in the vehicle, the world coordinate system fixed to the vehicle is used. As a reference, an object whose position does not change with time may be treated as a stationary object, and an object that changes with time as a moving object.
- the subclass classifying unit 105 divides a plurality of movement trajectories into each of the plurality of movement trajectories obtained by the movement trajectory calculation unit 102 based on the similarity between the trajectories or the speed, and at least one movement trajectory. Cluster into multiple subclasses containing That is, the subclass classification unit 105 moves a plurality of movement trajectories, each of which is a trajectory of a block made up of one or more pixels constituting a picture, across a plurality of pictures constituting a moving image. Classify into multiple subclasses that are sets of trajectories.
- the weighted distance calculation unit 106 is based on the stationary index associated with each movement trajectory that is the output of the stationary index calculation unit 103, the distance between the movement trajectories that is the output of the distance calculation unit 104, and the output of the subclass classification unit 105.
- a geodesic distance between subclasses weighted with a stationary index (hereinafter referred to as “intersubclass geodetic distance” as appropriate) is calculated based on the label information indicating which subclass each movement trajectory belongs to.
- the weighted distance calculation unit 106 calculates the geodetic distance between the movement trajectories based on the distance between the movement trajectories, and before weighting each calculated geodetic distance based on the stationary index of the movement trajectory. Compared to the geodetic distance between the movement trajectory of any stationary object, weighting is performed by weighting the ratio of the geodetic distance between the movement trajectory of any stationary object and the movement trajectory of any mobile object to be large. Calculate geodetic distance.
- the weighted distance calculation unit 106 first obtains a geodetic distance between each movement locus, and then obtains a representative geodetic distance between subclasses. In addition, the weighted distance calculation unit 106 obtains a representative value of the stationary index in each subclass. The weighted distance calculation unit 106 determines whether each subclass seems to be a moving object or a stationary object depending on whether the subclass stationary index exceeds or does not exceed a predetermined stationary index threshold, and uses the obtained result as a weight. The geodetic distance between subclasses is calculated by multiplying the representative geodetic distance between them.
- the weighted distance calculation unit 106 determines that a subclass having a stationary index equal to or lower than the stationary index threshold is a subclass of a stationary object by comparing the stationary index of the subclass with the stationary index threshold, and is larger than the stationary index threshold.
- a subclass having a stationary index is determined as a subclass of a moving object. The condition of the stationary index, the weighting standard, etc. will be described later in detail.
- the region dividing unit 107 performs region integration of a plurality of subclasses based on the geodesic distance between subclasses calculated by the weighted distance calculating unit 106, and the region division result finally obtained is identified, for example.
- Image processing is performed so that different display modes are provided for each region, and the data is output in a format that can be displayed on the display 120 or the like. That is, the area dividing unit 107 classifies the movement trajectories having weighted geodesic distances equal to or more than a predetermined threshold based on the weighted geodesic distance calculated by the weighted distance calculating unit 106 into different classes. Divide the area of the stationary object above and the area of the moving object.
- area extraction refers to a detection technique that extracts an image area in which a specific target object exists, and an area division technique that divides an area in a picture without distinguishing what the target object is. , Including both. Since the detection technique and the area division technique have many common parts, they are not distinguished in this specification.
- moving object detection is a movement that is relatively different from a detection technique that specifies only an image area where an object moving relative to a reference coordinate system is present. It includes both region segmentation techniques that segment regions within a picture for each object.
- each component image input unit 101, movement trajectory calculation unit 102, stationary index calculation unit 103, distance calculation unit 104, subclass classification unit 105, weighted distance calculation unit 106, area division constituting the mobile object detection device 100 is divided.
- the unit 107) may be realized by software such as a program executed on a computer, or may be realized by hardware such as an electronic circuit or an integrated circuit.
- FIG. 2 is a diagram showing a hardware configuration of the moving object detection device according to the present embodiment realized by software.
- the camera 110 captures and outputs a picture
- the computer 200 acquires the picture and performs a moving body extraction process to generate a picture that displays the region extraction result.
- the display 120 acquires and displays a picture generated by the computer 200.
- the computer 200 includes an I / F (interface) 201, a CPU (Central Processing Unit) 202, a ROM (Read Only Memory) 203, a RAM (Random Access Memory) 204, an HDD (Hard Disk Drive) 205, and a video card 206.
- a program for operating the computer 200 is stored in advance in the ROM 203 or the HDD 205.
- the program is read out from the ROM 203 or HDD 205 to the RAM 204 and expanded by the CPU 202 as a processor.
- the CPU 202 executes each coded instruction in the program expanded in the RAM 204.
- the I / F 201 captures a picture taken by the camera 110 into the RAM 204 in accordance with the execution of the program.
- the video card 206 outputs a picture generated according to the execution of the program, and the display 120 displays the picture.
- the computer program is not limited to the semiconductor ROM 203 or the HDD 205, and may be stored in, for example, a CD-ROM.
- the data may be transmitted via a wired or wireless network, broadcasting, or the like and taken into the RAM 204 of the computer.
- FIG. 3 is a flowchart showing the operation of the moving object detection apparatus 100 according to the present embodiment.
- steps S301 to S307 correspond to the respective processing units 101 to 107 in FIG. That is, the image input unit 101 is an image input step S301, the movement track calculation unit 102 is a movement track calculation step S302, the stationary index calculation unit 103 is a stationary index calculation step S303, the distance calculation unit 104 is a distance calculation step S304, and the subclass classification unit 105.
- the subclass classification step S305, the weighted distance calculation unit 106 executes the distance calculation step S306, and the region division unit 107 executes the region division step S307.
- the image input step S301 is executed by the image input unit 101. That is, the image input unit 101 acquires a plurality of pictures constituting a moving image from the camera 110.
- the moving image acquired from the camera 110 is a moving image of 30 frames / second.
- FIG. 4 is a diagram illustrating an example of a shooting situation.
- 5A shows the picture of the first frame
- FIG. 5C shows the picture of the T frame
- FIG. 5B shows the picture of the frame between the first frame and the T frame.
- the movement trajectory calculation step S302 is executed by the movement trajectory calculation unit 102. That is, the movement trajectory calculation unit 102 receives a plurality of pictures from the image input unit 101, detects pixel movement information (corresponding points) between pictures, and generates and outputs a movement trajectory. As a method for obtaining pixel motion information (corresponding points) between a plurality of pictures, here, all pixels (I) on one frame picture are used as a reference, and (T-1) frames from 2 frames to T frames. Find the corresponding pixel on the picture.
- the movement trajectory calculation unit 102 uses the two pictures of the t frame and the t + 1 frame, and the pixel coordinate value of the corresponding point as the motion vector of the pixel i. (X t i , y t i , x t + 1 i , y t + 1 i ) are estimated.
- the frames do not necessarily have to be continuous.
- the motion of the pixels may be obtained using two pictures of t frames and t + n frames.
- n is an integer of 1 or more.
- Non-Patent Document 1 calculates an optical flow based on hierarchical block matching. Since smoothness between pixels is used as a constraint, an optical flow is obtained in which a motion vector changes smoothly between adjacent optical flows. An efficient and accurate corresponding point is required particularly when there is no steep movement or shielding.
- the reliability of the estimation can be calculated, as will be described later, by removing the corresponding points whose reliability is lower than the threshold value from the subsequent processing, the ratio of the erroneous motion vectors to the total motion vectors can be reduced, and more accurately. There is an effect that the mobile object can be detected.
- Non-Patent Document 2 is a graph cut-based optical flow calculation method. Although the calculation cost is high, accurate corresponding points can be obtained densely on a picture. In addition, according to this method, since the occlusion area can also be estimated, as described later, by removing the corresponding points located in the occlusion area from the subsequent processing, the ratio of erroneous motion vectors to all motion vectors can be reduced. There is an effect that accurate moving body detection can be performed. Since further details are described in each document, detailed description thereof is omitted. P. Anandan, “A Computational Framework and an Algorithm for the Measurement of Visual Motion”, International Journal of Computer Vision, Vol. 2, pp. 283-310, 1989 Vladimir Kolmogorov and Ramin Zabih, “Computing Visual Correspondence with Occlusions via Graph Cuts”, International Conference on Computer Vision, 2001
- the movement trajectory calculation unit 102 may estimate an affine parameter instead of the motion vector as the pixel motion.
- motion information may be obtained for all pixels.
- the motion information may be obtained only for pixels on the grid at a fixed interval by dividing the picture into grids. You may also ask for motion information.
- Non-Patent Document 1 when calculating a motion vector using the technology disclosed in Non-Patent Document 1, only the pixels having motion information with high reliability may be used because the reliability can be calculated as described above. Further, when the motion vector is calculated using the technology disclosed in Non-Patent Document 2, the occlusion can be estimated as described above. For this reason, only the motion information of pixels that are not shielded may be used.
- a method of calculating a motion vector assuming affine deformation of a block may be used instead of the method of calculating a motion vector assuming the translational movement of the block.
- a method of calculating a motion vector on the assumption of affine deformation can be realized using the technology disclosed in Non-Patent Document 3. Jianbo Shi and Carlo Tomasi “Good Features to Track”, IEEE Conference on Computer Vision and Pattern Recognition, pp 593-600, 1994
- the affine parameter A t i corresponding to the motion in the vicinity of the pixel i of the picture of the t frame and the t + 1 frame is estimated.
- the pixel positions x t i and x t + 1 i on the picture in the t frame and the t + 1 frame have the relationship shown in (Formula 1).
- the movement trajectory calculation unit 102 calculates a movement trajectory i from the corresponding points of the pixels calculated between T pictures that are temporally different.
- the movement locus of the pixel i is referred to as a movement locus i.
- the movement trajectory calculation unit 102 uses the motion vector information 602 calculated in step S302 based on the pixel i 603a and the pixel k 603b of the input picture 601 of the t frame to calculate the motion of the pixel i 603a and the pixel k 603b.
- the corresponding points of the pixel i 603a and the pixel k 603b are obtained by tracking.
- the movement trajectory calculation unit 102 has a coordinate value (x 1 i , y 1 i ) of a pixel i on a picture of one frame and a pixel coordinate value (x t i , y t i ), the movement trajectory x i is calculated as in (Expression 2).
- the movement trajectory x i is a corresponding point between T pictures from 1 frame to T frame.
- FIG. 6B shows an example of the movement trajectory.
- the moving image input to the movement locus calculation unit 102 includes T pictures 604.
- the movement trajectories x i 606a and x k 606b are a collection of corresponding points on the picture from the 2nd frame to the Tth frame corresponding to the pixel i605a and the pixel k605b in one frame, respectively.
- the movement trajectories x i 606a and x k 606b are represented by vectors having the picture coordinate values of each picture as elements.
- the corresponding point between pictures is obtained by the movement trajectory calculation unit 102, the corresponding point is obtained for each of a plurality of adjacent pixels (blocks) in the picture, instead of obtaining the corresponding point for every pixel of the picture. Also good. In the present specification, it is not distinguished whether a certain corresponding point is obtained for each pixel or whether one corresponding point is obtained for a plurality of pixels. Also, the corresponding point of another picture corresponding to the pixel i of a certain picture and the corresponding point of another picture corresponding to the block i of the certain picture are both referred to as a movement locus of the pixel i.
- the stationary index calculation step S303 is executed by the stationary index calculation unit 103. That is, the stationary index calculation unit 103 calculates a stationary index for each of the plurality of movement trajectories calculated by the movement trajectory calculation unit 102.
- the stationary index of the movement trajectory is an index representing the likelihood that a certain movement trajectory is a movement trajectory on a stationary object.
- the stationary index calculation unit 103 estimates a geometric constraint to be satisfied by the movement locus on the stationary object, obtains an error indicating a degree that each movement locus satisfies the obtained geometric constraint, and uses this error as a stationary index. That is, the smaller the value of the stationary index of a certain movement locus (the smaller the error), the higher the likelihood that the movement locus on the stationary object is. Conversely, the greater the value of the stationary index of a certain movement trajectory (the greater the error), the lower the likelihood of being a trajectory on a stationary object.
- FIG. 7 shows the configuration of the stationary index calculation unit 103.
- the stationary index calculation unit 103 estimates the geometric constraint that the corresponding point on the stationary object should satisfy between the frames, and the degree of deviation of the moving trajectory from the stationary object based on the geometric constraint, that is, And an error calculation unit 702 that calculates a geometric constraint error indicating the degree of the moving object likeness (hereinafter referred to as “moving object likeness”) of the moving locus.
- “2 eyes” or “3 eyes” indicates a condition regarding the number of frames or the number of cameras. “Binocular” indicates a constraint established between two images. For example, even if a single camera is used, if there are a plurality of images taken at different times such as moving images, Good. Similarly, “three eyes” indicates a constraint that is established between three images. The plurality of images may be acquired from consecutive different frames, or when there are a plurality of cameras, one or a plurality of images may be acquired and used from each. In the present embodiment, it is assumed that an image acquired from continuous frames is used.
- Epipolar constraint is the most standard constraint condition among camera geometric constraints.
- the epipolar constraint works well for sparse corresponding points, but has an advantage that accurate corresponding points are preferable, and that the base length of the camera position between frames is preferably large.
- FIG. 9A is a diagram for explaining epipolar restraint.
- the stationary point X projected on the point p 1 i on the first frame image is on the point p 2 i existing on the straight line l 2 passing through the epipole e 2 called epipolar line on the second frame image. Projected. Therefore, when the corresponding point p 2 i deviates from the epipolar line l 2 by a distance equal to or greater than a certain threshold, it can be determined that the corresponding point is not a stationary object but a point on the moving body.
- the constant threshold value is generally set within a range of 0 to several pixels, although it depends on noise such as an error of corresponding points.
- the epipolar constraint is given by the basic matrix F 1,2 represented by a 3 ⁇ 3 matrix as follows.
- p 1 i is the pixel position of the pixel i on the first frame image shown in the format of (x 1 i , y 1 i , 1)
- p 2 i is the same as (x 2 i , y 2 i , 1) is the corresponding point pixel position on the second frame image corresponding to p 1 i
- l 1 i and l 2 i are epipolar lines.
- the basic matrix F (the basic matrix F 1,2 in (Expression 5)) is estimated by iterative calculation such as RANSAC (RANdom Sample Consensus) by the 8-point method, the 6-point method, or the like.
- RANSAC Random Sample Consensus
- the eight-point method is used in which the basic matrix F is calculated and estimated from the eight corresponding points, and when there are many points on the plane, a homography matrix is used. If the 6-point method for obtaining the epipole is used, the basic matrix F can be estimated more stably.
- FIG. 9B is a diagram for explaining the homography constraint.
- the homography constraint is a constraint condition that holds when, for example, the subject is assumed to be a plane or when the camera moves only in rotation.
- the homography constraint is suitable for close corresponding points because the camera motion may be small and it is also resistant to noise.
- the point p 1 i in the image of the first frame and the point p 2 i in the image of the second frame in which the point X ⁇ on a certain scene plane ⁇ is projected are expressed in a 3 ⁇ 3 matrix.
- the following conversion formula is established by the represented homography matrix H 1,2 .
- p 1 i is the pixel position of the point p 1 on the first frame image shown in the format of (x 1 i , y 1 i , 1)
- p 2 i is the same as (x 2 i , y
- the homography matrix H H 1,2 in (Expression 6)
- H can be estimated from an image using RANSAC or the like using four or more points on a plane.
- the stationary index calculation unit 103 determines that the corresponding point that is out of the homography conversion formula is a point on the moving object.
- the trilinear constraint is an epipolar constraint that is established between two eyes and is developed into a constraint condition between three eyes.
- the triple linear constraint is a straight line connecting the camera center in the first frame and the corresponding point on the first frame image when the corresponding point on the image of the three frames corresponding to the stationary point in a certain space is considered.
- a straight line connecting the camera center in the second frame and the corresponding point on the second frame image, and a straight line connecting the camera center in the third frame and the corresponding point on the third frame image are one point. It gives the constraint conditions for crossing.
- This constraint is expressed by a tensor called a trifocal tenor.
- Equation 7 The point p 1 i on the first frame image, the point p 2 i on the second frame image, and the point p 3 i on the third frame image.
- T j can be obtained by solving a linear equation using seven corresponding points and performing iterative calculation using LMedS or the like.
- Non-Patent Document 4 More detailed explanations about epipolar constraint, homography constraint, and triple linear constraint are described in Non-Patent Document 4, and the details are omitted.
- FIG. 10 is a diagram for explaining the structure consistency constraint.
- the structure coincidence constraint is a development of the homography constraint, and uses a relative depth from the homography plane ⁇ called a projective depth, as a constraint condition for a stationary object.
- a point obtained by directly projecting the three-dimensional point X reflected on the point x on the image of the first frame onto the image of the second frame is the point x ′, but if X is X ⁇ ′ Even a point on the plane ⁇ shown is projected to a point x ′ on the second frame image. At this time, on the first frame image, X ⁇ ′ appears at points x 1 to .
- x, x 1 to x ′, and epipoles e and e ′ are respectively the position C1 of the camera 110 in the first frame, the position C2 of the camera 110 and the point X in the second frame. It must exist on the plane formed by That is, by focusing on the ratio between the distance between the point x and the epipole e and the distance between the point x 1 and the epipole e, a relative depth expression (projection depth) based on the reference plane ⁇ can be realized.
- the projection depth value k 12 i is calculated by the following equation.
