WO2011013299A1 - 移動体検出装置及び移動体検出方法 - Google Patents
移動体検出装置及び移動体検出方法 Download PDFInfo
- Publication number
- WO2011013299A1 WO2011013299A1 PCT/JP2010/004378 JP2010004378W WO2011013299A1 WO 2011013299 A1 WO2011013299 A1 WO 2011013299A1 JP 2010004378 W JP2010004378 W JP 2010004378W WO 2011013299 A1 WO2011013299 A1 WO 2011013299A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- distance
- unit
- region
- movement
- geodetic
- Prior art date
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/292—Multi-camera tracking
-
- 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
- G06T2207/10—Image acquisition modality
- G06T2207/10016—Video; Image sequence
-
- 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
- G06T2207/20—Special algorithmic details
- G06T2207/20021—Dividing image into blocks, subimages or windows
-
- 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
- G06T2207/30196—Human being; Person
-
- 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 an apparatus for detecting a moving object in an image, and more particularly to a moving object such as a person who moves while changing its shape based on motion information of an image from a moving image composed of a plurality of images.
- the present invention relates to an apparatus for detecting a moving body by extracting a region that specifies all or part of a region.
- a method for extracting candidate areas from an image, and an extracted target There is a method combined with a method of applying an object model prepared in advance to an object candidate region.
- a silhouette image of a target object such as a person is extracted from a plurality of images as a target object candidate area, and a model related to the target object such as a person in which the part of the target is parameterized in advance from knowledge about the target object
- a method of applying the model to the extracted silhouette image using the above is disclosed.
- a parameterized model can be applied to a moving body such as a person who moves while changing its shape, so that it is possible to detect the moving body and extract a region.
- Non-Patent Document 1 an image obtained by photographing a fixed subject from a plurality of viewpoints is input, and a Euclidean distance representing similarity between images is calculated based on a luminance value of each image, and a geodetic survey is performed on the calculated Euclidean distance.
- a method is disclosed in which, after performing distance conversion, by performing dimensional compression, it is possible to project an image taken from a similar viewpoint so as to be a short distance in a two-dimensional space.
- PCA Principal Component Analysis
- Non-Patent Document 1 needs to perform geodetic distance conversion and dimension compression using a matrix having the number of elements of N 2 when the total number of data handled is N. Therefore, it is known that there is a problem that the amount of calculation becomes enormous when the number of data N is large.
- Non-Patent Document 2 and Non-Patent Document 3 the number of landmark points smaller than the number of data points is set from among the data points, the matrix is generated using the set landmark points, and geodetic distance conversion is performed.
- a method for reducing the amount of calculation by performing dimension compression has been proposed.
- Patent Document 1 has a problem in that a moving body cannot be correctly extracted, particularly in a scene where a moving body such as a person passes around the street.
- the region extraction method represented by Patent Document 1 needs to extract a target object candidate region from an image.
- the target object candidate area cannot be correctly extracted, it is impossible to accurately apply the model obtained by parameterizing the target object to the target object candidate area.
- it is difficult to correctly extract an object candidate region in a crowded scene.
- there is a problem that a plurality of moving objects are erroneously extracted as target object areas as one moving object, or an area in which no moving object to be extracted exists does not exist as a candidate object area.
- a first object of the present invention is to solve the problem that the region extraction of Patent Document 1 is not performed correctly.
- the property described in the following Non-Patent Document 1 that can efficiently represent non-linearly distributed data is used.
- Non-Patent Document 1 In the image processing technique represented by Non-Patent Document 1, it is possible to project image data into a low-dimensional space that is efficiently compressed by performing nonlinear processing using the distance between images as an input. Furthermore, data distributed continuously and nonlinearly can be efficiently expressed by geodetic distance conversion and dimensional compression.
- the main purpose of Non-Patent Document 1 is to visualize the similarity between images by projecting a plurality of still images to a low-dimensional space. A method for correctly extracting a moving object while responding to a change in posture is not disclosed. There is also a problem that the amount of calculation becomes enormous as the number of input data increases.
- Non-Patent Document 1 represented by Non-Patent Document 2 and Non-Patent Document 3, it was not selected as a data point existing between the landmark points, that is, the landmarks.
- linear interpolation is performed using landmark points.
- a moving body such as a person whose shape changes has a different movement depending on the part even if it is the same moving body, so that the movements of the head and feet are different.
- a second object of the present invention is to solve the problem of calculation amount in Non-Patent Document 1.
- the object of the present invention is to solve these two problems, and the present invention can be applied to an image including a moving body such as a person who moves while changing its shape, even if the posture or size of the moving body is large. It is an object of the present invention to provide a moving body detection apparatus and the like that can extract a region at high speed and without being affected by the influence.
- an embodiment of a moving body detection apparatus is a movement that detects a moving body in a moving image by extracting all or a part of the area of the moving body in the moving image.
- a body detection apparatus comprising: an image input unit that accepts a plurality of pictures constituting a moving picture; and a block composed of one or more pixels that constitute the accepted picture, between two different pictures.
- a motion analysis unit that detects a motion of an image, calculates a plurality of movement trajectories obtained by connecting the detected motions for the plurality of pictures, and divides the plurality of calculated movement trajectories into a plurality of subsets.
- an area dividing unit that sets a part of the movement trajectory as a common point in the plurality of subsets, and representing the similarity between the plurality of movement trajectories
- an embodiment of a vector data classification device is a vector data classification device that classifies a plurality of vector data into a class that is a collection of similar vector data,
- a vector data input unit for receiving a plurality of vector data, and dividing the received plurality of vector data into a plurality of subsets used for calculating a geodetic distance, and at least one of the plurality of divided subsets
- a region dividing unit that sets a part of the vector data included in the common point, a distance calculating unit that calculates a distance representing similarity between a plurality of vector data for each of the divided subsets; The calculated distance is traced from one movement locus to another movement locus while following the movement locus as a relay point.