- p 1 i indicates the pixel position of a point on the image of the first frame shown in the format of (x 1 i , y 1 i , 1).
- e 1 indicates the pixel position of the epipole on the image of the first frame shown in the format of (u, v, 1).
- H 21 represents a 3 ⁇ 3 homography matrix for projecting a point on the image of the second frame to a point on the image of the first frame with respect to a point on the reference plane ⁇
- p 2 i is (x A pixel position of a point on the image of the second frame corresponding to p 1 i shown in the form of 2 i , y 2 i , 1) is shown. From these, the projection depth k 12 i can be obtained for all pixels having corresponding points.
- G 1 , 2, and 3 can be estimated by using 15 points and performing iterative calculation using LMedS or the like.
- Non-Patent Document 5 A more detailed description of the structural coincidence constraint is described in Non-Patent Document 5, and therefore a further detailed description is omitted.
- Richard Hartley and Andrew Zisserman “Multi-View Geometry in Computer Vision”, second ed. Cambridge Univ. Press, 2003 Chang Yuan, Gerard Medioni, Jinman Kang and Isaac Cohen, "Detecting Motion Regions in the Presence of a Strong Parallax from a Moving Camera by Multiview Geometric Constraints", IEEE Transactions On Pattern Analysis and Machine Intelligence, Vol. 29, no. 9, September 2007
- the geometric constraint estimation unit 701 estimates the geometric constraints described above. Which geometric constraint to use may be selected by the user depending on the scene and situation.
- the homography H is estimated from the image.
- the basic matrix F is estimated.
- the matrix G is estimated, and error calculation is performed. Output to the unit 702.
- the error calculation unit 702 calculates a stationary index E representing the stationary object likeness of the movement locus corresponding to each pixel based on the geometric constraint estimated by the geometric constraint estimation unit 701.
- Points that do not satisfy the geometric constraints are unlikely to be stationary objects, so based on the estimated geometric constraints, the degree to which the moving trajectory corresponding to each pixel deviates from the stationary object, that is, the degree of the moving object Is calculated as a stationary index E of the stationary object-likeness.
- the stationary index E uses an evaluation value that is ideally 0 for a stationary object.
- epipolar constraints are used as geometric constraints for calculating the stationary index E.
- the following (Equation 10) is used as the stationary index E (i) for a certain pixel i from the first frame to the Tth frame.
- any index that is 0 for a stationary object and takes a value other than 0 for a moving object may be used as the stationary index E for the likelihood of a stationary object. That is, an evaluation value that is ideally 0 with respect to a moving trajectory on a stationary object, and is not like a stationary object, for example, a moving trajectory on a moving object that moves at a higher speed. If the index takes an evaluation value that becomes larger in the + direction from 0, the evaluation value can be used as the stationary index E.
- the stationary index E can be obtained by the following (Equation 13).
- ⁇ 1 and ⁇ 2 are values for designating the balance of the respective geometric constraint error values. For example, both can be set to 1. A portion with low error sensitivity occurs in a single geometric constraint error, but by using a plurality of errors, variations in error sensitivity can be covered, and the value of the stationary index E can be obtained more stably.
- FIG. 11A shows an example of the stationary index E indicating the stationary object likeness of the movement locus.
- the size of the stationary index E indicating the stationary object likeness of the movement trajectory is represented by the thickness of the arrow line. That is, as shown in FIG. 11B, the thinner the arrow line, the smaller the stationary index E is, and it seems to be a stationary object.
- the camera itself moves forward, a large movement occurs in the movement locus of the background.
- the direction of the arrow indicates the direction of the movement trajectory.
- the stationary index E has a large value on the moving trajectory on the moving body, and the stationary index E has a small value on the moving trajectory on the background.
- the stationary index Ei in the movement trajectory on the leg of the walking person 1101 on the right side is large (for example, when using KLT (Kanade-Lucas-Tomasi) corresponding point and epipolar constraint,
- the stationary index Ei is about 5
- the stationary index Ej on the movement locus on the background is small (for example, when the KLT corresponding point and epipolar constraint are used, the stationary index Ej is about 0.5 at most).
- the movement trajectory on the moving object and the movement trajectory on the stationary object are greatly different in size, the movement trajectory on the moving object and the movement trajectory on the background are based on the stationary index E. Can be distinguished.
- the stationary index E may become smaller.
- the stationary index Ek of the movement trajectory on the left walking person 1102 accidentally becomes a trajectory similar to the background, so the value of the stationary index E is small.
- the value of the stationary index E may increase due to an error in the movement trajectory calculation.
- the stationary index E indicating the likelihood of a moving object as a stationary object is close to the values of the stationary index (for example, Ek) of the moving path on the moving body and the stationary index (for example, Ej) of the moving path on the background. There is a case. As a result, even a movement trajectory on a moving object may be erroneously detected as a background.
- the moving locus in addition to the stationary index indicating the stationary object likeness of the moving locus, the moving locus is clustered in consideration of the similarity between the moving loci, thereby stably moving. Aim to detect the body. More specifically, a stationary index indicating the likelihood of a moving locus as a stationary object is used as a weighting for clustering based on the similarity between moving tracks described later.
- the distance calculation step S304 is executed by the distance calculation unit 104. That is, the distance calculating unit 104 uses a plurality of movement trajectory x i of the movement trajectory calculation unit 102 has calculated, to calculate a distance representing a similarity between movement trajectories. That is, the Euclidean distance f (i, j) and the geodetic distance g (i, j) between the movement locus of the pixel i and the movement locus of the pixel j are calculated stepwise.
- both the Euclidean distance and the geodetic distance are distances representing the similarity of the movement trajectories, both are 0 between the completely identical movement trajectories, and conversely, the lower the similarity between the movement trajectories, the larger the positive. It is a distance that takes a distance value (including ⁇ ).
- FIG. 12A is a diagram illustrating an example of the configuration of the distance calculation unit 104.
- the distance calculation unit 104 includes an Euclidean distance calculation unit 1201 and a geodesic distance calculation unit 1202 between moving trajectories.
- the Euclidean distance calculation unit 1201 calculates the Euclidean distance f (i, j) between the movement locus of the pixel i and the movement locus of the pixel j using (Equation 14).
- the Euclidean distance f (i, j) calculated by (Equation 14) is defined between all the movement trajectories for convenience of description, but it is N movements that have a finite value as the Euclidean distance. It is only the between the trajectory x i.
- the Euclidean distance in this Embodiment was calculated by (Formula 14), it is not limited to this formula. Similar to (Equation 14), the Euclidean distance may be an index representing the geometric similarity such as the position, motion, acceleration, rotational speed, etc. in the picture coordinates between the movement trajectories. Equation 15) may be used.
- w is a weighting factor and is a parameter set by the designer.
- the Euclidean distance f (i, j) between the movement trajectories in (Expression 15) is obtained by adding a time fluctuation component of the distance of the picture coordinates to the time average of the distance of the picture coordinates between the movement trajectories.
- the time fluctuation component of the distance between the moving trajectories indicates the similarity of the movement of the moving trajectories, so that even when there is a shape change, the similarity between the moving trajectories can be captured more accurately. .
- a group of Euclidean distances f (i, j) between the movement trajectories calculated by the above procedure is represented as a Euclidean distance matrix Fdist .
- the geodesic distance calculation unit 1202 between the movement trajectories of the distance calculation unit 104 calculates the geodetic distance g (i, j) from the Euclidean distance f (i, j) between the movement trajectories.
- the geodesic distance calculation unit 1202 between the movement trajectories uses a threshold R predetermined for the Euclidean distance f (i, j) calculated by the Euclidean distance calculation unit 1201, and performs nonlinearization expressed by (Equation 17).
- the calculated distance f ′ (i, j) is calculated.
- the geodesic distance calculation unit 1202 between the movement trajectories calculates a geodetic distance from the non-linearized distance f ′ (i, j).
- “Geodetic distance” means that for a plurality of data points defined in a certain space, the connection between these data points and the distance between the connected data points are obtained. It is the shortest distance among the distances of all the routes that can be connected.
- the geodesic distance calculation unit 1202 between the movement trajectories calculates the geodetic distance from the i- th movement trajectory x i to the j-th movement trajectory x j using any one of the other plurality of movement trajectories as a relay point.
- i shortest path among all the paths from the movement trajectory x i reach the movement locus x j of the j of is calculated as geodesic distances.
- the route connecting the two points of the movement locus x i and the movement locus x j includes a route for relaying another movement locus x s in addition to the node directly connecting the two points.
- the distance of this route is assumed to be f ′ (i, s) + f ′ (s, j).
- There are a plurality of paths connecting the two points of the movement trajectory x i and the movement trajectory x j is calculated as the geodetic distance g (i, j) (Equation 18).
- min (x, y,...) Is a function that returns the smallest value among the values x and y.
- s is a movement trajectory x s, which is a relay point for tracing (tracing) from the movement trajectory x i to the movement trajectory x j .
- the relay point s at f ′ (i, s) + f ′ (s, j) is not limited to one point.
- the geodesic distance calculation unit 1202 between the trajectories calculates the geodesic distance g (i, j) from the Euclidean distance f (i, j) between the trajectories that continue for a long time.
- the calculation method of geodesic distance is not limited to the said (Formula 17) and (Formula 18).
- the most different point between the Euclidean distance and the geodetic distance is the relationship between the two data points for which the distance is obtained and other data points.
- the Euclidean distance is defined only from two data points regardless of the state of other data points.
- two data points and a path that can connect the two data points can be connected.
- Geodesic distance is defined as a distance including some other data point, that is, it may be influenced by the state of another data point.
- a set of calculated geodetic distances g (i, j) between the movement trajectories is expressed as a geodetic distance matrix G dist (Equation 19).
- the geodesic distance calculation unit 1202 between the movement trajectories calculates a geodetic distance g (i, j) representing the similarity between the N movement trajectories, and outputs it as a geodetic distance matrix G dist .
- FIG. 13A is a diagram illustrating an example of a plurality of movement trajectories on a picture. Although the movement trajectory is also calculated in the background area, the movement trajectory of the background area is not shown here for ease of description.
- FIG. 13B is a diagram illustrating a distribution of data of a plurality of movement trajectories each represented by (Expression 2).
- Each data point " ⁇ " mark in FIG. 13B corresponds to the movement trajectory x i of the pixel i shown in equation (2).
- the movement trajectory x i is a vector composed of independent T ⁇ 2 variables. Therefore, the movement trajectory is originally (T ⁇ 2) dimensional space data at maximum, but in FIG. 13B, it is represented as a point in the three-dimensional space for convenience of description.
- An arrow 1301 in FIG. 13B represents the Euclidean distance f (i, j) between the movement trajectory x i and the movement trajectory x j obtained by (Expression 14). That is, the Euclidean distance 1301 between the data point i and the data point j is a distance directly connecting the data.
- an arrow 1302 in FIG. 13C represents a geodetic distance g (i, j) between the movement trajectory x i and the movement trajectory x j obtained by (Equation 18).
- the geodetic distance 1302 between the data point i and the data point j is a distance obtained by tracing the relay data point s.
- the Euclidean distance between the head movement locus x i and the hand movement locus x j is indicated by an arrow 1401 in FIG.
- the Euclidean distance 1401 between the movement trajectories depends only on the two movement trajectories x i and x j for which the distance is to be obtained, and is a distance unrelated to the other movement trajectories.
- the geodetic distance 1402 between the movement trajectory x i and hand portion movement trajectory x j of the head shown in FIG. 14 (b).
- the geodetic distance 1402 is the sum of the distances between the plurality of travel trajectories that have passed through, the geodesic distance 1402 is affected by the travel trajectories other than the travel trajectories x i and x j .
- the Euclidean distance 1401 in FIG. 14A does not reflect the distribution of other moving trajectories at all. For this reason, in a mobile body connected by a joint like a person, the distance between the movement trajectories takes a value independent of the shape.
- the geodetic distance 1402 shown in FIG. 14B is a distance reflecting another movement locus. Therefore, in a mobile body connected by a joint, the distance between the movement trajectories takes a value depending on the shape of the mobile body. That is, since the connection at the joint is included in the distance as information, it can be used for detecting a moving body whose shape changes like a person.
- the movement trajectory x i represented by (Equation 2) is mathematically data of a maximum (T ⁇ 2) dimensional space.
- the movement trajectory actually obtained from the picture has the property of being localized in a very small part of the (T ⁇ 2) dimensional space as shown in FIGS. 13B and 13C. ing.
- the Euclidean distance which is obtained only from the distance between two data regardless of the data distribution
- the geodetic distance reflecting the density of neighboring data is more suitable than the Euclidean distance 1401 in FIG.
- predetermined K threshold values Rk instead of the predetermined K threshold values Rk, predetermined K threshold values Nk are used.
- the following processing may be performed as processing for obtaining a non-linear distance f′k (i, j) from Euclidean distance f (i, j). That is, the Euclidean distance threshold Rk or Euclidean distance f (i, j) to replace the infinity in place of the processing of equation (17), and a mobile trajectory x i and another (I-1) pieces of movement trajectory
- the non-linearized distance f′k (i, j) may be calculated by replacing Euclidean distance larger than the Nk-th Euclidean distance from the smallest one among f (i, j) by infinity.
- the non-linearized distance f′k is obtained by replacing the Euclidean distance with the movement locus larger than the Euclidean distance with the kth movement locus from the smallest with infinity. (I, j) may be calculated.
- the subclass classification step S305 is executed by the subclass classification unit 105. That is, the subclass classification unit 105 generates a subclass by clustering the collection of movement trajectories calculated by the movement trajectory calculation unit 102 with a certain index such as luminance or similarity of movement trajectories.
- the subclass classification unit 105 includes a Euclidean distance calculation unit 1501 that obtains the Euclidean distance between the movement trajectories, and a clustering unit 1502 that performs clustering based on the Euclidean distance.
- the clustering unit 1502 uses the Euclidean distance f (i, j) between the movement trajectory i and the movement trajectory j calculated by the Euclidean distance calculation unit 1501, and uses the movement trajectory i and the movement trajectory j in ascending order of f (i, j). Clustering is performed by repeating the process of bundling as a same class.
- the movement trajectory in the vicinity retains high similarity.
- high similarity means that the Euclidean distance f (i, j) between the movement trajectory i and the movement trajectory j is small.
- the fact that the Euclidean distance f (i, j) is small can be interpreted that the movement trajectory i and the movement trajectory j are distributed at a short distance in a high-dimensional space composed of the movement trajectories.
- FIG. 16B shows a conceptual diagram of a high-dimensional space made up of moving trajectories.
- a three-dimensional space is used for ease of explanation, but in reality, each element of the vector shown in (Expression 2) corresponds to each dimension. That is, the high-dimensional space is a space having 2 ⁇ T dimensions.
- the eight movement trajectories a to h are used, but actually, the movement trajectory obtained for each pixel may be used, or the movement trajectory obtained in units of blocks is used. May be.
- each data point on the high-dimensional space formed of the movement locus shown in FIG. 16B corresponds to one movement locus shown in (Equation 2). In other words, this corresponds to the result of tracking pixels not only over an area on one picture but over a plurality of pictures that differ in time.
- each class can be expected to correspond to an individual subject or a part of the subject, and subject detection and area division can be performed.
- each area to be divided is expressed as follows.
- M is the number of areas, and is determined empirically according to the scene to be used.
- the subclass classification unit 105 performs processing of setting different movement trajectories i and movement trajectories j to the same region label ⁇ m in ascending order of Euclidean distance f (i, j).
- one of the movement trajectory i and movement trajectory j is the case already belongs to the region theta k is still to belong to the even region theta k pixels area label is not granted. Further, if the movement trajectory i and the movement trajectory j already belong to different areas, the area labels are integrated. Next, it is determined whether all the movement trajectories are labeled and the number of areas is a predetermined M.
- FIGS. 16C and 16D A specific example of the subclass classification process will be described with reference to FIGS. 16C and 16D.
- a larger M that sufficiently divides the area of the moving object into fine units is used.
- f (a, b) when the distance between the movement locus a and the movement locus b is f (a, b), f (a, b) ⁇ f (g, h) ⁇ f (d E) ⁇ f (b, c) ⁇ f (f, g) ⁇ f (c, d).
- the subclass classification unit 105 assigns the same area label ⁇ 1 to the movement trajectory a and the movement trajectory b.
- the subclass classification unit 105 assigns the same area label ⁇ 2 to the movement trajectory g and the movement trajectory h.
- the subclass classification unit 105 assigns the same region label ⁇ 3 to the movement trajectory d and the movement trajectory e which are the third smallest distance.
- the next smallest distance is the distance f (b, c) between the movement trajectory b and the movement trajectory c.
- the subclass classification unit 105 assigns the same area label ⁇ 1 as the movement trajectory b to the movement trajectory c.
- the next smallest distance is a distance f (f, g) between the movement locus f and the movement locus g.
- the subclass classification unit 105 gives the same region label ⁇ 3 as the movement locus g to the movement locus f.
- the subclass classification unit 105 again integrates the area label ⁇ 1 to which the movement locus c belongs and the area label ⁇ 3 to which the movement locus d belongs with respect to the movement locus c and the movement locus d that form the next smallest distance. Further, the subclass classification unit 105 gives the region label ⁇ 1 to the movement trajectories a to e.
- the movement trajectory that is continuous in the high-dimensional space is determined as one class, and the distance between the movement trajectories is large.
- Each class can be separated using as a discontinuous point. And it becomes possible to utilize the movement locus which belongs to each class for a mobile body detection.
- the subclass classification unit 105 clusters the collection of the movement trajectory calculated by the movement trajectory calculation unit 102 according to a certain index such as luminance or similarity of the movement trajectory, thereby subclasses.
- the subclass does not necessarily need to include a plurality of movement trajectories. That is, subclass classification in which each subclass is composed of one movement locus may be performed.
- the distance calculation step S306 is executed by the weighted distance calculation unit 106.
- the weighted distance calculation unit 106 includes a representative geodetic distance calculation unit 1701 and a stationary index addition unit 1702.
- the representative geodetic distance calculation unit 1701 calculates a representative value (representative geodetic distance) between the subclasses generated by the subclass classification unit 105 based on the geodetic distance and the Euclidean distance calculated by the distance calculation unit 104. . Subsequently, the stationary index adding unit 1702 obtains a stationary index indicating the stationary object likeness of the movement trajectory included in each subclass based on the stationary index of each pixel calculated by the stationary index calculating unit 103, and indicates the calculated stationary object likeness. Based on the stationary index, the representative geodetic distance calculation unit 1701 calculates the representative geodetic distance between the subclasses.
- FIG. 18 shows two adjacent subclasses ⁇ i and ⁇ j among the plurality of classes generated by the subclass classification unit 105.
- “subclass” is expressed only as “class”.
- x i included in (Expression 21) and (Expression 22) is a movement trajectory expressed in the form of a multidimensional vector, as in (Expression 2).
- the distance obtained between the movement locus included in I and the movement locus included in J is defined as an interclass distance.
- distance is a concept including both Euclidean distance and geodetic distance.
- FIG. 18 the conceptual diagram of the representative value (representative geodetic distance) of the geodesic distance between classes is shown.
- g 31 g (i 3 , j 1)
- the moving object region detection is performed on a pixel-by-pixel basis by focusing on a collection of single movement trajectories in a class as shown in I and J and operating based on a macro distance in class units. Therefore, it is possible to cope with noise / false detection of a stationary index that occurs due to this. Therefore, it is desirable to calculate a representative value of the distance between classes. That is, the representative value of the distance between classes is desirably a representative value that can approximate the movement or positional relationship between the classes for a plurality of classes.
- the average value of the geodesic distance between the movement trajectories of each class can be used as the representative value. This is obtained by calculating a plurality of geodetic distances corresponding to all combinations between the movement trajectories included in each class and averaging them.
- the representative geodetic distance G ( ⁇ i , ⁇ j ) can be calculated by the following (Equation 24).
- only the geodetic distance may be calculated again so that g (i, j) ⁇ ⁇ .
- this representative geodetic distance is not limited to the average value of the geodetic distance.
- the median value of the geodesic distance between the movement trajectories of each class can be used as a representative value. This is obtained by obtaining a plurality of geodetic distances corresponding to all combinations between movement trajectories included in each class and taking the median among the plurality of classes.
- the representative value of the Euclidean distance can be similarly obtained and used.
- the representative geodetic distance G ( ⁇ i , ⁇ j ) can be calculated by the following (Equation 25).
- those points are excluded from the median calculation as in the case of the above average calculation.
- only geodetic distance calculation may be performed again so that g (i, j) ⁇ ⁇ .
- the mode value of the geodesic distance between the movement trajectories of each class can be used as the representative geodesic distance.
- These representative values are values that appear most frequently among a plurality of classes when a plurality of geodetic distances corresponding to all combinations of movement trajectories included in each class are obtained.
- those points are excluded from the mode value calculation as in the case of the above average value calculation.
- only geodetic distance calculation may be performed again so that g (i, j) ⁇ ⁇ .
- the stationary index adding unit 1702 weights the representative geodetic distance G between the classes based on the stationary index of the movement locus belonging to each class.
- the stationary index adding unit 1702 weights the representative geodetic distance G between the classes based on the stationary index of the movement locus belonging to each class.
- FIG. 19A is a conceptual diagram showing an example of movement trajectories respectively belonging to a moving object and a background superimposed on a scene image.
- the movement trajectory corresponding to the left person A is indicated by “ ⁇ ”
- the movement trajectory corresponding to the right person B is indicated by “ ⁇ ”
- the movement trajectory corresponding to the background is indicated by “ ⁇ ”.
- FIG. 19A shows only eight representative movement trajectories, and there are actually a larger number of movement trajectories. Here, other movement trajectories are omitted for easy viewing.
- FIG. 19B shows, in a high-dimensional space, the result of class classification performed by the subclass classification unit 105 on the movement trajectory existing in the scene shown in FIG. 19A. Again, for ease of viewing, typical movement trajectories are shown in a two-dimensional space.
- the movement trajectory in the scene is observed as the movement component of the camera added to the movement component of the subject itself such as the moving object and the background. For this reason, the larger the camera motion, the more dominant the camera motion component in the movement trajectory distribution (for example, the movement trajectory distribution shown in FIG. 19B), making it difficult to distinguish between the movement trajectory on the moving object and the background movement trajectory. become.
- the background becomes a radial movement locus centering on the spring point.
- the degree of similarity between them may be relatively high, that is, the geodesic distance may be shortened.
- the greater the camera movement the closer the background movement locus distribution and the moving body movement locus distribution in the movement locus distribution shown in FIG. 19B, the background subclass and the moving body subclass are: Cannot be separated and integrated.
- FIG. 19C shows a result of the stationary index adding unit 1702 determining whether the stationary trajectory is stationary or moving with respect to each moving locus.
- the same movement trajectory as shown in FIG. 19B is displayed at the same position in the two-dimensional space as the movement trajectory shown in FIG. 19C.
- stillness determination value E MS used in the determination of the stationary or moving exhibited here is 0 or a binary 1, this is only an example, the stationary object, the above-mentioned binary if the different values mobile Not necessarily.
- the stationary index adding unit 1702 evaluates stationary or movement in units of subclasses and corrects erroneous determination.
- each subclass is often obtained as a small area on the moving object or a small area on the background. That is, it can be assumed that the movement trajectories in the subclass are all moving objects or all stationary objects. Therefore, the stationary index adding unit 1702 corrects the above-described erroneous determination as seen in the movement trajectories 1901 and 1902 of FIG. 19C by evaluating the stationary or moving in units of subclasses.
- the evaluation value indicating the evaluation determination result of stationary or moving in units of subclasses is set as an inter-subclass stationary determination value E SUB-MS .
- Subclasses between stillness determination value E SUB-MS, like stillness determination value E MS movement locus shall take two values 0 or 1.
- the inter-subclass stillness determination value E SUB-MS is binary
- the inter-subclass stillness determination value E SUB-MS is obtained by determining the magnitude of the average value of the multi-level stillness evaluation value E and a predetermined threshold value TH E. (Equation 31).
- the object of the present invention is to separate the moving body from the background.
- FIG. 19B in order to make the subclass on the moving object and the subclass on the background close to each other in a high-dimensional space separable and integrated, the subclasses that are stationary backgrounds are closer to each other and vice versa.
- the stationary index adding unit 1702 sets a weighting rule as shown in FIG. 19E and sets a weighted geodetic distance G w ( ⁇ i , which is a new subclass distance defined between the subclasses ⁇ i and ⁇ j . ⁇ j ) is calculated (Expression 32, Expression 33).
- the stationary index adding unit 1702 applies a weight W that shortens (closes) the distance to the representative geodetic distance G ( ⁇ i , ⁇ j ) between the subclasses. That is, a weight Wmin is applied so that W ⁇ 1.
- G w ( ⁇ i , ⁇ j ) Wmin ⁇ G ( ⁇ i , ⁇ j ) ⁇ G ( ⁇ i , ⁇ j )
- the stationary index adding unit 1702 applies a weight W for increasing (decreasing) the distance to the representative geodetic distance G ( ⁇ i , ⁇ j ) between the subclasses. That is, a weight Wmax is applied so that W> 1.
- G w ( ⁇ i , ⁇ j ) Wmax ⁇ G ( ⁇ i , ⁇ j )> G ( ⁇ i , ⁇ j )
- a subclass between stillness determination value E SUB-MS was an evaluation value of 2 values determined from the stillness determination value E MS, multi subclasses between stillness determination value E SUB-MS It may be a value.
- the inter-subclass stillness determination value E SUB-MS may be obtained directly from the stillness evaluation value E as in the following Expression 34.
- (Equation 33) defining the weighting rule becomes (Equation 35).
- the weight Wmin in the case of both subclasses of the stationary object is applied, and the subclass between the two subclasses is multiplied.
- the weights W ( ⁇ i , ⁇ j ) are also set to three values of Wmin, Wmax, and 1 for ease of explanation, but may be multivalued.
- the inter-subclass stillness determination values E SUB-MS ( ⁇ i ) and E SUB-MS ( ⁇ j ) both approach 0 and the weight W ( ⁇ i , ⁇ j ) approaches Wmin, and the inter-subclass stillness determination
- the weight W ( ⁇ i , ⁇ j ) approaches 1
- the inter-subclass stationary determination value E SUB-MS ( ⁇ i) ) E SUB-MS ( ⁇ j ) may be a multi-value so that the weight W ( ⁇ i , ⁇ j) approaches Wmax as one approaches 0 and the other increases.
- Equation 35 when both subclasses are stationary objects, weighting is performed to shorten the distance between the subclasses (Equation 36), or one of the two subclasses is the background and the other is the stationary object. In such a case, weighting for increasing the distance between subclasses (Equation 37) may be used.
- the weight W ( ⁇ i , ⁇ j ) in (Equation 36) described above is a binary value of Wmin, 1, but may be multivalued.
- the weight W ( ⁇ i , ⁇ j ) approaches Wmin as the subclass static determination values E SUB-MS ( ⁇ i ) and E SUB-MS ( ⁇ j ) both approach 0, and in other cases Any multi-value may be used so that the weight W ( ⁇ i , ⁇ j ) approaches 1.
- the weight W ( ⁇ i , ⁇ j ) of (Expression 37) described above is a binary value of Wmax and 1, but may be multivalued.
- the weight W ( ⁇ i , ⁇ j ) approaches Wmax as one of the subclass static determination values E SUB-MS ( ⁇ i ) and E SUB-MS ( ⁇ j ) approaches 0 and the other increases.
- it may be a multivalued value such that the weight W ( ⁇ i , ⁇ j ) approaches 1.
- the moving object can be more correctly separated from the background by changing the weight according to the reliability of the inter-subclass static determination value E SUB-MS .
- FIG. 19F shows a diagram showing the weighted geodetic distances G w ( ⁇ i , ⁇ j ) calculated by the above weighting rules in a high-dimensional space. As in FIG. 19B, the high dimension is displayed in two dimensions for convenience.
- Region dividing section 107 as the evaluation value of the region division of the subclass theta p, based weighted geodesic distance G w ( ⁇ p, ⁇ q ) between subclasses calculated by the weighted distance calculation unit 106, the subclass theta p It is determined whether to divide the area division candidates as separate clusters.
- w ( ⁇ p , ⁇ q ) is used as the distance between the subclasses ⁇ p and ⁇ q so far, but for the sake of explanation, the “weighted geodesic distance G between the subclasses will be described below.
- the region dividing unit 107 has a sufficient distance between the two subclasses ⁇ p and ⁇ q. Select as a separate class and confirm as an individual class.
- the area dividing unit 107 determines the corresponding two subclasses ⁇ p and ⁇ q as the same class. That is, in this case, it is decided not to divide. Then, after determining whether or not to divide all subclasses of the region division candidates, the region dividing unit 107 assigns different labels ⁇ m to the movement trajectories belonging to different classes, and the region division information of the movement trajectory Output as.
- the number of subclasses may be larger, but here only seven subclasses are displayed for ease of viewing.
- FIG. 20A It is shown in Figure 20A, the geodesic distance between subclasses of class number 2 corresponding to the person A in FIG. 19A and h 1, 2.
- h 1,2 ⁇ Ht the corresponding subclasses ⁇ 1 and ⁇ 2 are divided.
- h 1,2 ⁇ Ht the corresponding subclasses ⁇ 1 and ⁇ 2 are divided.
- the region dividing unit 107 generates an image from the moving locus cluster ⁇ p to which the label is assigned by the threshold value Ht and displays the image on the display 120 by the above procedure.
- the region dividing unit 107 performs image processing on the input picture so that the result of the integrated subclass can be visually recognized with respect to the moving image received by the image input unit 101, and outputs it. This is displayed on the display 120.
- FIG. 20B to 20D show examples of pictures generated by the area dividing unit 107.
- FIG. 20B to 20D show examples of pictures generated by the area dividing unit 107.
- the threshold value Ht is set in advance, but is not limited to this. Specifically, the threshold value Ht may be changed according to the magnitude of the movement of the moving object to be extracted, or the threshold value Ht may be changed according to whether the moving object is a person or a car. .
- the threshold value Ht corresponds to an evaluation criterion for setting two different region division candidates as different regions or the same region. For example, when two area division candidates correspond to two moving objects, respectively, by reducing the threshold value Ht, even when the relative position or movement difference between the two moving objects is small, Can be extracted. Conversely, by increasing the threshold value Ht, it is possible to extract two areas only when the relative position and movement of the two moving bodies are large. That is, there is an effect that the region extraction target can be changed depending on the threshold value Ht.
- 20B to 20D show examples in which the above processing is performed and the results are displayed on the display.
- the subclass classification unit 105 For each subclass calculated by the subclass classification unit 105, for each subclass extracted as the same mobile object, when color classification is performed as a single mobile object, the subclass classification result displayed as shown in FIG.
- the subclass integration result can be displayed as shown in FIG. 20C.
- two moving objects are displayed in two different colors.
- a display form By adopting such a display form, there is an effect that a moving body in a picture can be detected correctly and the result can be displayed in an easy-to-understand manner.
- the pixel at the picture coordinate position of the movement trajectory is drawn with a color corresponding to the segment region as a procedure for generating a picture.
- the method for generating a picture is not limited to this. Specifically, if the number of movement trajectories and the number of pixels in the entire picture are the same, all the pixels in the picture can be drawn with a color corresponding to the class by the method described above. On the other hand, when the number of moving loci is smaller than the number of pixels in the entire picture, there are pixels that do not match the picture coordinate positions of any moving loci.
- Such pixels that do not match the picture coordinate position of the movement locus may be drawn by another method.
- the movement trajectory calculation unit 102 may draw the pixels belonging to the block used for motion detection for generating a certain movement trajectory in the same color as the subclass to which the movement trajectory belongs.
- a pixel that does not match the picture coordinate position of the movement locus it may be drawn in the same color as the subclass to which the nearest movement locus belongs.
- the picture coordinate positions (points) of a plurality of movement trajectories are connected by the Delaunay triangular mesh generation method and belong to the same class.
- the pixels included in the triangle surrounded by the points may be drawn with the same color as the class.
- steps S301 to S307 described above may be repeated each time T pictures are input.
- the movement based on the representative value of the geodesic distance of the movement trajectory between the subclasses in the picture and the stationary index of the movement trajectory of each subclass.
- the moving body area in the picture can be detected regardless of the posture of the moving body as a result of temporally tracking the area of the object moving in the picture.
- the region dividing processing based on the subclass is not necessarily performed.
- the geodetic distances corresponding to all the geodesic distances between all the movement trajectories corresponding to each subclass are supported.
- the weighting is performed using the same weight as the weight given to the representative geodesic distance between subclasses to calculate the geodesic distance between the weighted movement trajectories, and the area dividing unit 107 performs the weighted movement instead of the weighted geodesic distance Gw.
- the area may be divided based on the geodesic distance between the trajectories.
- the same processing as that of the subclass classification unit 105 and its modification is performed on the geodesic distance between weighted movement trajectories, thereby dividing the picture into areas and separating the moving object from the background. be able to.
- the above-mentioned weighting is applied to the geodetic distance, that is, the movement trajectory of any stationary object with respect to the geodetic distance between the movement trajectories of any stationary object based on the stationary index of the movement trajectory and the distance between the movement trajectories.
- the geodetic distance between the moving trajectories is calculated based on the distance between the moving trajectories, and each of the calculated geodetic distances based on the stationary index of the moving trajectory is compared to that before weighting.
- Calculate the geodetic distance between weighted moving trajectories by giving a weight that increases the ratio of the geodesic distance between the moving trajectory of any stationary object and the moving trajectory of any moving object to the geodetic distance between the moving trajectories is doing.
- the moving object region corresponding to the moving object's moving locus is detected by separating the moving object's moving locus and the moving object's moving locus.
- a moving body can be correctly detected even in an image that includes a moving body such as a person that moves while changing its shape and is captured by a moving camera.
- the distance calculation unit 104 calculates the distance and inputs it to the weighted distance calculation unit 106, but a time change of the distance can be used instead of the distance.
- the difference between the distance obtained from the movement trajectory of t frame to 2t frame and the distance obtained from the movement trajectory of 2t frame to 3t frame is called a time change of distance, and this is replaced with the movement trajectory of t frame to 2t frame.
- the same effect can be obtained by using it.
- the number T of pictures for obtaining the movement trajectory extracted by the distance calculation unit 104 is predetermined and has been described as being constant.
- the number of pictures T used for calculating the trajectory may be dynamically changed.