- Approximate geodetic distance that spans the subset by integrating the geodetic distance conversion unit that converts the geodetic distance that is the distance of the route and the geodetic distance that shares the common point among the converted geodetic distances And an approximate geodetic distance calculation unit for calculating a block, and using the calculated approximate geodetic distance to perform clustering that identifies blocks having similar vector data as one region, at least for the moving image And a data classification unit for generating one class.
- the present invention can be realized not only as the mobile object detection apparatus and the vector data classification apparatus, but also as a mobile object detection method and vector data classification method using the respective constituent elements as steps, a program for causing a computer to execute the steps, It can also be realized as a computer-readable recording medium such as a CD-ROM storing the program.
- the moving object can be detected at high speed and correctly without being affected by the posture or size of the moving object even for an image including a moving object such as a person who moves while changing its shape.
- a region can be extracted from the region.
- FIG. 1 is a diagram showing a basic configuration of a moving object detection apparatus according to Embodiment 1 of the present invention.
- FIG. 2 is a hardware configuration diagram in the case where the moving object detection apparatus according to Embodiment 1 of the present invention is realized by software.
- FIG. 3 is a flowchart showing the basic operation of the moving object detection device according to Embodiment 1 of the present invention.
- FIG. 4 is a diagram illustrating a processing example of the motion analysis unit according to Embodiment 1 of the present invention.
- 5 (a) to 5 (d) are diagrams showing a processing example of the area dividing unit in the first embodiment of the present invention.
- FIGS. 6 (a) to 6 (c) are diagrams showing an example of the effect of the geodetic distance of the geodetic distance conversion unit in the first embodiment of the present invention.
- FIGS. 7A and 7B are diagrams illustrating an example of the effect of the geodetic distance of the geodetic distance conversion unit in the first embodiment of the present invention.
- FIG. 8 is a diagram illustrating a processing example of the approximate geodesic distance calculation unit according to Embodiment 1 of the present invention.
- FIGS. 9A to 9C are diagrams showing a processing example of the region extraction unit in the first embodiment of the present invention.
- 10 (a) to 10 (c) are diagrams showing a processing example of the area dividing unit in the first modification of the first embodiment of the present invention.
- FIG. 11 is a flowchart showing the basic operation of the moving object detection device according to the second modification of the first embodiment of the present invention.
- FIG. 12 is a diagram showing a dimension compression result of the region extraction unit in the second modification of the first embodiment of the present invention.
- FIGS. 13A and 13B are diagrams showing the moving object detection results of the region extraction unit in the second modification of the first embodiment of the present invention.
- FIG. 14 is a diagram illustrating a dimension compression result of the region extraction unit in the second modification of the first embodiment of the present invention.
- FIG. 15 is a diagram showing a basic configuration in a third modification of the first embodiment of the present invention.
- FIG. 16 is a flowchart showing the basic operation of the moving object detection device according to the third modification of the first embodiment of the present invention.
- FIG. 17 (a) to 17 (e) are diagrams showing a processing example of the region extraction unit in the third modification of the first embodiment of the present invention.
- FIG. 18 is a diagram showing a basic configuration in the fourth modification of the first embodiment of the present invention.
- FIGS. 19A and 19B are diagrams showing a display example of the image display unit in the fourth modification of the first embodiment of the present invention.
- FIG. 20 is a diagram illustrating a configuration example of the moving object detection device according to the fifth modification of the first embodiment of the present invention.
- FIG. 21 is a diagram showing an example of recording / transmission data in the fifth modification example of the first embodiment of the present invention.
- FIG. 22 is a diagram illustrating a configuration example of the moving object detection device according to the second embodiment of the present invention.
- FIG. 23 is a flowchart showing the basic operation of the moving object detection device according to the second embodiment of the present invention.
- FIG. 24 is a diagram illustrating an example of motion prediction according to Embodiment 2 of the present invention.
- FIG. 25 is a diagram illustrating a configuration example of the moving object detection device according to the third embodiment of the present invention.
- FIGS. 26A and 26B are diagrams showing an example of the camera arrangement of the moving object detection device according to the third embodiment of the present invention.
- FIG. 27 is a flowchart showing the basic operation of the moving object detection device according to the third embodiment of the present invention.
- FIGS. 28A and 28B are diagrams showing a processing example of the area dividing unit in the third embodiment of the present invention.
- FIG. 29 is a diagram illustrating a configuration example of the moving object detection device according to the fourth embodiment of the present invention.
- FIG. 30 is a flowchart showing the basic operation of the moving object detection device according to the fourth embodiment of the present invention.
- FIGS. 31A to 31C are diagrams showing examples of figures that can be separated by the method according to the second embodiment of the present invention.
- FIG. 32 is a diagram illustrating a configuration example of a moving object detection device according to a modification of the embodiment of the present invention.
- One embodiment of a moving body detection apparatus is a moving body detection apparatus that detects a moving body in a moving image by extracting all or a part of the area of the moving body in the moving image,
- An image input unit that accepts a plurality of pictures constituting an image, and a motion of an image between two different pictures for each block composed of one or more pixels that constitute the accepted picture, and is detected.
- a motion analysis unit that calculates a plurality of movement trajectories obtained by connecting the motions of the plurality of pictures, and divides the calculated plurality of movement trajectories into a plurality of subsets, and a part of the plurality of divided subsets.
- An area dividing unit that sets the movement trajectory as a shared point, and a distance calculation unit that calculates a distance representing the similarity between the plurality of moving trajectories for each of the divided subsets By integrating the calculated geodetic distance to the geodetic distance and the geodetic distance sharing the common point among the converted geodetic distances, the approximate geodesic across the subset Using the approximate geodetic distance calculation unit that calculates the distance and the calculated approximate geodetic distance, clustering that identifies blocks having similar movement trajectories as the same region is performed on the moving image.