- an upper limit value of the number of movement trajectories is determined in advance, and the number of pictures T is dynamically changed so as not to greatly exceed these upper limit values. Also good. Specifically, when the upper limit value Nmax of the number of moving tracks is determined in advance, and the number of moving tracks extracted from the moving track over the predetermined number T of pictures in the distance calculation unit 104 exceeds Nmax.
- the number of pictures T may be increased.
- the large number of movement trajectories means that many corresponding points across all pictures can be calculated for reasons such as small changes in the moving image, small movement of the subject, and small blockage area due to movement. It corresponds to. In such a case, by increasing the number of pictures T, it can be expected that the moving object can be correctly detected by including more motion information in the moving region.
- a lower limit value of the movement locus number is determined in advance, and the number of pictures T is dynamically changed so as not to fall below these lower limit values. It may be changed.
- a lower limit value Nmin of the number of moving tracks is determined in advance, and when the number of moving tracks over a predetermined number of pictures T falls below Nmin in the distance calculation unit 104, the number of moving tracks is Nmin.
- a smaller number of pictures T may be used so as to exceed.
- the Euclidean distance is used as the linear distance, but as described above, the Euclidean distance is not limited. A form obtained by using an arbitrary distance index defined by the linear distance is also included in the present invention.
- the mobile object detection apparatus and method according to the present invention are not limited to the above-described embodiments, and are obtained by modifying the above-described embodiments by a person skilled in the art and various modifications described below. Forms obtained by arbitrarily combining components in a plurality of forms including examples are also included in the present invention.
- Modification 1 Next, a moving body detection apparatus according to Modification 1 of the embodiment of the present invention will be described.
- the moving body detection apparatus includes an interface for adjusting the threshold value TH MS used when obtaining the stationary determination value E MS .
- the optimum value of the threshold value TH MS used when obtaining the stillness determination value E MS varies depending on the size of the camera movement, the moving speed of the moving body, and the like. Therefore, in practice, when a shooting scene or camera setting changes, a predetermined fixed threshold value TH MS is not optimal, and when a fixed threshold value TH MS is used, it is accurate. There may be a case where the moving object cannot be detected.
- the moving body detection device 100 calculates the inter-subclass stillness determination value E SUB-MS based on the fixed threshold TH MS , and weights the inter-subclass distance based on the calculated inter-subclass stillness determination value E SUB-MS.
- the result of area division is displayed on the display 120.
- the user inputs the threshold value TH MS by trial and error after confirming the displayed region division result.
- region division by the moving body detection apparatus 100 is performed again, and after confirming the displayed region division result, the procedure of inputting the threshold value TH MS again is repeated.
- the trial and error threshold TH MS by repeating the procedure of input Segmentation and threshold TH MS, it is necessary to find the optimal threshold TH MS.
- the threshold value TH MS is a threshold value applied to the value of the stationary index E. That is, it can be said that the threshold value TH MS largely depends on the distribution of the value of the stationary index E. Therefore, if the value of the stationary index E and its distribution on the image can be confirmed, the threshold value TH MS can be adjusted without performing the region division processing to the end. Thereby, the optimal threshold value TH MS can be determined earlier.
- FIG. 21 is a block diagram illustrating a configuration of a moving object detection device 100L according to the first modification.
- the mobile object detection device 100L includes a stationary index calculation unit 103A and a weighted distance calculation unit 106A in place of the static index calculation unit 103 and the weighted distance calculation unit 106 in the configuration of the mobile object detection device 100 illustrated in FIG. Each configuration is used.
- FIG. 22 is a block diagram showing a detailed configuration of the stationary index calculation unit 103A.
- the stationary index calculation unit 103A has a configuration in which a stationary index image creation unit 703 is further added to the stationary index calculation unit 103 illustrated in FIG.
- the still index image creating unit 703 causes the display 120 to display each block constituting the picture in a display mode according to the calculated value of the still index. Details of processing executed by the stationary index image creation unit 703 will be described later.
- FIG. 23 is a block diagram showing a detailed configuration of the weighted distance calculation unit 106A.
- the weighted distance calculation unit 106A has a configuration in which a threshold value input unit 1703 is added to the weighted distance calculation unit 106 illustrated in FIG. 17 and a stationary index adding unit 1702A is used instead of the stationary index adding unit 1702.
- the threshold value input unit 1703 is a processing unit that receives a stationary index threshold value input by the user, and may be configured by a keyboard, a mouse, or the like, for example.
- the stationary index adding unit 1702A obtains a stationary index by the same method as the stationary index adding unit 1702, and weights the representative geodetic distance between the subclasses based on the stationary index.
- the point threshold TH MS still indicators used in this case is one in which the threshold value input unit 1703 has received is different from the stationary index adding unit 1702.
- FIG. 24 is a diagram illustrating an example of an image created by the stationary index image creation unit 703 of the stationary index calculation unit 103A.
- the stationary index image creation unit 703 classifies the values of the stationary index E into five levels, and displays an image in which each block is hatched according to the level of the stationary index E on the display 120.
- the still index image creating unit 703 may perform hue gradation display on the display 120 according to the value of the still index E.
- the stationary index image creation unit 703 includes a threshold value input unit, and can binarize and display the value of the stationary index E using a threshold value separately input by the user from the threshold value input unit.
- a pixel having a stationary index E larger than a threshold input by the user may be displayed in red, and a pixel having a stationary index E having a value other than that may be displayed in blue. By doing so, it is possible to determine the actual threshold when you enter there, stillness determination value value how made or easily visible while threshold is E MS.
- the threshold value input unit 1703 receives a threshold value TH MS input by the user at an arbitrary timing.
- the stationary index adding unit 1702A calculates the stationary determination value E MS using the input threshold value TH MS , the region dividing unit 107 performs region division, and the result is displayed on the display 120. For example, as shown in FIG. 25A, a slider 2211 is displayed on the display 120. The user inputs the threshold value TH MS by moving the slider 2211 left and right.
- the threshold value TH MS is set to a large value as shown in FIG. 25A, for example, the moving body 2111 having a large stationary index E is extracted. In contrast, as shown in FIG.
- Modification 2 Next, a moving body detection apparatus according to Modification 2 of the embodiment of the present invention will be described.
- the subclass classification unit 105 has been described as obtaining a subclass based on the Euclidean distance f (i, j) calculated in (Equation 14).
- the operation of the subclass classification unit 105 is not limited to this. That is, the subclass classification unit 105 may classify a plurality of movement trajectories into a plurality of subclasses based on the color similarity between blocks belonging to each movement trajectory between the movement trajectories.
- movement trajectory clustering is performed based on the color similarity of pixels will be described.
- FIG. 26A is a diagram illustrating a configuration of a moving object detection device according to Modification 2 of the embodiment.
- the moving object detection device 100A includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 2101, a weighted distance calculation unit 106, and An area dividing unit 107 is included.
- processing units other than the subclass classification unit 2101 are the same as those in the above embodiment, the description thereof is omitted.
- a method for calculating the subclass in the subclass classification unit 2101 instead of the method of calculating the subclass by labeling the similar movement locus described in the above embodiment, so-called “superpixel” is used based on the color similarity of the pixel.
- a method of dividing a picture into a plurality of called subclasses may be used.
- a method for calculating the superpixel a graph-based method or the like can be used. A detailed description of the processing procedure is omitted because it is described in Non-Patent Document 7 and the like, but by estimating the boundary between the regions based on the graphical representation of the picture, the picture is maintained while maintaining efficient and global features.
- FIG. 26B shows the configuration of subclass classification unit 2101 in the second modification of the present embodiment.
- the subclass classification unit 2101 includes a clustering unit 2102.
- the clustering unit 2102 classifies pictures into a plurality of subclasses based on the above-described color similarity.
- the configuration of the subclass classification unit 2101 of this modification does not require the Euclidean distance calculation unit, so classification into subclasses can be performed at a high speed with a simpler configuration. There is an effect that can be performed.
- pictures can be separated into subclasses, and the movement trajectory belonging to each subclass can be used for moving object area detection.
- the subclass classification unit 105 has been described as obtaining a subclass based on the Euclidean distance f (i, j) calculated in (Equation 14).
- the operation of the subclass classification unit 105 is not limited to this.
- an example in which classification into subclasses is performed by performing dimensional compression of the geodetic distance g (i, j) will be described.
- FIG. 27A is a diagram illustrating a configuration of a moving object detection device according to a third modification of the embodiment.
- the moving object detection apparatus 100B includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 2201, a weighted distance calculation unit 106, and an area division unit 107.
- processing units other than the subclass classification unit 2201 are the same as those in the above embodiment, the description thereof is omitted.
- FIG. 27B shows the configuration of the subclass classification unit 2201 in Modification 3 of the present embodiment.
- the subclass classification unit 2201 includes a second distance calculation unit 2202 and a clustering unit 2203. Unlike the configuration of the subclass classification unit 105 illustrated in FIG. 15, the subclass classification unit 2201 includes a second distance calculation unit 2202 instead of the Euclidean distance calculation unit 1501.
- the second distance calculation unit 2202 calculates the Euclidean distance f (i, j) from the movement locus calculated by the movement locus calculation unit 102 according to (Expression 14), and then the geodetic distance according to (Expression 17) and (Expression 18). Find g (i, j). Note that the procedure for calculating the geodetic distance g (i, j) here is the same as the operation in the distance calculation unit 104 described in the above embodiment, and thus the description thereof is omitted.
- the clustering unit 2203 performs dimension compression of the calculated geodetic distance g (i, j), and then uses the dimension-compressed geodetic distance to give the number of classes and minimize the intra-class variance. Cluster movement trajectories.
- Dimensional compression can be realized by obtaining Eigen system after performing Young-Householder conversion. This is a method for efficiently projecting data distributed in a multidimensional space to a low-dimensional space by dimensional compression.
- a procedure for clustering the movement trajectory by performing dimension compression of the geodetic distance g (i, j) in the clustering unit 2203 will be described.
- the matrix formed by the geodetic distance g (i, j) is the geodetic distance matrix G (formula 38).
- the clustering unit 2203 first performs Young-Householder transformation on the geodetic distance matrix G by applying the centering matrix H from both sides. This is performed in order to convert the distance matrix into a distance matrix having the center of gravity as the origin, whereas the distance matrix is a distance matrix composed of distances between points.
- I is a unit matrix
- N is the number of movement trajectories.
- the clustering unit 2203 calculates P eigenvectors e p and corresponding eigenvalues ⁇ p for ⁇ (G) in order to perform dimensional compression.
- e p i is the i-th element of the p-th eigenvector e p .
- the number of eigenvectors P may be determined experimentally according to the scene to be used, or may be determined based on the contribution rate a p calculated from the eigenvalue ⁇ p as follows.
- P is the number of eigenvectors to be used, that is, the number of dimensions of the compressed space.
- N is the number of all eigenvectors. Therefore, P when the contribution rate a p is equal to or greater than a certain value may be the number of eigenvectors.
- the clustering unit 2203 performs dimension compression of the geodetic distance g (i, j) by the processing of (Equation 39) to (Equation 44).
- the geodetic distance g i calculated by the second distance calculation unit 2202 and the corresponding virtual movement trajectory can be associated with the data z p i in the dimension-compressed space spanned by the eigenvector e p .
- FIG. 28A is a diagram illustrating a data distribution of a movement trajectory before dimension compression in a multidimensional space.
- the multidimensional space is a three-dimensional space, but in reality, each element of the vector shown in (Equation 2) corresponds to each dimension.
- FIG. 28B shows a space in which the multidimensional space of the movement locus shown in FIG. 28A is dimensionally compressed.
- the horizontal and vertical axes in FIG. 28B are eigenvectors e 1 and e 2 , respectively.
- the two-dimensionally projected points (z 1 i , z 2 i ) are projections of the geodetic distance g i .
- (z 1 i , z 2 i ) can be regarded as corresponding to the movement trajectory x i of the pixel i.
- the number of dimensions of the nonlinear space is set to two. However, as described above, the number of dimensions does not necessarily need to be two, and a higher number of dimensions has higher accuracy. Data can be projected.
- the clustering unit 2203 by performing clustering on the data z p i that the movement locus dimensional compression as shown in FIG. 28B, performs clustering of movement trajectory.
- a method of clustering the movement trajectory so as to minimize the intra-class variance given the number of classes is used.
- M is the number of subclasses and is determined empirically according to the scene to be used.
- Each subclass ⁇ m is a parameter
- the initial value may be determined at random, or the coordinate value of the intersection may be set as the initial value by dividing the compressed nonlinear space into equal intervals by a grid.
- C m is the number of data belonging to the subclass ⁇ m on the compressed nonlinear space.
- the subclass ⁇ m to which the data z i belongs is obtained using the distance function of the following (formula 48).
- ⁇ m (z i ) indicates the distance between the data z i corresponding to the movement locus of the pixel i and each subclass ⁇ m .
- Each data belongs to a subclass ⁇ m in which the distance ⁇ m (z i ) has a minimum value.
- ⁇ m (z i ) is the Mahalanobis distance
- ⁇ m (z i ) may be used in place of ⁇ m (z i ).
- P (z i ) is a priori probability of z i in the likelihood function framework. Therefore, p (z i ) may be a constant value, or if it is known that the target scene includes a fixed subject such as a person, it may be set in advance based on the shape or area ratio of the person part. good. This is particularly effective when the density of the data z i is uneven. For example, when it is known that the density of the data z i is high, if the dense data z i is desired to be in the same subclass, the corresponding prior probability p (z i ) may be set large.
- the corresponding prior probability p (z i ) may be set small.
- the density of the data z i may be a density on an image space or a density on a compressed nonlinear space.
- Equation (48) using the data z i that belongs to a subclass theta m, subclasses theta m as follows:
- z cm is data on a compressed nonlinear space belonging to the subclass ⁇ m .
- the subclass ⁇ m to which each piece of data in the nonlinear space belongs can be obtained by repeating the distance calculation and the parameter update of (Expression 48) to (Expression 51) a specified number of times.
- other clustering methods such as k-means and competitive learning may be used.
- FIG. 28C shows the result of applying the above clustering processing to the dimensionally compressed data as shown in FIG. 28B. It can be seen that the data points are subclassed.
- FIG. 28D shows an example in which the clustering process is similarly performed on the moving person data in the nonlinear space. Note that the corresponding person area is also shown in the dimensionally compressed data distribution diagram. Looking at the correspondence on the picture for subclasses ⁇ 1 to ⁇ 2 in the compressed nonlinear space, ⁇ 1 is the human head, ⁇ 2 is the human torso, and so on. It corresponds.
- each part of the human body corresponds to a subclass in a compressed non-linear space is that pixels are tracked over a plurality of temporally continuous pictures.
- clustering By performing clustering on the compressed nonlinear space, an image region can be extracted for each moving subject in the picture as a result of temporally tracking the region of the object moving in the picture.
- the Euclidean distance is used as the linear distance connecting two points.
- the linear distance is not limited to the Euclidean distance.
- a form obtained by using an arbitrary distance index defined by the linear distance is also included in the present invention.
- the subclass classification unit 105 has been described as calculating a subclass based on the Euclidean distance f (i, j).
- the operation of the subclass classification unit 105 is not limited to this.
- an example will be described in which a plurality of geodesic distances are generated to generate subclass candidates, and the subclass classification is performed by selecting from these candidates.
- FIG. 29A is a diagram illustrating a configuration of a moving object detection device according to a fourth modification of the embodiment.
- the moving object detection device 100C includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 2401, a weighted distance calculation unit 106, and An area dividing unit 107 is included.
- FIG. 29B shows the configuration of subclass classification section 2401 in Modification 4 of the present embodiment.
- the subclass classification unit 2401 selects, for each of a plurality of movement trajectories, a distance that is equal to or smaller than a predetermined distance threshold among a plurality of distances from the movement trajectory to another movement trajectory, and sets the unselected distance to infinity. After delinearizing to change, calculate the geodetic distance by finding the shortest path from the moving trajectory to another moving trajectory, and the same set of moving trajectories where the geodesic distance between the moving trajectories is a finite value By classifying into subclasses, each movement trajectory is classified into one of a plurality of subclasses.
- the subclass classification unit 2401 includes a third distance calculation unit 2402, a subclass candidate generation unit 2403, and a subclass candidate selection unit 2404.
- the third distance calculation unit 2402 calculates the Euclidean distance f (i, j) from the movement trajectory calculated by the movement trajectory calculation unit 102.
- Geodetic distance conversion is performed by setting a determination criterion to obtain a geodetic distance g (i, j).
- the subclass candidate generation unit 2403 detects discontinuous points in the distribution of the distance between the moving trajectories using a threshold value, so that moving trajectories separated by a geodetic distance smaller than the detected discontinuous points become one class. Then, subclass candidates for the threshold are generated by clustering continuously distributed movement trajectories.
- the subclass candidate selection unit 2404 obtains an instruction about the number of classes, and extracts a plurality of region extraction candidates generated by the subclass candidate generation unit 2403 from the region extraction candidates divided into the number of regions close to the acquired number of classes.
- the candidate selected from the candidates is output as a result of subclassing the selected subclass candidate from the movement trajectory calculated by the movement trajectory calculation unit 102. That is, the clustering result closest to the predetermined number of classes is selected from the region extraction candidates for each of the threshold values generated by the subclass candidate generation unit 2403.
- the third distance calculation unit 2402 calculates the Euclidean distance f (i, j). Since this procedure is the same as that of the Euclidean distance calculation unit 1501 described in the above embodiment, the description thereof is omitted.