- a region extracting unit that extracts at least one region.
- the approximate geodetic distance across the subset is calculated by integrating the geodetic distances using the shared points, so if the movement trajectory is not divided into subsets
- the geodetic distances for the combinations between all two movement trajectories are calculated with a smaller calculation amount, and high-speed moving object detection becomes possible. That is, since the amount of calculation required for geodesic distance calculation can be reduced, a moving body can be detected at high speed. Furthermore, it is not necessary to hold or learn advance information such as shape information relating to a moving body to be extracted in advance.
- the image input unit accepts the plurality of pictures for each of a plurality of moving images from a plurality of cameras
- the motion analysis unit displays the plurality of movement trajectories for each of the plurality of moving images.
- the region dividing unit holds the plurality of movement trajectories calculated for each of the plurality of moving images as the plurality of divided subsets, and the region extracting unit performs the clustering. At least one region may be extracted for the plurality of moving images. Thereby, the video obtained from a plurality of cameras can be processed in an integrated manner.
- the region dividing unit converts the plurality of movement trajectories into the plurality of subsets in a space related to the position of the block on the image corresponding to the movement trajectory. It may be divided. Thereby, it becomes easy to divide the movement locus located near the position on the image as one subset.
- geodesic distance calculation can be performed at higher speed.
- the region dividing unit may divide the plurality of movement loci into the plurality of subsets according to a space division designated by the user for the picture.
- the movement trajectory is expressed by a multidimensional vector
- the region dividing unit displays the plurality of movement trajectories on a multidimensional vector space expressing the multidimensional vector.
- the region dividing unit performs the division so that a part of a nearby subset overlaps, and sets a movement trajectory included in the overlapping region as the common point.
- the region dividing unit may set, for each of the plurality of subsets, a movement locus close to a boundary with another subset among the movement locus belonging to the subset as the shared point. Good.
- the geodetic distance is fast and accurate Can be calculated.
- the geodetic distance can be calculated at high speed and accurately.
- the geodetic distance conversion unit is calculated by the motion analysis unit by connecting small distances that satisfy a predetermined condition among the distances calculated by the distance calculation unit. It is preferable to convert each of the distances calculated by the distance calculation unit into a geodetic distance by obtaining a shortest path that reaches one other movement locus from the one movement locus that has been determined. Specifically, the geodetic distance conversion unit is predetermined for each of a plurality of movement trajectories included in the plurality of subsets, in order from the smallest of a plurality of distances from the movement trajectory to another movement trajectory.
- Each of the distances calculated by the distance calculation unit is converted into a geodetic distance by selecting the number of distances and performing non-linearization to change the unselected distance to infinity and then obtaining the shortest path. Is preferred.
- the distance selected by the threshold and the distance not selected are in a non-linear relationship, the similarity / dissimilarity between the movement trajectories is emphasized compared to the linear distance, like a person. It becomes possible to correctly express the movement of the object connected by the joint.
- the geodetic distance conversion unit for each of a plurality of movement trajectories included in the plurality of subsets, out of a plurality of distances from the movement trajectory to another movement trajectory, a distance below a predetermined threshold value
- Each of the distances calculated by the distance calculation unit may be converted into a geodetic distance by obtaining the shortest path after making a non-linearization by selecting and changing the distance not selected to infinity.
- the region extraction unit detects at least one discontinuous point in the geodetic distance distribution calculated by the approximate geodetic distance calculation unit, and detects the detected defect. It is preferable to perform the clustering so that the movement trajectories separated by a geodetic distance smaller than the continuous points form one cluster. As a result, movement trajectories separated by a distance smaller than the discontinuous points form one cluster, and the number of generated clusters can be controlled by changing the discontinuous point conditions.
- the region extraction unit performs dimensional compression by obtaining eigenvalues and eigenvectors for the geodetic distance calculated by the approximate geodetic distance calculation unit,
- the clustering may be performed on a dimensionally compressed space.
- the geodetic distance conversion unit generates a plurality of determination criteria for converting the distance into the geodetic distance, and each of the generated plurality of determination criteria
- the geodetic distance corresponding to each of the plurality of judgment criteria is generated by converting the distance into the geodetic distance using the judgment criteria, and the approximate geodesic distance calculation unit
- the integration is performed on the geodetic distances corresponding to each of the plurality of geodetic distances, and the region extracting unit extracts the region by performing the clustering on the integrated approximate geodetic distances corresponding to each of the plurality of determination criteria.
- an area extraction candidate generation unit that generates the area extraction result as an area extraction candidate and an instruction about the number of classes are acquired.
- a region extraction candidate from which a number of regions close to the acquired number of classes is extracted is selected from a plurality of region extraction candidates generated by the region extraction candidate generation unit, and the selected region extraction candidate is selected by the region extraction unit.
- An area extraction candidate selection unit that outputs as a result of extraction can be used.
- clustering is performed for each of the plurality of judgment criteria generated in advance, so that a desired number of clusters is generated from the plurality of clustering.
- the geodetic distance conversion unit generates a plurality of threshold values as the plurality of determination criteria, and the distance calculation unit calculates each of the generated plurality of threshold values. It is preferable to generate a geodetic distance corresponding to each of the plurality of determination criteria by connecting distances smaller than the threshold among the distances. As a result, the similarity of the movement trajectory is determined by using the geodetic distance obtained by connecting the small distances. Therefore, compared with the clustering using the Euclidean distance that is a linear distance, the continuity regarding the similarity between the movement trajectories is improved. Clustering in consideration is performed, and it is reliably discriminated whether each block in the picture belongs to the same object (or part) or separate object (or part).
- the moving body detection device further includes an image display unit that displays the region extraction result obtained by the region extraction unit on a picture received by the image input unit.
- an image display unit that displays the region extraction result obtained by the region extraction unit on a picture received by the image input unit.
- the moving body detection device further identifies a region in the picture received by the image input unit in correspondence with a result of region extraction by the region extraction unit, and for each identified region, a corresponding region extraction is performed. It is good also as a structure provided with the recording and transmission part which records and hold
- the recording and transmission part which records and hold
- the image input unit receives a moving image including two or more moving objects
- the region extracting unit detects the two or more moving objects by extracting the region of the two or more moving objects. May be. Thereby, it is possible to correctly detect a plurality of moving objects not only for one moving body that moves while changing its shape but also for an image including a plurality of moving bodies.
- the moving body detection device further calculates a movement locus representing the area from the movement locus of the block included in the area extracted by the area extraction unit, and predicts that the area moves according to the calculated movement locus.
- a motion prediction unit that predicts the motion of the moving body.
- Another embodiment of the present invention is a vector data classification device that classifies a plurality of vector data into a class that is a collection of similar vector data, a vector data input unit that receives the plurality of vector data, The received plurality of vector data is divided into a plurality of subsets used for calculation of geodetic distance, and a part of the vector data included in at least one of the plurality of divided subsets is used as a common point For each of the plurality of divided subsets to be set, a distance calculation unit that calculates a distance representing the similarity between a plurality of vector data, and the calculated distance as a relay point Geodetic distance conversion that converts the distance from one movement locus to another one while following the locus.
- an approximate geodetic distance calculation unit that calculates an approximate geodetic distance that spans the subset by integrating the geodetic distances that share the common point among the converted geodesic distances, and the calculated approximate
- a data classifying unit that generates at least one class for the moving image by performing clustering that identifies blocks having similar vector data as one region using a simple geodetic distance.
- Similar vector data is grouped and classified. For example, when vector data related to three-dimensional motion is input, an articulated object such as a person composed of a plurality of parts having different motions is moved. Even in the case of detection as a body, it is classified as one region, so that accurate region extraction is performed and a moving body is reliably detected. That is, it is possible to correctly classify a moving body such as a person who moves while changing its shape, that is, extract a region.
- the approximate geodetic distance across the subsets is calculated by integrating the geodetic distances using the shared points, so when the vector data is not divided into subsets In comparison, geodesic distances for all combinations of two vector data are calculated with a smaller amount of calculation, and high-speed data classification becomes possible.
- Embodiment 1 First, the mobile body detection apparatus and mobile body detection method according to Embodiment 1 of the present invention will be described.
- FIG. 1 is a diagram showing a configuration of the moving object detection apparatus 100 according to the first embodiment.
- the moving body detection apparatus 100 includes an image input unit 101, a motion analysis unit 102, a region division unit 103, a distance calculation unit 104, a geodetic distance conversion unit 105, an approximate geodetic distance calculation unit 106, and a region An extraction unit 107 and an output unit 108 are provided. And this moving body detection apparatus 100 detects the moving body in a moving image by extracting the area
- the image input unit 101 is a processing unit that receives input of a plurality of temporally different images (pictures) constituting a moving image, and is, for example, a video camera or a communication interface connected to the video camera.
- the motion analysis unit 102 performs image motion between two different pictures (for example, two temporally adjacent pictures) for each block including one or more pixels constituting the picture received by the image input unit 101. Is a processing unit that calculates a movement trajectory by connecting the detected motions for a plurality of pictures.
- the area dividing unit 103 is a processing unit that divides the movement trajectory calculated by the motion analysis unit 102 into a plurality of subsets, and sets a part of the movement trajectories as shared points in the plurality of divided subsets.
- the area dividing unit 103 sets a common point so that at least some of the movement trajectories overlap or are connected between the subsets. That is, in the present embodiment, the area dividing unit 103 divides a plurality of movement trajectories into a plurality of subsets in a space related to a position on an image. For example, a plurality of movement trajectories are divided into a plurality of subsets in a space related to the position on the image so that blocks having close positions on the picture corresponding to the movement trajectory belong to the same subset. Divide into
- the distance calculation unit 104 is similar to the combination of the two movement tracks in the plurality of movement tracks included in the subset. It is a processing part which calculates the distance showing.
- the distance calculation unit 104 captures the movement trajectory of the block i included in the subset calculated by the motion analysis unit 102 and the subset other than the block i in order to capture the shape change of the moving object in the image.
- a distance representing the similarity of motion between the blocks is calculated using the movement trajectory of the included block. For example, when the movement trajectory of N blocks is targeted, the calculated distance is a distance matrix composed of N ⁇ N distances.
- the distance between the blocks changes depending on the movement, in particular, an object such as a person that moves while changing its shape like a joint object. It is possible to express movement as a distance matrix.
- the movement locus of the block i is referred to as the movement locus i.
- the “distance” in this specification includes not only the distance between two points in a two-dimensional space but also the arithmetic distance between multidimensional data, and a single value or a set of a plurality of values. (Distance matrix).
- the geodetic distance conversion unit 105 is a processing unit that performs geodetic distance conversion on the distance matrix in each subset calculated by the distance calculation unit 104.
- the approximate geodetic distance calculation unit 106 calculates an approximate geodetic distance by integrating the geodetic distances across the subsets using the shared points among the geodetic distances converted by the geodetic distance conversion unit 105. It is a processing unit that calculates an approximate geodetic distance for a combination of two movement trajectories in a plurality of movement trajectories calculated by the motion analysis unit 102 by performing processing. That is, the approximate geodetic distance calculation unit 106 calculates an approximate geodetic distance that spans between the subsets by integrating the geodetic distance matrices in the respective subsets calculated by the geodetic distance conversion unit 105.