- the third distance calculation unit 2402 determines K threshold values R k for the obtained Euclidean distance f (i, j).
- the third distance calculation unit 2402 performs a non-linearization process on each determined threshold value R k and calculates g k (i, j) which is a geodetic distance with respect to the threshold value R k .
- the calculation procedure of the geodetic distance g k (i, j) corresponding to each threshold value R k is the same as the operation in the distance calculation unit 104 described in the above embodiment, and thus the description thereof is omitted.
- the subclass candidate generation unit 2403 generates a subclass candidate by detecting discontinuous points using the geodetic distance matrix g k (i, j) corresponding to each threshold value R k . Specifically, the subclass candidate generation unit 2403 sets a discontinuity point between the movement trajectory i and the movement trajectory j where g k (i, j) is infinite.
- FIGS. 30A to 30E show the movement trajectories a to h
- FIG. 30B shows a conceptual diagram of a multidimensional space consisting of the movement trajectories a to h.
- the threshold value R k is a sufficiently large value, for example, if the threshold value R k is larger than the maximum value of the Euclidean distance f (i, j), the geodetic distance g k (i, j) The combination of i and j is not infinite. That is, since there are no discontinuous points, it can be determined that there is one subclass as shown in FIG. 30C.
- the threshold R k is sufficiently small, specifically, when the threshold R k is smaller than the minimum value of f (i, j), g k (i, j ) Becomes infinite. That is, the number of subclasses is the same as the number of movement trajectories.
- FIGS. 30D and 30E show examples in which the threshold value thus determined is applied.
- the subclass candidate generation unit 2403 determines that the interval between the movement locus e and the movement locus f is a discontinuous point. As a result, the geodesic distance between each of the movement trajectories a to d and the movement trajectory e does not pass through discontinuous points, and thus does not take an infinite value. Geodesic distances from a to e for each movement locus are infinite because they pass through the discontinuous point g 1 (e, f).
- FIG. 30E shows an example in which another threshold is defined as R 2 (where R 1 > R 2 ). It is determined that there are discontinuous points between the movement locus c and the movement locus d, between the movement locus e and the movement locus f, and between the movement locus f and the movement locus g, respectively, as in FIG. 30D.
- a group in which the geodesic distance is infinite and a group in which the geodesic distance is not infinite are arranged and separated into four subclasses of ⁇ 1 , ⁇ 2 , ⁇ 3 , ⁇ 4 .
- the subclass candidate generating unit 2403 to a plurality of thresholds R k, a set of moving track geodesic distance is not infinite, it can be determined that the same subclass by a continuous.
- a plurality of subclass candidates can be generated based on the discontinuous points.
- the subclass candidate selection unit 2404 selects, as the final result, the subclass candidate that has the number of classes closest to the preset number from the subclass candidate generation result in the subclass candidate generation unit 2403.
- the subclass candidate (FIG. 30E) at the threshold R 2 is selected.
- the number of classes M 2 is set, a subclass candidate (FIG. 30D) at the time of the threshold R 1 is selected, and the subclass classification is executed so that each movement locus is closest to the designated number of subclasses.
- the number of classes M is 2 or 4 here, a larger number is actually desirable, and can be arbitrarily determined according to the situation and experience of the scene.
- subclass classification considering the pixel position and the similarity of motion is possible by using the Euclidean distance and the geodetic distance.
- the upper thigh and the lower thigh are separate subclasses.
- threshold values R k may be prepared. If there are no subclass candidates equal to the set number of classes, the closest subclass number may be selected, or the closest number of subclasses from among the smaller class numbers The subclass candidate that has become may be selected. Similarly, a subclass candidate having the closest subclass number may be selected from among the class numbers larger than the set class number.
- subclass candidate generation is performed based on discontinuous points calculated using geodetic distances, and subclass classification is performed by selecting a subclass candidate closest to the specified number of classes. Can do.
- the subclass classification unit 105 has been described as obtaining a subclass based on the Euclidean distance f (i, j) calculated in (Equation 14).
- the operation of the subclass classification unit 105 is not limited to this. That is, in the subclass classification unit 105, the Euclidean distance calculation unit 1501 calculates the Euclidean distance, but the distance calculation unit 104 also calculates the Euclidean distance. For this reason, in the fifth modification of the present embodiment, the distance calculation in the subclass classification unit is omitted, and the subclass classification process is performed using the distance calculated by the distance calculation unit 104.
- FIG. 31A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 5 of the embodiment.
- the moving object detection device 100D includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 2601, a weighted distance calculation unit 106, and An area dividing unit 107 is included.
- a subclass classification unit 2602 may be used instead of the subclass classification unit 2601 or a subclass classification unit 2603 may be used.
- processing units other than the subclass classification unit 2601, 2602, or 2603 are the same as those in the above embodiment, the description thereof is omitted.
- 31B to 31D show the configurations of the subclass classification units 2601 to 2603, respectively. These correspond to the embodiment, the third modification of the embodiment, and the fourth modification of the embodiment, respectively. Each will be described in turn.
- 31B includes a Euclidean distance load unit 2604 and a clustering unit 1502.
- the operation of the clustering unit 1502 is the same as that of the clustering unit 1502 shown in FIG.
- the Euclidean distance calculation unit 1201 included in the distance calculation unit 104 stores the calculated Euclidean distance f (i, j) in a memory (not shown) first, and the Euclidean distance included in the subclass classification unit 2601.
- the load unit 2604 loads the Euclidean distance f (i, j) stored in the memory from the memory.
- the clustering unit 1502 performs clustering of the movement trajectory using the Euclidean distance f (i, j) loaded from the memory by the Euclidean distance loading unit 2604. Thereby, the trouble of newly calculating the Euclidean distance can be saved, and higher speed processing can be realized.
- 31C includes a distance load unit 2605 and a clustering unit 2203.
- the operation of the clustering unit 2203 is the same as that of the clustering unit 2203 in Modification 3 of the embodiment shown in FIG.
- the Euclidean distance calculation unit 1201 and the movement locus geodetic distance calculation unit 1202 included in the distance calculation unit 104 store the calculated Euclidean distance f (i, j) and geodetic distance g (i, j) in advance.
- the distance load unit 2605 included in the subclass classification unit 2602 loads the Euclidean distance f (i, j) and geodetic distance g (i, j) stored in the memory from the memory. To do.
- the clustering unit 2203 clusters the movement trajectory using the Euclidean distance f (i, j) and the geodetic distance g (i, j) loaded from the memory by the distance loading unit 2605. Thereby, the trouble of newly calculating the Euclidean distance and the geodetic distance can be saved, and higher-speed processing can be realized.
- 31D includes a distance load unit 2605, a subclass candidate generation unit 2403, and a subclass candidate selection unit 2404.
- the operations of the subclass candidate generation unit 2403 and the subclass candidate selection unit 2404 are the same as those of the subclass candidate generation unit 2403 and the subclass candidate selection unit 2404 in the modification 4 of the embodiment shown in FIG. To do.
- the Euclidean distance calculation unit 1201 and the movement locus geodetic distance calculation unit 1202 included in the distance calculation unit 104 store the calculated Euclidean distance f (i, j) and geodetic distance g (i, j) in advance.
- the distance load unit 2605 included in the subclass classification unit 2603 loads the Euclidean distance f (i, j) and geodetic distance g (i, j) stored in the memory from the memory. To do.
- the subclass candidate generation unit 2403 performs movement trajectory clustering together with the subclass candidate selection unit 2404 using the Euclidean distance f (i, j) and the geodetic distance g (i, j) loaded from the memory by the distance load unit 2605. . Thereby, the trouble of newly calculating the Euclidean distance and the geodetic distance can be saved, and higher-speed processing can be realized.
- the geometric constraint estimation unit 701 included in the stationary index calculation unit 103 has been described as obtaining geometric constraints based on the movement trajectory obtained by the movement trajectory calculation unit 102.
- the stationary index calculation unit 103 may estimate the geometric constraint using a movement trajectory obtained separately from a moving image received as an input by the image input unit 101, in addition to the movement trajectory calculation unit 102.
- FIG. 32A is a diagram illustrating a configuration of a mobile object detection device according to Modification 6 of the embodiment.
- the moving body detection device 100E includes an image input unit 101, a movement locus calculation unit 102, a geometric constraint estimation movement locus calculation unit 2701, a stationary index calculation unit 2702, a distance calculation unit 104, and a subclass classification unit. 105, a weighted distance calculation unit 106, and an area dividing unit 107.
- FIG. 32B shows the configuration of the stationary index calculation unit 2702. This will be described below, including the flow from the geometric constraint estimation moving trajectory calculation unit 2701.
- the stability of the estimation and the estimation result depend on the accuracy of the corresponding points extracted from the movement locus. That is, more accurate geometric constraints can be obtained by using data with as little error as possible than data with much error.
- the inventors' experiments have also confirmed that the estimation results and stability are lowered when the corresponding point accuracy is poor.
- a sparse corresponding point acquisition method such as KLT or LK (Lucas-Kanade) has a demerit that it is sparse, but instead, it is possible to acquire corresponding points with high accuracy.
- the geometric constraint is estimated more accurately by performing only the estimation of the geometric constraint using the separately acquired sparse and highly accurate corresponding points.
- a sparse corresponding point acquisition method there is a KLT (Kanade-Lucas-Tomasi) tracker.
- the KLT tracker calculates the similarity between images based on the following (formula 52).
- p is a pixel position shown in the format of (u, v)
- z is a movement amount between pixels of corresponding points shown in the format of (u, v)
- e is an error component to be minimized.
- the amount of movement z is obtained by solving z that minimizes (Formula 52). Since the method for acquiring corresponding points is shown in more detail in Non-Patent Document 8, further detailed description is omitted. “An Implementation of the Kanade-Lucas-Tomasi Feature Tracker”, http: // www. ces. creson. edu / ⁇ stb / klt /, 2006
- the stationary index calculation unit 2702 includes a geometric constraint estimation unit 701 and an error calculation unit 702.
- the operations of the geometric constraint estimation unit 701 and the error calculation unit 702 are the same as those of the geometric constraint estimation unit 701 and the error calculation unit 702 shown in FIG.
- the geometric constraint estimation unit 701 of the stationary index calculation unit 2702 receives the movement trajectory calculated by the geometric constraint estimation movement trajectory calculation unit 2701 as an input, and estimates the geometric constraint. Subsequently, the error calculation unit 702 applies the geometric constraint estimated by the geometric constraint estimation unit 701 to the movement trajectory calculated by the movement trajectory calculation unit 102, and obtains a stationary index E for each movement trajectory. . The subsequent weighted distance calculation unit 106 and subsequent processes are executed using the stationary index obtained by the above procedure.
- the weighted distance calculation unit 106 applies the geodetic distance g (i, j) calculated by the distance calculation unit 104 and the stationary index E calculated by the stationary index calculation unit 103. Based on this, it has been explained that the geodesic distance between subclasses is obtained by weighting.
- the operations of the distance calculation unit 104 and the weighted distance calculation unit 106 are not limited to this. That is, the distance calculation unit 104 obtains the Euclidean distance between the movement trajectories, and the weighted distance calculation unit 106 obtains the inter-subclass Euclidean distance, and determines whether the subclass is stationary or moving based on the stationary index of the intra-subclass movement trajectory. Then, based on the determination result, the Euclidean distance between subclasses may be weighted, and finally, the geodesic distance between subclasses may be obtained from the Euclidean distance between subclasses.
- FIG. 33A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 7 of the embodiment.
- the moving object detection apparatus 100F includes an image input unit 101, a movement locus calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 2801, a subclass classification unit 105, a weighted distance calculation unit 2802, and An area dividing unit 107 is included.
- the distance calculation unit 2801 calculates the Euclidean distance between the movement trajectories.
- FIG. 33B shows the configuration of the distance calculation unit 2801.
- the distance calculation unit includes a Euclidean distance calculation unit 1501.
- the processing of the Euclidean distance calculation unit 1501 is the same as that of the Euclidean distance calculation unit 1501 described in FIG.
- the weighted distance calculation unit 2802 moves the arbitrary stationary object relative to the Euclidean distance between the movement trajectories of the arbitrary stationary object with respect to the Euclidean distance calculated by the distance calculation unit 2801 based on the stationary index as compared to before the weighting.
- a weighted Euclidean distance is calculated by assigning a weight that increases the ratio of the Euclidean distance between the locus and the moving locus of an arbitrary moving object, and a weighted geodetic distance is calculated from the calculated weighted Euclidean distance. .
- FIG. 33C shows the configuration of the weighted distance calculation unit 2802.
- the weighted distance calculation unit 2802 of this modification includes an inter-subclass Euclidean distance calculation unit 2803, a stationary index addition unit 2804, and an inter-subclass weighted geodetic distance calculation unit 2805.
- FIG. 34 shows two adjacent subclasses ⁇ i and ⁇ j among the plurality of classes generated by the subclass classification unit 105.
- “subclass” is expressed only as “class”.
- x i included in (Expression 21) and (Expression 22) is a movement trajectory expressed in the form of a multidimensional vector, as in (Expression 2).
- the distance obtained between the movement locus included in I and the movement locus included in J is defined as an interclass distance.
- distance is a concept including both Euclidean distance and geodetic distance.
- FIG. 34 shows a conceptual diagram of a representative value of the Euclidean distance between classes (representative Euclidean distance).
- f 31 f (i 3 , j 1)
- the moving object region detection is performed on a pixel-by-pixel basis by focusing on a collection of single movement trajectories in a class as shown in I and J and operating based on a macro distance in class units. Therefore, it is possible to cope with noise / false detection of a stationary index that occurs due to this. Therefore, it is desirable to calculate a representative value of the distance between classes. That is, as shown in FIG. 34, the representative value of the distance between classes is desirably a representative value that can approximate the movement or positional relationship between classes for a plurality of classes.
- the average value of the Euclidean distance between the movement trajectories of each class can be used as the representative value. This is obtained by obtaining a plurality of Euclidean distances corresponding to all combinations between movement trajectories included in each class and averaging them.
- the representative Euclidean distance F ( ⁇ i , ⁇ j ) can be calculated by the following (Formula 54).
- the representative Euclidean distance is not limited to the average value of the Euclidean distance.
- the median value of the Euclidean distance between the movement trajectories of each class can be used as a representative value. This is obtained by obtaining a plurality of Euclidean distances corresponding to all combinations between movement trajectories included in each class and taking the median among them.
- the representative Euclidean distance F ( ⁇ i , ⁇ j ) can be calculated by the following (Expression 55).
- the mode value of the Euclidean distance between the movement trajectories of each class can be used as the representative value as the representative Euclidean distance.
- These representative values are values that appear most frequently among a plurality of classes when a plurality of Euclidean distances corresponding to all combinations of movement trajectories included in each class are obtained.
- the stationary index adding unit 2804 weights the representative Euclidean distance F ( ⁇ i , ⁇ j ) between the subclasses based on the stationary index of the movement trajectory belonging to each class.
- the weighting procedure and criteria are the same as those described with reference to FIG. After weighting, the representative Euclidean distances between subclasses are close to each other, and the distance between the mobile and the background is long.
- the inter-subclass weighted geodesic distance calculation unit 2805 calculates the interclass subclass geodesic distance for the weighted representative Euclidean distance F ( ⁇ i , ⁇ j ).
- the procedure for obtaining the geodetic distance from the representative Euclidean distance is described in detail in the description of the distance calculation unit 104. That is, the same processing as that performed by the distance calculation unit 104 may be performed in the same way as the geodesic distance calculation unit 1202.
- the geodesic distance can be obtained by the same processing only by the difference between the processing unit of the movement trajectory or the subclass.
- the weighted distance calculation unit 2802 does not necessarily include the inter-subclass weighted geodesic distance calculation unit 2805.
- the geodetic distance is particularly suitable for expressing a mobile object such as a person whose deformation is severe, but depending on the degree of deformation of the mobile object in a moving image, detection may be possible only with the Euclidean distance without using the geodetic distance.
- the configuration of the weighted distance calculation unit 2806 in this case is shown in FIG. 33D.
- the weighted distance calculation unit 2806 moves the arbitrary stationary object with respect to the Euclidean distance between the movement trajectories of the arbitrary stationary object with respect to the Euclidean distance calculated by the distance calculation unit 2801 based on the stationary index as compared with before the weighting.
- a weighted Euclidean distance is calculated by assigning a weight such that the ratio of the Euclidean distance between the trajectory and the moving trajectory of an arbitrary moving body is increased.
- the weighted distance calculation unit 2806 of this modification includes an inter-subclass Euclidean distance calculation unit 2803 and a stationary index addition unit 2804.
- a subclass is generated by the subclass classification unit 105 by clustering a collection of movement trajectories calculated by the movement trajectory calculation unit 102 according to a certain index such as luminance or similarity of movement trajectories.
- the subclass does not necessarily need to include a plurality of movement trajectories. That is, you may perform the subclass classification
- the weighted distance calculation unit 106 applies the geodetic distance g (i, j) calculated by the distance calculation unit 104 and the stationary index E calculated by the stationary index calculation unit 103. Based on the above description, the inter-subclass geodesic distance is weighted.
- the operations of the stationary index calculation unit 103 and the weighted distance calculation unit 106 are not limited to this. That is, the stationary index calculation unit 103 may include a camera movement acquisition unit that acquires movement information of the camera 110 from the image, and output the camera movement detected from the image to the weighted distance calculation unit 106.