- the reason why the “approximate” geodetic distance is calculated is that the geodetic distance across the subsets is compared to the case where the geodetic distances are calculated using all the movement trajectories belonging to the subsets. This is because the geodesic distance is generally calculated using the movement trajectory set as the common point.
- the region extraction unit 107 is a processing unit that performs region extraction by specifying a region composed of blocks having similar movement trajectories based on the distance calculated by the approximate geodesic distance calculation unit 106.
- the image input unit 101 performs clustering that identifies blocks having movement loci that are separated by a geodetic distance smaller than a certain threshold among the geodetic distances integrated by the approximate geodetic distance calculation unit 106 as one region. At least one region is extracted from the received moving image. More specifically, in the present embodiment, this region extraction unit 107 uses the distance matrix calculated by the distance calculation unit 104 to use the distance between the subsets obtained by the geodetic distance conversion unit 105 and the approximate geodetic distance calculation unit 106.
- Discontinuity in the distribution of the distance between the trajectories is detected from the geodetic distance across the two, and the discontinuous points are also arranged so that the trajectories separated by a distance smaller than the detected discontinuity become one cluster.
- the moving trajectories that are continuously distributed to each other are clustered to detect a moving object in the image and extract an area of the image, and as a result, detect a moving object in the image and extract an area of the image.
- the output unit 108 is a processing unit that outputs a detection result of a moving object in a moving image performed by the region extraction unit 107 or a region extraction result of the image, a writing unit to a memory or a hard disk, a display device, or the like.
- Unit 107 and output unit 108) are, as shown in FIG. 2, a CPU 1005, a RAM 1007, a ROM 1006, a communication interface 1004 for connecting to the camera 1001, an I / O port (video card 1009, etc.), a hard disk 1008, and a display 1003.
- the program may be realized by software such as a program executed on a computer 1002 provided with the above, or may be realized by hardware such as an electronic circuit.
- the components other than the input / output device such as the display device may be realized by a program and data (that is, software) executed by a computer, an electronic circuit, It may be realized by hardware such as a memory and a recording medium, or may be realized by mixing them.
- the functional elements of the present invention are realized by executing a program using computer hardware resources such as a CPU, a memory, and an input / output circuit. . That is, the CPU reads out (takes out) data to be processed from the memory or the input / output circuit, and temporarily stores (outputs) the operation result in the memory or the input / output circuit.
- the function is realized.
- the present invention when the present invention is realized by hardware, it may be realized by a one-chip semiconductor integrated circuit, or may be realized by mounting a plurality of semiconductor chips on one circuit board. It may be realized as one device in which all the components are housed in one housing, or may be realized by cooperation of a plurality of devices connected by a transmission path.
- the present invention can be realized by a server / client system by providing the storage unit in the present embodiment in a server device and providing the processing unit in the present embodiment in a client device that wirelessly communicates with the server device. Good.
- step S201 the image input unit 101 receives a plurality of pictures.
- step S202 the motion analysis unit 102 calculates a block motion from at least two pictures.
- pixel motion is calculated as an example of calculating block motion.
- pixel unit processing will be described as an example of block unit processing.
- Non-Patent Document 4 the method disclosed in Non-Patent Document 4 or Non-Patent Document 5 can be used to calculate a motion vector by optical flow calculation.
- 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 motion vector (u i t , v i t ) of the pixel i is estimated using the pictures input at time t and time t + 1.
- the frames do not necessarily have to be continuous.
- the motion of the pixels may be obtained using pictures input at time t and time t + n.
- n is an integer of 1 or more.
- an affine parameter may be estimated as the pixel motion instead of the two-dimensional motion vector. At this time, motion information may be obtained for all pixels.
- the image may be divided into grids to obtain motion information only for pixels on the grid at regular intervals, or as described above, the image may be divided into blocks and each block may be obtained. You may also ask for motion information. Furthermore, when a motion vector is calculated using the method disclosed in Non-Patent Document 4, since the reliability can be calculated, only pixels having motion information with high reliability may be used. Further, when the motion vector is calculated using the method disclosed in Non-Patent Document 5, the occlusion can be estimated. 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.
- the method disclosed in Non-Patent Document 6 below can be used. Jianbo Shi and Carlo Tomasi, “Good Features to Track”, IEEE Conference on Computer Vision and Pattern Recognition, pp 593-600, 1994.
- the above method estimates the affine parameter A i t corresponding to the motion in the vicinity of the pixel i of the picture input at time t and time t + 1.
- pixel positions x i t and x i t + 1 on the picture at time t and time t + 1 have the following relationship.
- the above method can estimate the motion of the pixel i with higher accuracy than the method of calculating a motion vector assuming a translational movement, particularly for an object that rotates.
- step S203 the motion analysis unit 102 uses the motion information calculated in step S202 to calculate the movement trajectory i from the pixel motion for a plurality of temporally different pictures.
- the movement locus of the pixel i is referred to as a movement locus i.
- the motion of the pixel i is tracked using the motion information 302 calculated in step S202 from the pixel i303 of the input image 301 at time t.
- the movement locus i (here, x i ) is calculated using the pixel position (x i t , y i t ) on the picture at the time t when the pixel i passes.
- T is the number of pictures used to calculate the movement trajectory. Furthermore, in the expression 2, it is possible not to perform the following processing for a movement trajectory having no movement or a movement trajectory having a small movement.
- the area dividing unit 103 divides the space related to the position of the image 401 into areas as P subsets 402 as shown in FIG.
- the subset 402 is also a set of movement trajectories corresponding to the divided image regions (positions on the image).
- the subsets are set or the common points are set so that the subsets have a common point (movement trajectory that logically belongs to both subsets).
- a part of the space related to the position of the image 401 is set to overlap and belongs to the overlapping region.
- the movement trajectory corresponding to the image region may be set as a shared point, or the movement trajectory corresponding to the position on the image close to the boundary of the subset as shown in FIG.
- the shared points 403 may be set, or as shown in FIG. 5D, the subsets may be set spatially at high density, and the shared points 403 may be set spatially sparse.
- FIG. 5 (a), FIG. 5 (c), FIG. By making it square as shown in (d), there is an effect of making it difficult to be affected by the rotation of the subject or the camera.
- a subject that is long in the vertical direction such as a person
- the shape of the person can be extracted more accurately by using a vertically long rectangle as shown in FIG.
- an animal such as a dog
- the amount of calculation is further reduced by reducing the overlap.
- the size of the subset at the top of the image is made smaller than the subset at the bottom of the image, so that the subject at the bottom of the image is large and the subject at the top of the image is small.
- the effect of reducing the subset at the top of the image instead of the square can also be realized by dividing the subset into trapezoids. The number and size of each subset will be described later.
- the movement locus corresponding to the image area is obtained by dividing the movement locus i into a subset based on the position on the image in the divided image area, for example, When the position (x i 1 , y i 1 ) on the image belongs to the subset p, the movement shown in the above equation 2 through the position (x i 1 , y i 1 ) on the image at time 1 Assign trajectory i to subset p. Further, when the position (x i t , y i t ) on the image at a certain time t belongs to the subset p, the movement trajectory i passing through (x i t , y i t ) is assigned to the subset p. You may do it. Furthermore, the time average position on the image of the movement trajectory is calculated by calculating the time average for the movement trajectory i of Equation 2 above.
- the movement locus i may be assigned to the subset p to which the time average position belongs.
- the shared point is a movement locus corresponding to the image region (position on the image). As described above, each of the movement trajectories is divided into a plurality of subsets by dividing the image into P pieces based on the image regions (positions on the image). Will be divided.
- the distance calculation unit 104 calculates, for each subset p divided in step S204, a distance matrix including similarity of pixel motion using the movement trajectory i belonging to the subset p. .
- the linear distance f p (i, j) representing the similarity between the movement trajectory i and the movement trajectory j can be calculated by the following equations 3 to 13. It should be noted that the distances represented by Equations 3 to 13 may include a nonlinear function, but in order to distinguish from the geodetic distance calculated by the nonlinear transformation and the route searching method described later, This is called a linear distance.
- the movement trajectory i and the movement trajectory j belong to the same subset p, and the following calculation is performed for each subset p.
- w is a weighting coefficient, which is a parameter set by the designer.
- w is a weighting coefficient, which is a parameter set by the designer.
- the linear distance f p (i, j) may be defined as:
- ptn ij and mtn ij are the time average value of the distance between the moving trajectories and the time fluctuation component of the distance between the moving trajectories, and the definitions thereof will be shown below.
- the time variation component of the distance between moving tracks shown in Equation 6 is changed to the linear distance f p ( i, j).
- the fluctuation component of the distance between the movement trajectories shown in Expression 6 indicates the similarity of the movements of the pixels, so that not only the rigid body in which the relationship between the distances between the pixels does not change with time but also the joint object. Etc. can be captured.
- the same effect can be expected by using time-variable components such as Expression 8 to Expression 13 instead of Expression 6.
- u t i is a motion vector from time t to t + 1 (u i t, v i t), ⁇ u t i ⁇ u t i> is the inner product.
- the following calculation may be used as the time variation component.
- step S206 the geodetic distance conversion unit 105 uses the threshold R for the linear distance f p (i, j) in the subset p calculated by the above formula 3 or the above formula 4 as follows: Nonlinear processing is performed to calculate f ′ p (i, j).
- the R movement trajectories j are selected in ascending order of the linear distance f p (i, j) to the movement trajectory i, and the distance to the selected movement trajectory j is selected without being changed.
- the distance from the missing movement locus j is changed to infinity.
- the linear distance is selected in ascending order, but the threshold value R may be set as in the following equation.
- the geodetic distance conversion unit 105 selects a predetermined number (R) of movement trajectories in ascending order of distance for each of a plurality of movement trajectories belonging to the subset p calculated by the motion analysis unit 102, After non-linearizing the distance to the unselected movement trajectory to infinity, each of the plurality of distances may be converted into a geodetic distance, or belongs to the subset p calculated by the motion analysis unit 102 For each of the plurality of movement trajectories, after selecting a movement trajectory whose distance is equal to or less than a predetermined threshold and performing non-linearization to change the distance from the unselected movement trajectory to infinity, each of the plurality of distances is changed. You may convert into geodesic distance.
- the distance non-linearization is not limited to the above function, and any distance can be used as long as the nonlinear transformation is performed on the distances related to the movement trajectory i and the movement trajectory j.
- linear distance f p (i, j) is weighted by multiplying the weight calculated by using Expression 16 and Expression 17 as follows, and then the processing of Expression 14 or Expression 15 is performed. Also good.
- NN indicates that processing is performed on points that are close to the movement trajectory and belong to the same subset p.
- a movement trajectory within a certain distance from the movement trajectories j and i, or The calculation is performed using N movement trajectories in ascending order of distance. That is, N a and N b are the number of movement trajectories belonging to the same subset p and within a certain distance, or N. Note that z is set by the designer.
- dispersion may be used as in the following equation.
- the geodetic distance g p (i, j) is calculated using the non-linearized distance f ′ p (i, j) as in the following equation.
- min (x, y) is a function that returns the smaller of the values x and y.
- s is a movement trajectory s, which is a relay point for tracing from the movement trajectory i to the movement trajectory j.
- the relay point s in f ′ p (i, s) + f ′ p (s, j) is not limited to one point.
- p corresponds to each subset. This method is a shortest path search method called the Dijkstra method, and is described in Non-Patent Document 7 below.