- the weighted distance calculation unit 106 determines whether the subclass is stationary or moving based on the interclass subclass geodesic distance and the stationary index of the movement trajectory within the subclass, and based on the determination result and the camera motion information, It is good also as what calculates
- FIG. 35A is a diagram illustrating a configuration of a moving object detection device according to Modification Example 8 of the embodiment.
- the moving object detection device 100G includes an image input unit 101, a movement trajectory calculation unit 102, a stationary index calculation unit 3001, a distance calculation unit 104, a subclass classification unit 105, a weighted distance calculation unit 3002, and An area dividing unit 107 is included.
- FIG. 35B shows the configuration of the stationary index calculation unit 3001.
- the stationary index calculation unit 3001 of this modification includes a camera motion acquisition unit 3003, a geometric constraint estimation unit 3004, and an error calculation unit 702.
- the operation of the error calculation unit 702 is the same as the operation of the error calculation unit 702 described with reference to FIG.
- the camera motion acquisition unit 3003 estimates camera motion information from image motion information. That is, the camera motion acquisition unit 3003 extracts corresponding points between the frames from the movement trajectory calculated by the movement trajectory calculation unit 102, and estimates camera motion information between the frames.
- a basic matrix F is obtained by an 8-point method and motion estimation is performed therefrom.
- the 8-point method in which 8 samples are selected from corresponding points and a matrix is estimated by RANSAC is known as a method often used for basic matrix estimation.
- this matrix is obtained from the estimated basic matrix and the camera calibration matrix, and the singular value decomposition of the E matrix is performed to determine the camera motion. Information can be estimated.
- Non-Patent Document 9 Since the above motion estimation method is described in detail in Non-Patent Document 4, further detailed description thereof is omitted. Of course, in addition to the above method, for example, as described in Non-Patent Document 9, motion estimation may be performed from corresponding points and a plane. In addition, any method for estimating camera motion information between frames from an image can be used for motion estimation in the present embodiment.
- Image Understanding-Mathematical Principles of 3D Recognition Kenichi Kanaya, Morikita Publishing, 1990
- the geometric constraint estimation unit 3004 can estimate the geometric constraint using the camera motion information estimated by the camera motion acquisition unit 3003. For example, in the epipolar constraint equations shown in (Equation 4) and (Equation 5), the basic matrix F estimated by the camera motion acquisition unit 3003 can be used as it is, and in the homography constraint or the structure matching constraint equation The homography matrix and epipole included in the projection depth used (Equation 8) can also be obtained from the camera motion information estimated by the camera motion acquisition unit 3003. Since detailed conversion is also detailed in Non-Patent Document 4, description thereof is omitted. By the above procedure, there is an effect that the process of the geometric constraint estimation unit 3004 is simplified.
- the error calculation unit 702 calculates the stationary index E of each movement locus.
- the error calculation unit 702 outputs the obtained stillness index E to the weighted distance calculation unit 3002, and the camera motion acquisition unit 3003 outputs the obtained camera motion information to the weighted distance calculation unit 3002.
- the weighted distance calculation unit 3002 calculates the inter-subclass geodesic distance in the same manner as the weighted distance calculation unit 106 in FIG. Up to this point, the weighted distance calculation unit 3002 and the weighted distance calculation unit 106 perform the same operation.
- FIGS. 36A and 36B show conceptual diagrams of the distribution of subclasses in a high-dimensional space when the camera motion is large and small. Although it is actually a high-dimensional space, it is displayed in two dimensions for ease of viewing.
- FIG. 36A shows a distribution when the camera motion is large
- FIG. 36B shows a distribution when the camera motion is small.
- the background motion component relatively increases and the background distribution spreads, so that the distance between the moving body and the background inevitably decreases as shown in FIG. 36A.
- the background is distributed at a position far from the moving body as shown in FIG. 36B.
- the above-described weighting rule is changed according to the magnitude of the camera motion.
- W B W B ⁇ W S ⁇ 1 (W B : Weight when camera movement is large, W S : Weight when camera movement is small) Subjected to subclass between geodesic distance weight W B or W S to be.
- W B Weight when camera movement is small
- W B Weight when camera movement is small
- W B or W S Subjected to subclass between geodesic distance weight W B or W S to be. That is, when the camera motion is large, weights are set such that the subclasses of the stationary object are closer to each other and the subclass of the moving object is separated from the subclass of the stationary object.
- the criterion for whether the camera motion is large or small depends on the moving speed of the moving object that is desired to be detected from the moving image. For example, if a person is to be detected as a moving object, the average movement speed of the person is 3 km / h.
- the camera movement is 3 km / h or more, it is determined that the camera movement is large, and the camera movement is more
- the larger the weight the smaller the weight W applied to the geodesic distance between moving trajectory pairs whose stationary determination value E MS representing the determination result of stationary or moving subclass is 0 (stationary) (W ⁇ 1), and the stationary determination value E MS
- the weight W applied to the geodesic distance between the moving trajectory pairs of 0 and 1 (stationary and moving) may be increased (W> 1).
- FIG. 36D and 36E show examples of results of changing the subclass geodesic distance by the weighting set as described above.
- FIG. 36D shows the distribution of subclasses after weighting the inter-subclass geodesic distance according to the weight when the camera motion is large
- FIG. 36E shows the weight of the inter-subclass geodesic distance according to the weight when the camera motion is small. The distribution of subclasses after doing is shown. It can be seen that the geodetic distance between the subclasses is obtained by appropriately changing the weighting in this way, regardless of the magnitude of the camera movement.
- the weighting rule can be determined by multiplying or dividing the weight W ref by the ratio of the absolute value of the camera motion to the weight W ref set in advance under a specific camera motion. It may be changed.
- the camera motion acquisition unit determines the camera motion information itself, but instead of the camera motion, the average value of the size of the movement locus (the amount of movement obtained from the movement locus) is obtained. It may be used instead of movement. As the camera movement increases, the size of the movement trajectory on the background increases on average, so that a value approximately proportional to the magnitude of the camera movement can be obtained. For example, by acquiring the relationship between the actual camera motion and the approximate average value of the moving trajectory in advance, the average value of the moving trajectory size can be used as an evaluation value corresponding to the camera motion. Can be used.
- the camera motion acquisition unit 3003 is included in the stationary index calculation unit 3001, and has been described as estimating camera motion information from an image.
- the operation of the camera motion acquisition unit 3003 is not limited to this.
- the camera motion acquisition unit 3003 may acquire the information of the camera motion electronically or physically by a sensor installed in the camera, or may acquire the information of the camera motion from the operation control signal of the camera. Good.
- FIG. 37 is a diagram illustrating a configuration of a moving object detection device according to Modification Example 9 of the embodiment.
- the moving body detection device 100H includes an image input unit 101, a movement locus calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 105, a weighted distance calculation unit 3102, and An area dividing unit 107 is included.
- processing units other than the camera motion acquisition unit 3101 and the weighted distance calculation unit 3102 are the same as those in the modified example 8, description thereof will be omitted.
- the camera motion acquisition unit 3101 acquires camera motion information electronically or physically from a sensor installed in the camera. Subsequently, the weighted distance calculation unit 3102 performs the camera motion in addition to the stationary index E from the stationary index calculation unit 103, the distance between the movement trajectories from the distance calculation unit 104, and the subclass classification information (label information) from the subclass classification unit 105.
- the camera movement information acquired by the camera movement acquisition unit 3101 is received from the acquisition unit 3101. Similar to the modified example 8, the weight W for the subclass geodesic distance is changed based on the camera motion information. The details of the processing are the same as in Modification 8 and will not be described. According to this configuration, the weighted distance calculation unit 3102 can acquire information on actual camera motion, and thus can detect a moving object more correctly.
- the acquisition of camera movement information by the camera movement acquisition unit 3101 does not necessarily have to be a sensor installed in the camera, but may be a sensor separately installed in a moving vehicle or the like in which the camera is installed.
- a sensor separately installed in a moving vehicle or the like in which the camera is installed.
- an in-vehicle sensor can be used. If the positional relationship between the camera and the separately installed sensor is known, the information on the moving body motion acquired by the sensor can be easily converted into the information on the camera motion and used by performing coordinate conversion.
- FIG. 38 shows a configuration diagram of the moving object detection device when the sensor is installed separately from the camera.
- the moving body detection device 100I includes a camera motion acquisition unit 3201 instead of the camera motion acquisition unit 3101 in the mobile body detection device 100H illustrated in FIG.
- a camera motion acquisition unit 3201 acquires sensor information from a moving body sensor 3202 provided on a moving vehicle such as a car, performs coordinate conversion to a camera coordinate system, and uses the motion information of the moving body as camera motion information. Just output.
- the moving body sensor 3202 used when the moving vehicle is a vehicle may be a sensor that detects a travel distance and a steering angle.
- the mounting destination of the camera is a moving vehicle.
- the moving destination is not limited to a vehicle as long as the camera can be moved and mounted and the camera motion can be detected.
- Modification 10 a moving object detection device according to Modification 10 of the embodiment of the present invention will be described with reference to FIGS. 39A and 39B.
- the weighted distance calculation unit 106 includes a representative geodetic distance calculation unit 1701 and a stationary index adding unit 1702, and weights the representative geodetic distance based on the stationary index.
- the area dividing unit 107 has been described as performing area division on the weighted representative geodetic distance.
- the operations of the weighted distance calculation unit 106 and the region division unit 107 are not limited to this.
- the weighted distance calculation unit 106 may obtain a representative geodetic distance and a representative stationary index of each subclass and output them to the region dividing unit, and the region dividing unit 107 may weight the clustering threshold based on the representative stationary index. .
- FIG. 39A is a diagram illustrating a configuration of a mobile object detection device according to Modification Example 10 of the embodiment.
- the moving body detection device 100J includes an image input unit 101, a movement locus calculation unit 102, a stationary index calculation unit 103, a distance calculation unit 104, a subclass classification unit 105, a weighted distance calculation unit 3401, and An area dividing unit 3402 is included.
- the weighted distance calculation unit 3401 calculates the geodetic distance between the movement trajectories based on the distance between the movement trajectories.
- the area dividing unit 3402 assigns a weight based on the stationary index to the geodesic distance threshold used to determine whether to classify the moving area of the stationary object and the moving area of the moving object into different classes.
- the area dividing unit 3402 uses a weight for the geodesic distance threshold when the stationary indices of the two movement trajectories are values representing “stationary object” and “stationary object”, respectively. It is larger than the weight for the threshold of the geodetic distance when the value represents the “body” and “stationary object”.
- the weighted distance calculation unit 3401 includes a representative geodetic distance calculation unit 1701 and a representative stationary index calculation unit 3403. Since the operation of the representative geodetic distance calculation unit 1701 is the same as that in the above embodiment, the description thereof is omitted.
- the representative stationary index calculation unit 3403 obtains a representative value of the stationary index of the movement trajectory belonging to each class.
- the movement trajectories in a certain subclass there are more movement trajectories determined as “still”, or the number of movement trajectories determined as “still” and the number of movement trajectories determined as “movement” are the same.
- the subclass between stillness determination value E SUB-MS 1
- one inter-subclass stillness determination value E SUB-MS that is, a value of evaluation of stillness or movement in units of subclasses is given between the subclasses.
- the weighted distance calculation unit 3401 outputs the inter-subclass stillness determination value E SUB-MS and the inter-subclass geodesic distance hp , q to the region dividing unit 3402.
- the area dividing unit 3402 uses the inter-subclass geodesic distance hp , q calculated by the weighted distance calculating unit 3401 as the evaluation value of the area division of the subclass ⁇ p , and uses the inter-subclass as the weighting of the threshold for the area division.
- E SUB-MS it is determined whether or not to divide the region division candidate of the subclass ⁇ p as a separate cluster.
- the area dividing unit 3402 has a sufficient distance between the two subclasses ⁇ p and ⁇ q. Select as a separate class and confirm as an individual class.
- the region dividing unit 3402 determines the corresponding two subclasses ⁇ p and ⁇ q as the same class. That is, in this case, it is decided not to divide. Then, after determining whether or not to divide all subclasses of region division candidates, the region division unit 3402 assigns different labels ⁇ m to the movement trajectories belonging to different classes, and the region division information of the movement trajectory Output as.
- the following weighting rule is set for the predetermined threshold value Ht.
- the inter-subclass stillness determination value E SUB-MS of two subclasses is 1, both are subclasses of the moving object.
- the weighting which makes it easy to integrate between subclasses when the two subclasses are both stationary objects, and the weighting which makes it difficult to integrate between subclasses when one of the two subclasses is the background and the other is a stationary object, as described above. May perform either one or both at the same time.
- the moving object can be more correctly separated from the background by changing the weight according to the reliability of the inter-subclass static determination value E SUB-MS of the subclass.
- the same effect as the weighting on the geodetic distance in the above embodiment can be obtained by weighting the threshold value for the region division between the subclasses.
- the area dividing unit 3402 uses the weighted threshold value, and similarly to the process of the area dividing unit 107 of the above embodiment, the inter-subclass geodesic distance hp , q and the weighting threshold value Htw (p, q) If h p, q ⁇ Htw (p, q), it is determined that the corresponding subclass ⁇ p , ⁇ q is divided, and conversely, h p, q ⁇ Htw (p, q) If there is, it is determined that the corresponding subclasses ⁇ p and ⁇ q are not divided, that is, integrated.
- the same effect as that of the mobile object detection apparatus 100 of the above embodiment can be obtained only by changing the threshold value without directly changing the value of the geodetic distance.
- the geodetic distance after weighting is not retained, that is, when the number of subclasses is large, it is not necessary to retain both the geodetic distance before weighting and the geodetic distance after weighting in the memory. There is.
- subclassification is performed based on the distance between pixels or the similarity of the movement trajectory, and the representative distance between subclasses
- the region segmentation of the picture including the moving object can be performed from the stationary index of the moving track regardless of the posture of the moving object.
- FIG. 40 is a diagram illustrating a configuration of a moving object detection device including components essential to the present invention.
- the moving object detection apparatus 100K includes a stationary index calculation unit 103, a distance calculation unit 104, a weighted distance calculation unit 106, and an area division unit 107. That is, when the movement trajectory of a picture in each of a plurality of blocks constituting a moving image is calculated in advance, the moving object detection apparatus 100 acquires such a movement trajectory from the outside, and acquires the acquired movement trajectory.
- the processes of steps S303, S304, S306, and S307 may be executed.
- the weighted distance calculation unit 106 may calculate the geodetic distance between all the movement trajectories without classifying the movement trajectory into subclasses.
- the weighted distance calculation unit 106 and the region division unit 107 are collectively referred to as a region detection unit.
- the present invention is realized as a moving body detection apparatus.
- image processing for extracting or dividing a region of an object having articulated motion in a moving image Needless to say, it can be realized as a device.
- the Euclidean distance is used as the linear distance.
- the Euclidean distance is not limited. A form obtained by using an arbitrary distance index defined by the linear distance is also included in the present invention.
- the system LSI is an ultra-multifunctional LSI manufactured by integrating a plurality of components on a single chip, and specifically, a computer system including a microprocessor, ROM, RAM, and the like. .
- a computer program is stored in the RAM.
- the system LSI achieves its functions by the microprocessor operating according to the computer program.
- each of the above-described devices may be configured from an IC card or a single module that can be attached to and detached from each device.
- the IC card or module is a computer system that includes a microprocessor, ROM, RAM, and the like.
- the IC card or the module may include the super multifunctional LSI described above.
- the IC card or the module achieves its function by the microprocessor operating according to the computer program. This IC card or this module may have tamper resistance.
- the present invention may be the method described above. Further, the present invention may be a computer program that realizes these methods by a computer, or may be a digital signal composed of the computer program.
- the present invention provides a non-volatile recording medium that can read the computer program or the digital signal, such as a flexible disk, a hard disk, a CD-ROM, an MO, a DVD, a DVD-ROM, a DVD-RAM, a BD (Blu-ray). -Ray Disc (registered trademark)), recorded in a semiconductor memory, or the like.
- the digital signal may be recorded on these non-volatile recording media.
- the computer program or the digital signal may be transmitted via an electric communication line, a wireless or wired communication line, a network represented by the Internet, data broadcasting, or the like.
- the present invention may also be a computer system including a microprocessor and a memory.
- the memory may store the computer program, and the microprocessor may operate according to the computer program.
- the present invention provides a moving object detection apparatus that detects a moving object in a picture by extracting a region including a moving object such as a person who moves while changing its shape based on movement in a plurality of pictures. Further, it can be used as a motion analysis device, a monitoring device, a moving body detection device incorporated in an AV device such as a video camera or a TV.