- the geodetic distance conversion unit 105 connects one of the distances calculated by the distance calculation unit 104 with a small distance that satisfies a predetermined condition, thereby calculating one movement locus calculated by the motion analysis unit 102.
- Each of the distances calculated by the distance calculation unit 104 is converted into a geodetic distance by obtaining the shortest route from one to the other movement locus.
- FIG. 6A shows a two-dimensional data distribution.
- each data point corresponds to the movement trajectory i shown in the above equation 3 or 4.
- the distance between the data point i and the data point j is the distance between the data point i and the data point k. Smaller than.
- FIG. 6B shows a two-dimensional data distribution.
- the distance between the data point i and the data point j is not the Euclidean distance but the geodetic distance, for example, by performing the processing of Expression 15 and Expression 17 above. This is the distance that the data point is traced as indicated by an arrow. As a result, unlike the case where the Euclidean distance is used, the distance between the data point i and the data point j is larger than the distance between the data point i and the data point k.
- the distance between the movement trajectories at time t as shown in the above equation 5 is taken as an example.
- the fluctuation component of the distance between the movement trajectories as the similarity of the movement of the pixels as shown in the above equation 3, not only the shape of the joint object etc. but also the shape change Can also be captured.
- FIG. 7A shows an example of the case where the non-linearization process of Expressions 14 to 17 is not performed.
- the distance between the head pixel i 602 and the hand tip pixel j 603 is a distance indicated by a linear distance 601.
- the threshold value R is appropriately set by performing nonlinear processing such as Expression 14 to Expression 17, as shown in FIG. 7B
- the distance to j603 is a distance (that is, a geodetic distance) as a linear sum as indicated by an arrow from the pixel k604 to the pixel j.
- the linear distance 601 cannot continuously represent the shape of joints of joint objects such as a person as data, whereas the geodetic distance expresses continuity using the shape of joints as a distance. It becomes possible.
- the geodesic distance calculation method is not limited to the method that requires the nonlinear processing of Equation 14 to Equation 17.
- the distance in the linear distance and the geodetic distance is in a relationship opposite to the similarity, and the distance is small when the similarity is high, and the distance is large when the similarity is small. Therefore, when using the similarity instead of the distance described above, the reciprocal of the similarity is used as the distance, or the value obtained by subtracting the similarity from the value equal to or greater than the maximum value of the similarity is used as the distance. It may be converted into a distance so as to satisfy the conflicting relationship.
- step S207 the approximate geodetic distance calculation unit 106 calculates an approximate geodetic distance across the subsets by integrating the geodetic distance matrix g p (i, j) in each subset. Specifically, the integration process is performed as follows using the shared point set in step S204.
- c is a common point and is a movement trajectory belonging to both the subset p and the subset q. That is, the movement trajectory is overlapped between a plurality of subsets.
- the shared point does not necessarily belong to two subsets, and may belong to three or more subsets.
- i and h are indexes corresponding to movement trajectories belonging to different subsets p and q, respectively.
- the shared point c is not limited to one point as in the example shown in FIG.
- the present invention is not limited to two sets of subsets p and q. If a plurality of shared points are used and the geodetic distance is added so as to straddle each subset, with respect to Equation 21 above. A geodetic distance spanning between two or more subsets can be calculated.
- the linear distance f c (c i , c j ) has an effect of connecting the shared points.
- f c (c i , c j ) may be calculated for all shared points, or a distance between at least one shared point that is spatially close may be obtained.
- it is not limited to the two sets of subsets p and q, and the geodetic distance so as to straddle each of the subsets using the linear distance connecting the plurality of shared points with respect to Equation 22 above.
- step S204 the subsets are divided into different subsets a and b.
- step S205 geodesic distances in the respective subsets (for example, geodetic distance 704 in subset a and geodetic distance 705 in subset b) are calculated. To do. Then, in step S207, by performing the integration process using the shared point 703, for example, a movement locus i corresponding to the head point and a movement locus j corresponding to the toe point in FIG. An approximate geodetic distance g (i, j) across the subsets can be obtained.
- the geodetic survey is performed with the same accuracy as the case where the process of dividing into the subsets is not performed by performing the integration process.
- the distance can be calculated.
- the geodetic distance can be calculated with a small amount of calculation compared to the case where the geodetic distance is calculated without being divided into subsets. Specifically, it is shown in Non-Patent Document 3 that the amount of calculation required for geodetic distance calculation is O (N 2 logN), where N is the number of movement trajectories.
- the approximate calculation amount for the distance calculation and integration is as follows.
- M is the number of subsets
- C is the number of all sharing points, that is, the total number of sharing points between the partial areas shown in FIG. C 2 logC is the amount of calculation required for the geodetic distance integration processing in the above equation 21.
- N the number of movement trajectories
- the amount of calculation is approximately 0, compared with the case where it is not divided into subsets. .11 times.
- the amount of calculation reduction can be estimated by using two variables, the number M of subsets and the common point C. When accuracy is important, it is desirable to take many overlapping areas.
- the number C of all shared points it is preferable to increase the number C of all shared points. Furthermore, when importance is attached to the calculation amount, it is preferable to increase the number M of subsets. However, considering the accuracy, the number C of all shared points also needs to be increased accordingly. It is necessary to decide on the balance between movement and accuracy. As described above, by integrating geodetic distances using shared points, geodetic distance calculation processing can be performed at high speed.
- step S208 the region extraction unit 107 uses the geodetic distance g (i, j) integrated by the approximate geodetic distance calculation unit 106 to detect discontinuous points, and from blocks having similar movement trajectories. Region extraction is performed by specifying the region to be configured.
- g (i, j) is infinite.
- An example of the geodetic distance obtained for the threshold value R will be described with reference to FIGS. 9 (a) to 9 (c).