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Multimedia (AREA)
- Theoretical Computer Science (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Image Analysis (AREA)
Abstract
Description
P.Anandan,"A Computational Framework and an Algorithm for the Measurement of Visual Motion",International Journal of Computer Vision, Vol.2, pp.283-310,1989 Vladimir Kolmogorov and Ramin Zabih, "Computing Visual Correspondence with Occlusions via Graph Cuts", International Conference on Computer Vision,2001
Jianbo Shi and Carlo Tomasi "Good Features to Track", IEEE Conference on Computer Vision and Pattern Recognition,pp593-600,1994
Richard Hartley and Andrew Zisserman,"Multiple-View Geometry in Computer Vision",second ed.Cambridge Univ. Press, 2003 Chang Yuan,Gerard Medioni,Jinman Kang and Isaac Cohen,"Detecting Motion Regions in the Presence of a Strong Parallax from a Moving Camera by Multiview Geometric Constraints",IEEE Transactions On Pattern Analysis and Machine Intelligence,Vol.29,No.9,September 2007
E.W.Dijkstra,"A note on two problems in connexion with graphs",Numerische Mathematik,pp.269-271,1959
なお、本実施の形態では線形距離としてユークリッド距離を用いたが、前述した通り、ユークリッド距離に限るものではない。前述の線形距離で定義される任意の距離指標を用いて得られる形態も本発明に含まれるものとする。
次に、本発明の実施の形態の変形例1に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例2に係る移動体検出装置について説明する。
Pedro F. Felzenszwalb and Daniel P. Huttenlocher "Efficient Graph-Based Image Segmentation", International Journal of Computer Vision ,Vol.59,No.2,pp.167-181,Sept,2004
次に、本発明の実施の形態の変形例3に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例4に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例5に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例6に係る移動体検出装置について説明する。
"An Implementation of the Kanade-Lucas-Tomasi Feature Tracker",http://www.ces.clemson.edu/~stb/klt/,2006
次に、本発明の実施の形態の変形例7に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例8に係る移動体検出装置について説明する。
「画像理解-3次元認識の数理-」金谷健一著、森北出版、1990
WB<WS<1
(WB:カメラ動き大のときの重み、WS:カメラ動き小のときの重み)
となる重みWBまたはWSをサブクラス間測地距離にかける。サブクラスが「静止」と「移動」であった場合(異なる分類)、
WB>WS>1
(WB:カメラ動き大のときの重み、WS:カメラ動き小のときの重み)
となる重みWBまたはWSをサブクラス間測地距離にかける。つまり、カメラ動きが大きい場合に、より静止物のサブクラス同士が近づき、より移動体のサブクラスと静止物のサブクラスが離れるような重みを設定する。
次に、本発明の実施の形態の変形例9に係る移動体検出装置について説明する。
次に、本発明の実施の形態の変形例10に係る移動体検出装置について、図39Aおよび図39Bを参照しながら説明する。
Htw(p,q)=Ymin・Ht
となる。
Htw(p,q)=Ymax・Ht
となる。
Htw(p,q)=Yneu・Ht
となる。
101 画像入力部
102 移動軌跡算出部
103、103A、2702、3001 静止指標算出部
104 距離算出部
105、2101、2201、2401、2601、2602、2603 サブクラス分類部
106、106A、2802、2806、3002、3102、3401 重み付き距離算出部出部
107、3402 領域分割部
110 カメラ
120 ディスプレイ
200 コンピュータ
201 I/F
202 CPU
203 ROM
204 RAM
205 HDD
206 ビデオカード
701、3004 幾何拘束推定部
702 誤差算出部
703 静止指標画像作成部
1201、1501、2801 ユークリッド距離算出部
1202 移動軌跡間測地距離算出部
1502、2102、2203 クラスタリング部
1701 代表測地距離算出部
1702、1702A、2804 静止指標付加部
1703 閾値入力部
2202 第2距離算出部
2402 第3距離算出部
2403 サブクラス候補生成部
2404 サブクラス候補選択部
2604 ユークリッド距離ロード部
2605 距離ロード部
2701 幾何拘束推定用移動軌跡算出部
2803 サブクラス間ユークリッド距離算出部
2805 サブクラス間重み付き測地距離算出部
3003、3101、3201 カメラ動き取得部
3202 移動体上センサ
3403 代表静止指標算出部
Claims (32)
- 各々が動画像中の各領域に対応する複数の移動軌跡から、移動体領域を検出する移動体検出装置であって、
動画像を構成する2枚以上のピクチャ間における前記ピクチャを構成する1個以上の画素からなるブロックの動きの軌跡である複数の移動軌跡の各々について、当該移動軌跡の静止物の移動軌跡らしさを表す静止指標を算出する静止指標算出部と、
前記移動軌跡間の類似度を表す距離を算出する距離算出部と、
前記移動軌跡の静止指標および前記移動軌跡間の距離に基づいて、任意の静止物の移動軌跡間の距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の距離の比が、変換前と比べて大きくなるような変換処理を行ない、前記変換した移動軌跡間の距離に基づいて、静止物の移動軌跡と移動体の移動軌跡とを分離することにより、移動体の移動軌跡に対応する移動体領域を検出する領域検出部と
を備える移動体検出装置。 - 前記距離算出部は、前記移動軌跡間の距離に基づいて、2つの移動軌跡以外の移動軌跡を中継点として前記2つの移動軌跡の一方から他方にたどりつく経路の距離である測地距離を算出し、
前記領域検出部は、前記移動軌跡の静止指標および前記移動軌跡間の測地距離に基づいて、任意の静止物の移動軌跡間の測地距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の測地距離の比が、変換前と比べて大きくなるような変換処理を行ない、前記移動軌跡間の測地距離に基づいて、静止物の移動軌跡と移動体の移動軌跡とを分離することにより、移動体の移動軌跡に対応する移動体領域を検出する
請求項1記載の移動体検出装置。 - 前記静止指標算出部は、前記複数の移動軌跡から各移動軌跡が静止物の移動軌跡である場合に成立する幾何拘束を推定し、推定した前記幾何拘束を満たす度合いを前記静止指標として算出する
請求項1記載の移動体検出装置。 - 前記静止指標算出部は、複数の移動軌跡からエピポーラ拘束、ホモグラフィ拘束、三重線形拘束および構造一致性拘束のいずれかの幾何拘束を推定し、推定した前記幾何拘束を満たす度合いを前記静止指標として算出する
請求項3記載の移動体検出装置。 - 前記領域検出部は、
前記移動軌跡間の距離に基づいて、前記移動軌跡間の測地距離を算出し、前記移動軌跡の静止指標に基づいて、算出した各測地距離に対して、重み付け前に比べ、任意の静止物の移動軌跡間の測地距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の測地距離の比が大きくなるような重みを付けることにより、重み付き測地距離を算出する重み付き距離算出部と、
前記重み付き距離算出部が算出した前記重み付き測地距離に基づいて、前記重み付き測地距離が所定の閾値以上となる移動軌跡同士を異なるクラスに分類することにより、各ピクチャ上の静止物の領域と移動体の領域を分割する領域分割部とを含む
請求項1~4のいずれか1項に記載の移動体検出装置。 - 前記距離算出部は、移動軌跡間の類似度を表す測地距離を算出し、
前記重み付き距離算出部は、前記静止指標に基づいて、前記距離算出部が算出した前記測地距離に対して、重み付け前に比べ、任意の静止物の移動軌跡間の測地距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の測地距離の比が大きくなるような重みを付けることにより、前記重み付き測地距離を算出する
請求項5記載の移動体検出装置。 - 前記距離算出部は、移動軌跡間の類似度を表す線形距離を算出し、
前記重み付き距離算出部は、前記静止指標に基づいて、前記距離算出部が算出した前記線形距離に対して、重み付け前に比べ、任意の静止物の移動軌跡間の線形距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の線形距離の比が大きくなるような重みを付けることにより、重み付き線形距離を算出し、算出した前記重み付き線形距離から前記重み付き測地距離を算出する
請求項5記載の移動体検出装置。 - 前記領域検出部は、
前記移動軌跡間の距離に基づいて、前記移動軌跡間の測地距離を算出する重み付き距離算出部と、
静止物の移動領域と移動体の移動領域を異なるクラスに分類するか否かを判断するために用いられる前記測地距離の閾値に対して、前記静止指標に基づく重みを付けることにより得られる重み付き閾値と前記重み付き距離算出部が算出した前記測地距離との比較結果から静止物の移動軌跡と移動体の移動軌跡とを分離することにより、各ピクチャ上の静止物の領域と移動体の領域を分割する領域分割部とを含む
請求項1~4のいずれか1項に記載の移動体検出装置。 - 前記領域分割部は、2つの移動軌跡の静止指標がそれぞれ「静止物」および「静止物」を表す値であるときの前記測地距離の閾値に対する重みを、2つの移動軌跡の静止指標がそれぞれ「移動体」および「静止物」を表す値であるときの前記測地距離の閾値に対する重みよりも大きくする
請求項8記載の移動体検出装置。 - さらに、前記複数の移動軌跡を、各々が類似する移動軌跡の部分集合である複数のサブクラスに分類するサブクラス分類部を備え、
前記重み付き距離算出部は、移動軌跡の静止指標、移動軌跡間の距離、およびサブクラスの分類結果に基づいて、前記静止指標に基づく前記重みが付けられた前記移動軌跡間の測地距離を算出し、
前記領域分割部は、前記重み付き距離算出部が算出した前記重み付けされた測地距離に基づいて、静止物の移動軌跡と移動体の移動軌跡とを分離することにより、各ピクチャ上の静止物の領域と移動体の領域を分割する
請求項5記載の移動体検出装置。 - さらに、前記複数の移動軌跡を、各々が類似する移動軌跡の部分集合である複数のサブクラスに分類するサブクラス分類部を備え、
前記重み付き距離算出部は、移動軌跡の静止指標、移動軌跡間の距離、およびサブクラスの分類結果に基づいて、前記静止指標に基づく前記重みが付けられたサブクラス間の測地距離を算出し、
前記領域分割部は、前記重み付き距離算出部が算出した前記サブクラス間の測地距離に基づいて、静止物のサブクラスと移動体のサブクラスとを分離することにより、各ピクチャ上の静止物の領域と移動体の領域を分割する
請求項5記載の移動体検出装置。 - 前記距離算出部は、移動軌跡間の類似度を表す測地距離を算出し、
前記重み付き距離算出部は、サブクラス間の測地距離の代表値に対して、サブクラスに含まれる移動軌跡の静止指標の代表値に基づく重みをかけることにより、前記サブクラス間の測地距離を算出する
請求項11記載の移動体検出装置。 - 前記距離算出部は、移動軌跡間の類似度を表す線形距離を算出し、
前記重み付き距離算出部は、サブクラス間の線形距離の代表値に対して、サブクラスに含まれる移動軌跡の静止指標の代表値に基づく重みをかけ、重み付けされたサブクラス間の線形距離の代表値に基づいて、前記サブクラス間の測地距離を算出する
請求項11記載の移動体検出装置。 - 前記重み付き距離算出部は、2つのサブクラス間で、前記2つのサブクラスに含まれる移動軌跡の静止指標の代表値がそれぞれ「移動体」および「静止物」を表す値であったときに、前記サブクラス間の測地距離の前記重みを、1よりも大きい値に設定する
請求項11記載の移動体検出装置。 - 前記重み付き距離算出部は、2つのサブクラス間で、前記2つのサブクラスに含まれる移動軌跡の静止指標の代表値がそれぞれ「静止物」および「静止物」を表す値であったときに、前記サブクラス間の測地距離の前記重みを、1未満の値に設定する
請求項11または14に記載の移動体検出装置。 - 前記重み付き距離算出部は、2つのサブクラス間で、前記2つのサブクラスに含まれる移動軌跡の静止指標の代表値がそれぞれ「移動体」および「移動体」を表す値であったときに、前記サブクラス間の測地距離の前記重みを、1に設定する
請求項14または15に記載の移動体検出装置。 - 前記サブクラス分類部は、移動軌跡間の類似度に基づいて、各移動軌跡を前記複数のサブクラスのいずれかに分類する
請求項10または11記載の移動体検出装置。 - 前記サブクラス分類部は、移動軌跡間での、各移動軌跡に属するブロック同士の輝度の類似度に基づいて、各移動軌跡を前記複数のサブクラスのいずれかに分類する
請求項10または11記載の移動体検出装置。 - 前記サブクラス分類部は、
前記移動軌跡間の測地距離を算出する第2距離算出部と、
前記第2距離算出部が算出した前記移動軌跡間の測地距離の次元圧縮を行い、次元圧縮された前記移動軌跡間の測地距離に基づいて、各移動軌跡を前記複数のサブクラスのいずれかに分類するクラスタリング部とを含む
請求項10または11記載の移動体検出装置。 - 前記サブクラス分類部は、前記複数の移動軌跡のそれぞれについて、当該移動軌跡から他の移動軌跡までの複数の距離のうち、前記所定の距離閾値以下の距離を選択し、選択しなかった距離を無限大に変更する非線形化をした後に、当該移動軌跡から他の移動軌跡までの最短経路を求めることにより、前記測地距離を算出し、移動軌跡間の測地距離が有限の値となる移動軌跡の集まりを同一のサブクラスに分類することにより、各移動軌跡を前記複数のサブクラスのいずれかに分類する
請求項10または11記載の移動体検出装置。 - 前記静止指標算出部は、静止物の移動軌跡に対して成立する幾何拘束を推定するための移動軌跡から前記幾何拘束を推定し、推定した前記幾何拘束に基づいて、前記距離算出部において距離を算出するのに用いられる各移動軌跡の静止指標を算出する
請求項1~20のいずれか1項に記載の移動体検出装置。 - さらに、前記動画像を撮影するカメラの動き情報を取得するカメラ動き取得部を備え、
前記重み付き距離算出部は、前記カメラの動き情報の大きさに基づいて、重み付けの際の静止指標の重みを変化させる
請求項5記載の移動体検出装置。 - 前記重み付き距離算出部は、前記カメラの動き情報が大きいほど、静止物の移動軌跡と移動体の移動軌跡との間の距離の重み付けを大きくする
請求項22記載の移動体検出装置。 - さらに、前記動画像を撮影するカメラの動き情報を取得するカメラ動き取得部を備え、
前記重み付き距離算出部は、前記2つのサブクラスの静止指標の代表値がそれぞれ「移動体」および「静止物」を表す値であったときに、前記2つのサブクラス間の測地距離に重み付けを行い、
前記カメラの動き情報が所定の閾値以上の場合の重みをWBとし、前記カメラの動き情報が前記所定の閾値よりも小さい場合の重みをWSとした場合に、WB>WS>1の関係を満たす
請求項11記載の移動体検出装置。 - さらに、前記動画像を撮影するカメラの動き情報を取得するカメラ動き取得部を備え、
前記重み付き距離算出部は、前記2つのサブクラスの静止指標の代表値がそれぞれ「静止物」および「静止物」を表す値であったときに、前記2つのサブクラス間の測地距離に重み付けを行い、
前記カメラの動き情報が所定の閾値以上の場合の重みをWBとし、前記カメラの動き情報が前記所定の閾値よりも小さい場合の重みをWSとした場合に、WB<WS<1の関係を満たす
請求項11記載の移動体検出装置。 - 前記カメラ動き取得部は、カメラに対する操作制御信号から前記カメラの動き情報を取得する
請求項22~25のいずれか1項に記載の移動体検出装置。 - 前記カメラ動き取得部は、車載センサから前記カメラの動き情報を取得する
請求項22~25のいずれか1項に記載の移動体検出装置。 - 前記重み付き距離算出部は、さらに、前記移動軌跡の静止指標と静止指標閾値とを比較することにより、前記静止指標閾値以下の静止指標を有する前記移動軌跡を静止物の移動軌跡と判断し、前記静止指標閾値よりも大きい静止指標を有する前記移動軌跡を移動体の移動軌跡と判断する
請求項5~20および22~27のいずれか1項に記載の移動体検出装置。 - 前記重み付き距離算出部は、静止指標閾値を受け付ける閾値入力部を含み、前記移動軌跡の静止指標と前記閾値入力部が受け付けた前記静止指標閾値とを比較することにより、前記静止指標閾値以下の静止指標を有する前記移動軌跡を静止物の移動軌跡と判断し、前記静止指標閾値よりも大きい静止指標を有する前記移動軌跡を移動体の移動軌跡と判断し、
前記領域検出部は、さらに、検出した前記移動体領域を前記表示部に表示させる
請求項28に記載の移動体検出装置。 - 前記静止指標算出部は、さらに、前記ピクチャを構成する各ブロックを、算出した前記静止指標の値に応じた表示態様で、表示部に表示させる
請求項1~29のいずれか1項に記載の移動体検出装置。 - 各々が動画像中の各領域に対応する複数の移動軌跡から、移動体領域を検出する移動体検出方法であって、
動画像を構成する2枚以上のピクチャ間における前記ピクチャを構成する1個以上の画素からなるブロックの動きの軌跡である複数の移動軌跡の各々について、当該移動軌跡の静止物の移動軌跡らしさを表す静止指標を算出する静止指標算出ステップと、
前記移動軌跡間の類似度を表す距離を算出する距離算出ステップと、
前記移動軌跡の静止指標および前記移動軌跡間の距離に基づいて、任意の静止物の移動軌跡間の距離に対する任意の静止物の移動軌跡と任意の移動体の移動軌跡との間の距離の比が、変換前と比べて大きくなるような変換処理を行ない、前記移動軌跡間の距離に基づいて、静止物の移動軌跡と移動体の移動軌跡とを分離することにより、移動体の移動軌跡に対応する移動体領域を検出する領域検出ステップと
を含む、
移動体検出方法。 - 各々が動画像中の各領域に対応する複数の移動軌跡から、移動体領域を検出するためのプログラムであって、
請求項31に記載の移動体検出方法に含まれるステップをコンピュータに実行させるためのプログラム。
Priority Applications (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201180008077.8A CN102741884B (zh) | 2010-07-27 | 2011-07-25 | 移动体检测装置及移动体检测方法 |
JP2012501494A JP4979840B2 (ja) | 2010-07-27 | 2011-07-25 | 移動体検出装置および移動体検出方法 |
US13/454,577 US8599252B2 (en) | 2010-07-27 | 2012-04-24 | Moving object detection apparatus and moving object detection method |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2010-168704 | 2010-07-27 | ||
JP2010168704 | 2010-07-27 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US13/454,577 Continuation US8599252B2 (en) | 2010-07-27 | 2012-04-24 | Moving object detection apparatus and moving object detection method |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2012014430A1 true WO2012014430A1 (ja) | 2012-02-02 |
Family
ID=45529663
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2011/004173 WO2012014430A1 (ja) | 2010-07-27 | 2011-07-25 | 移動体検出装置および移動体検出方法 |
Country Status (4)
Country | Link |
---|---|
US (1) | US8599252B2 (ja) |
JP (1) | JP4979840B2 (ja) |
CN (1) | CN102741884B (ja) |
WO (1) | WO2012014430A1 (ja) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2014241012A (ja) * | 2013-06-11 | 2014-12-25 | 株式会社メガチップス | 動体検出装置、動体検出方法及び制御プログラム |
US9236090B2 (en) | 2012-07-20 | 2016-01-12 | Panasonic Intellectual Property Management Co., Ltd. | Video generating apparatus and video generating method |
JP2016505181A (ja) * | 2013-02-04 | 2016-02-18 | ハーマン インターナショナル インダストリーズ インコーポレイテッド | 移動物体を検出するための方法及びシステム |
WO2021166604A1 (ja) * | 2020-02-17 | 2021-08-26 | Necソリューションイノベータ株式会社 | 抽出装置 |
JP2022534556A (ja) * | 2019-05-30 | 2022-08-02 | モービルアイ ビジョン テクノロジーズ リミテッド | カメラを使用した車両環境モデル化 |
US11973921B2 (en) | 2020-05-21 | 2024-04-30 | Subaru Corporation | Image processing apparatus |
Families Citing this family (57)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
GB2488445A (en) * | 2009-08-06 | 2012-08-29 | Toshiba Res Europ Ltd | Correlated probabilistic trajectories pedestrian motion detection using a decision forest |
JP5475138B2 (ja) * | 2010-09-08 | 2014-04-16 | トヨタ自動車株式会社 | 移動物予測装置、仮想可動物予測装置、プログラム、移動物予測方法、及び仮想可動物予測方法 |
US8855406B2 (en) * | 2010-09-10 | 2014-10-07 | Honda Motor Co., Ltd. | Egomotion using assorted features |
AU2011265430B2 (en) * | 2011-12-21 | 2015-03-19 | Canon Kabushiki Kaisha | 3D reconstruction of partially unobserved trajectory |
US9567097B2 (en) * | 2012-02-03 | 2017-02-14 | Rosemount Aerospace Inc. | System and method for real-time aircraft performance monitoring |
JP5648647B2 (ja) * | 2012-03-21 | 2015-01-07 | カシオ計算機株式会社 | 画像処理装置、画像処理方法及びプログラム |
JP6013884B2 (ja) * | 2012-11-08 | 2016-10-25 | 日立オートモティブシステムズ株式会社 | 物体検出装置及び物体検出方法 |
US9224210B2 (en) * | 2013-02-06 | 2015-12-29 | University Of Virginia Patent Foundation | Systems and methods for accelerated dynamic magnetic resonance imaging |
US9165208B1 (en) * | 2013-03-13 | 2015-10-20 | Hrl Laboratories, Llc | Robust ground-plane homography estimation using adaptive feature selection |
CN104063740B (zh) * | 2013-03-21 | 2017-11-17 | 日电(中国)有限公司 | 办公室实体组识别系统、方法及装置 |
US9129161B2 (en) * | 2013-05-31 | 2015-09-08 | Toyota Jidosha Kabushiki Kaisha | Computationally efficient scene classification |
JP5886809B2 (ja) * | 2013-09-27 | 2016-03-16 | 富士重工業株式会社 | 車外環境認識装置 |
US9336436B1 (en) * | 2013-09-30 | 2016-05-10 | Google Inc. | Methods and systems for pedestrian avoidance |
JP6261266B2 (ja) * | 2013-10-02 | 2018-01-17 | 東芝アルパイン・オートモティブテクノロジー株式会社 | 移動体検知装置 |
KR101511450B1 (ko) * | 2013-12-09 | 2015-04-10 | 포항공과대학교 산학협력단 | 산등성이 데이터를 이용한 객체의 움직임 추적 장치 및 방법 |
CN104899894B (zh) * | 2014-03-05 | 2017-09-01 | 南京理工大学 | 一种采用多台摄像机进行运动目标跟踪的方法 |
GB2537780B (en) * | 2014-03-11 | 2019-04-17 | Mitsubishi Electric Corp | Person detecting device and person detecting method |
US9626802B2 (en) | 2014-05-01 | 2017-04-18 | Microsoft Technology Licensing, Llc | Determining coordinate frames in a dynamic environment |
US9761008B2 (en) * | 2014-05-08 | 2017-09-12 | The Trustees Of The University Of Pennsylvania | Methods, systems, and computer readable media for visual odometry using rigid structures identified by antipodal transform |
DE102014209137B4 (de) | 2014-05-14 | 2023-02-02 | Volkswagen Aktiengesellschaft | Verfahren und Vorrichtung zur Kalibrierung eines Kamerasystems eines Kraftfahrzeugs |
KR102051142B1 (ko) * | 2014-06-13 | 2019-12-02 | 현대모비스 주식회사 | 차량용 운전자 위험 지수 관리 시스템 및 그 방법 |
US10091411B2 (en) * | 2014-06-17 | 2018-10-02 | Lg Electronics Inc. | Mobile terminal and controlling method thereof for continuously tracking object included in video |
JP2016046642A (ja) * | 2014-08-21 | 2016-04-04 | キヤノン株式会社 | 情報処理システム、情報処理方法及びプログラム |
CN105518702B (zh) * | 2014-11-12 | 2018-06-26 | 深圳市大疆创新科技有限公司 | 一种对目标物体的检测方法、检测装置以及机器人 |
US9855890B2 (en) * | 2014-12-11 | 2018-01-02 | Toyota Motor Engineering & Manufacturing North America, Inc. | Autonomous vehicle interaction with external environment |
US9536320B1 (en) * | 2014-12-23 | 2017-01-03 | John H. Prince | Multiple coordinated detectors for examination and ranging |
JP6497987B2 (ja) * | 2015-03-12 | 2019-04-10 | キヤノン株式会社 | 画像処理装置及び画像処理方法、プログラム、記憶媒体 |
US20170109586A1 (en) * | 2015-10-16 | 2017-04-20 | Canary Connect, Inc. | Sensitivity adjustment for computer-vision triggered notifications |
CN105469361B (zh) * | 2015-12-24 | 2018-12-04 | 努比亚技术有限公司 | 一种局部图像处理方法和装置 |
JP6975945B2 (ja) * | 2016-02-24 | 2021-12-01 | パナソニックIpマネジメント株式会社 | 判定装置、判定方法、プログラムおよびプログラムを記録した記録媒体 |
TWI586936B (zh) * | 2016-05-20 | 2017-06-11 | 國立交通大學 | 實體影像與虛擬影像之間轉換的方法及其系統 |
RU2657308C2 (ru) * | 2016-06-03 | 2018-06-13 | Федеральное государственное бюджетное военное образовательное учреждение высшего образования "Военно-космическая академия имени А.Ф. Можайского" Министерства обороны Российской Федерации | Способ определения направления на источник оптического излучения по рассеянной в атмосфере составляющей |
CN107798272B (zh) * | 2016-08-30 | 2021-11-02 | 佳能株式会社 | 快速多目标检测与跟踪系统 |
AU2016421610B2 (en) * | 2016-09-05 | 2020-03-19 | Umitron Pte. Ltd. | Feeding system and feeding method |
CN106529482A (zh) * | 2016-11-14 | 2017-03-22 | 叶瀚礼 | 一种采用集合距离的交通路标识别方法 |
DE102016224095A1 (de) * | 2016-12-05 | 2018-06-07 | Robert Bosch Gmbh | Verfahren zum Kalibrieren einer Kamera und Kalibriersystem |
CN106515578A (zh) * | 2016-12-22 | 2017-03-22 | 百度在线网络技术(北京)有限公司 | 无人车驾驶状态的识别装置、识别方法及无人车 |
US20230107110A1 (en) * | 2017-04-10 | 2023-04-06 | Eys3D Microelectronics, Co. | Depth processing system and operational method thereof |
JP6822906B2 (ja) * | 2017-06-23 | 2021-01-27 | 株式会社東芝 | 変換行列算出装置、位置推定装置、変換行列算出方法および位置推定方法 |
CN110473232B (zh) * | 2017-07-14 | 2024-02-09 | 腾讯科技(深圳)有限公司 | 图像识别方法、装置、存储介质及电子设备 |
WO2019089591A1 (en) * | 2017-10-30 | 2019-05-09 | Mobileye Vision Technologies Ltd. | Vehicle navigation based on human activity |
US10460512B2 (en) * | 2017-11-07 | 2019-10-29 | Microsoft Technology Licensing, Llc | 3D skeletonization using truncated epipolar lines |
JP6587006B2 (ja) * | 2018-03-14 | 2019-10-09 | エスゼット ディージェイアイ テクノロジー カンパニー リミテッドSz Dji Technology Co.,Ltd | 動体検出装置、制御装置、移動体、動体検出方法、及びプログラム |
CN110299063B (zh) * | 2018-03-22 | 2022-01-25 | 阿里巴巴集团控股有限公司 | 轨迹数据的可视化展示方法及装置 |
US10896531B1 (en) * | 2018-05-18 | 2021-01-19 | Tableau Software, Inc. | Using visual distortion to emphasize targeted regions of data visualizations according to geodesic curves |
CN109255315B (zh) * | 2018-08-30 | 2021-04-06 | 跨越速运集团有限公司 | 一种出车途中人车分离判断方法及装置 |
US11270132B2 (en) * | 2018-10-26 | 2022-03-08 | Cartica Ai Ltd | Vehicle to vehicle communication and signatures |
US10817733B2 (en) * | 2019-02-13 | 2020-10-27 | Sap Se | Blind spot implementation in neural networks |
US10984546B2 (en) * | 2019-02-28 | 2021-04-20 | Apple Inc. | Enabling automatic measurements |
JP2020201674A (ja) * | 2019-06-07 | 2020-12-17 | キヤノン株式会社 | 映像解析装置及びその制御方法及びプログラム |
TWI730509B (zh) * | 2019-11-22 | 2021-06-11 | 英業達股份有限公司 | 影像偵測區域取得方法及空間使用情況的判定方法 |
CN111009012B (zh) * | 2019-11-29 | 2023-07-28 | 四川沃洛佳科技有限公司 | 一种基于计算机视觉的无人机测速方法、存储介质及终端 |
JP7484574B2 (ja) * | 2020-08-26 | 2024-05-16 | トヨタ自動車株式会社 | 自律移動ロボット制御システム、その制御方法、その制御プログラム及び自律移動ロボット制御装置 |
US11670144B2 (en) | 2020-09-14 | 2023-06-06 | Apple Inc. | User interfaces for indicating distance |
US11854224B2 (en) | 2021-07-23 | 2023-12-26 | Disney Enterprises, Inc. | Three-dimensional skeleton mapping |
CN115578862B (zh) * | 2022-10-26 | 2023-09-19 | 中国建设基础设施有限公司 | 交通流量转换方法、装置、计算设备和存储介质 |
CN116972831B (zh) * | 2023-09-25 | 2024-02-02 | 山东亚历山大智能科技有限公司 | 一种基于显著特征的动态场景移动机器人定位方法及系统 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2008146185A (ja) * | 2006-12-06 | 2008-06-26 | Univ Waseda | 移動物体検出装置及び移動物体検出装置用プログラム |
WO2010050110A1 (ja) * | 2008-10-27 | 2010-05-06 | パナソニック株式会社 | 移動体検出方法及び移動体検出装置 |
WO2010079556A1 (ja) * | 2009-01-09 | 2010-07-15 | パナソニック株式会社 | 移動体検出方法および移動体検出装置 |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP3780017B2 (ja) * | 1994-12-06 | 2006-05-31 | オリンパス株式会社 | 時系列画像解析装置及びその解析方法 |
JP4477439B2 (ja) * | 2004-07-12 | 2010-06-09 | 日立ソフトウエアエンジニアリング株式会社 | 画像分割処理システム |
JP5054206B2 (ja) * | 2008-03-14 | 2012-10-24 | パナソニック株式会社 | 画像処理方法及び画像処理装置 |
-
2011
- 2011-07-25 WO PCT/JP2011/004173 patent/WO2012014430A1/ja active Application Filing
- 2011-07-25 CN CN201180008077.8A patent/CN102741884B/zh active Active
- 2011-07-25 JP JP2012501494A patent/JP4979840B2/ja active Active
-
2012
- 2012-04-24 US US13/454,577 patent/US8599252B2/en active Active
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2008146185A (ja) * | 2006-12-06 | 2008-06-26 | Univ Waseda | 移動物体検出装置及び移動物体検出装置用プログラム |
WO2010050110A1 (ja) * | 2008-10-27 | 2010-05-06 | パナソニック株式会社 | 移動体検出方法及び移動体検出装置 |
WO2010079556A1 (ja) * | 2009-01-09 | 2010-07-15 | パナソニック株式会社 | 移動体検出方法および移動体検出装置 |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US9236090B2 (en) | 2012-07-20 | 2016-01-12 | Panasonic Intellectual Property Management Co., Ltd. | Video generating apparatus and video generating method |
JP2016505181A (ja) * | 2013-02-04 | 2016-02-18 | ハーマン インターナショナル インダストリーズ インコーポレイテッド | 移動物体を検出するための方法及びシステム |
US9852341B2 (en) | 2013-02-04 | 2017-12-26 | Harman International Industries, Incorporation | Method and system for detecting moving objects |
JP2014241012A (ja) * | 2013-06-11 | 2014-12-25 | 株式会社メガチップス | 動体検出装置、動体検出方法及び制御プログラム |
JP2022534556A (ja) * | 2019-05-30 | 2022-08-02 | モービルアイ ビジョン テクノロジーズ リミテッド | カメラを使用した車両環境モデル化 |
US11803981B2 (en) | 2019-05-30 | 2023-10-31 | Mobileye Vision Technologies Ltd. | Vehicle environment modeling with cameras |
JP7383870B2 (ja) | 2019-05-30 | 2023-11-21 | モービルアイ ビジョン テクノロジーズ リミテッド | デバイス、方法、システムおよびコンピュータプログラム |
WO2021166604A1 (ja) * | 2020-02-17 | 2021-08-26 | Necソリューションイノベータ株式会社 | 抽出装置 |
JPWO2021166604A1 (ja) * | 2020-02-17 | 2021-08-26 | ||
JP7319013B2 (ja) | 2020-02-17 | 2023-08-01 | Necソリューションイノベータ株式会社 | 抽出装置 |
US11973921B2 (en) | 2020-05-21 | 2024-04-30 | Subaru Corporation | Image processing apparatus |
Also Published As
Publication number | Publication date |
---|---|
CN102741884A (zh) | 2012-10-17 |
US8599252B2 (en) | 2013-12-03 |
JPWO2012014430A1 (ja) | 2013-09-12 |
CN102741884B (zh) | 2016-06-08 |
JP4979840B2 (ja) | 2012-07-18 |
US20120206597A1 (en) | 2012-08-16 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP4979840B2 (ja) | 移動体検出装置および移動体検出方法 | |
Luiten et al. | Track to reconstruct and reconstruct to track | |
US8401243B2 (en) | Articulated object region detection apparatus and method of the same | |
US8340357B2 (en) | Moving object detection method and moving object detection apparatus | |
US11037325B2 (en) | Information processing apparatus and method of controlling the same | |
EP2548174B1 (en) | Method and apparatus for trajectory estimation, and method for segmentation | |
US8582822B2 (en) | Moving object detection apparatus and moving object detection method | |
US8929599B2 (en) | Articulation region display apparatus, articulation region detecting apparatus, articulation region belongingness calculating apparatus, articulated object region belongingness calculating apparatus, and articulation region display method | |
JP4668360B2 (ja) | 移動体検出方法及び移動体検出装置 | |
Vidas et al. | Real-time mobile 3D temperature mapping | |
Djelouah et al. | Sparse multi-view consistency for object segmentation | |
JP6985897B2 (ja) | 情報処理装置およびその制御方法、プログラム | |
Smith | ASSET-2: Real-time motion segmentation and object tracking | |
Seer et al. | Kinects and human kinetics: a new approach for studying crowd behavior | |
Madrigal et al. | Motion priors for multiple target visual tracking | |
Ghedia et al. | Moving Objects Detection Using Machine Learning | |
Liu et al. | Implementation of levels-of-detail in Bayesian tracking framework using single RGB-D sensor | |
Xuezhi et al. | Scene flow estimation methodologies and applications—A review | |
Chang | Robust Tracking of Multiple People Using Two Widely Separated Cameras | |
Jie | Robust moving object detection by information fusion from multiple cameras | |
Smith | Technical Report TR95SMS2b | |
Luo et al. | Multiple People Tracking Based on Dynamic Visibility Analysis |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 201180008077.8 Country of ref document: CN |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2012501494 Country of ref document: JP |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 11812032 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 11812032 Country of ref document: EP Kind code of ref document: A1 |