- FIG. 9 (a) shows the movement trajectories a to h
- FIG. 9 (a) shows the movement trajectories a to h
- FIG. 9 (b) is a conceptual diagram of a multidimensional space expressing the movement trajectories a to h shown in FIG. 9 (a).
- the number of the movement trajectories a to h shown in FIG. 9A is eight, actually, the movement trajectory corresponding to each pixel may be used, or the movement trajectory obtained in units of blocks may be used. It may be used.
- one point (point a to point h) in the multidimensional space shown in FIG. 9B corresponds to one moving locus shown in the above equation 2. In other words, this is a result of tracking pixels not only over an area on one picture but over a plurality of pictures that differ in time.
- the distance between the points does not correspond to the Euclidean distance between the vectors, but corresponds to the geodetic distance calculated from the above equations 20 and 21 (however, Except infinity).
- FIG. 9C An example of clustering shown in FIG. 9C will be described.
- the distance between the movement locus a and the movement locus b shown in the above equation 3 or 4 is f (a, b)
- the threshold value is set as R
- g p (e, f) is infinite even if the geodetic distance is obtained by the above equation 20.
- the result g (e, f) obtained by integrating the geodetic distances by the above formula 21 or 22 is also infinite. Therefore, the region extraction unit 107 determines that the distance between the movement locus e and the movement locus f is a discontinuous point. As a result, the geodesic distance between 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. Conversely, the movement trajectories f to h and the movement trajectories a to e The geodetic distance to each movement locus is infinite because it passes through the discontinuous point g (e, f).
- the set of the movement trajectory i and the movement trajectory j whose geodetic distance does not become infinite is the same cluster, and when it is infinite, it is a different cluster.
- the two clusters of ⁇ 1 and ⁇ 2 can be separated.
- the movement trajectories a to h in FIG. 9C the movement trajectories a to e regarding the upper body belong to the cluster ⁇ 1
- the movement trajectories f to h regarding the lower body belong to the cluster ⁇ 2 . That is, the result of separation into clusters is the result of direct region extraction.
- each region obtained as a result of performing region extraction without using a movement trajectory having no motion corresponds to each moving body. It is not necessary to distinguish from region extraction for extracting each mobile region.
- region extraction is performed using a movement trajectory that does not move, for example, the region having the maximum size can be used as the background, and other regions can be used as the detection result of the moving object.
- the method for detecting the moving object from the result is not limited to this.
- clustering is performed based on the distance between pixels or the similarity of movement trajectories, whereby similar movement trajectories are collected and region extraction is performed.
- a region that is close in distance and moves similarly is recognized as one region, and as a result of temporally tracking the region of the object moving in the moving image, regardless of the posture of the joint object, It is possible to detect a moving object or a part of the moving object in a moving image and extract an area of an image including the moving object.
- the geodetic distance across the subset is calculated by integrating the geodetic distance using the shared points, so compared to the case where the movement trajectory is not divided into subsets. With a small amount of calculation, geodesic distances for all combinations of two moving trajectories are calculated, and high-speed moving object detection is possible.
- the region extraction that is, the detection of the moving object is performed at high speed without being affected by the posture or size of the moving object. It can be performed.
- the moving body detection apparatus has the same configuration except for the processing of the area dividing unit 103 as compared with the first embodiment, description of the same components is omitted.
- the processing in the area dividing unit 103 is the processing in the area dividing unit 103.
- the space related to the position of the image 401 is divided into regions as P subsets 402, but here, the movement trajectory as shown in Equation 2 above, that is, multidimensional Divide into P subsets on the vector space. That is, in the present modification, the region dividing unit 103 performs a multi-dimensional vector space representing a movement trajectory (multi-dimensional vector) so that similar movement trajectories belong to the same subset. A plurality of movement trajectories are divided into a plurality of subsets.
- steps S201 to S203 are the same as those in the first embodiment, description thereof will be omitted.
- step S204 the region dividing unit 103 divides the moving trajectory calculated by the above equation 2 as shown in FIG. 10 (a) to FIG. 10 (c), that is, into a subset on the multidimensional vector space.
- FIG. 10A shows a movement locus 901
- FIG. 10B is a conceptual diagram of a multidimensional vector space expressing the movement locus 901 shown in FIG. 10A.
- a multidimensional vector one point in the multidimensional vector space
- FIG. 10B shows the result of tracking pixels over a plurality of pictures that differ not only in the area on one picture but also in time.
- the region dividing unit 103 divides the multidimensional vector space into a plurality of partial regions.
- the subset 903 is a set of points existing in the divided multidimensional vector space, that is, a movement locus.
- the black circle is a shared point 904.
- a part of the multidimensional vector space is set so as to overlap, and the movement trajectory belonging to the overlapping region may be used as the shared point.
- the movement trajectory that is close at the boundary of the subset in the multidimensional vector space may be set as the shared points 904, the subsets may be set spatially dense, and the shared points 904 may be set spatially sparse.
- a rectangular parallelepiped as shown in FIG. 10 (c) (in fact, it is multidimensional, so an n-dimensional rectangular parallelepiped). It may be a polyhedron. Moreover, it is desirable to set so as to include all the movement trajectories by a plurality of subsets.
- step S204 calculation is performed as follows for all the movement trajectories.
- i and j are all the movement trajectories calculated by Equation 2.
- Equation 7 the distance matrix f (i, j) between the movement trajectories calculated by the above equation 24 is regarded as several row vectors of movement trajectories. That is, the i-th row vector corresponds to the movement locus i.
- Each row vector can be considered as a multidimensional vector. Therefore, conceptually, the row vector of the distance matrix between the movement trajectories can be expressed as a point on the multidimensional space shown in FIG. Therefore, in the same manner as described in the above embodiment, the movement trajectory can be divided into subsets on the multidimensional vector space using the subsets as shown in FIG.
- PCA Principal Component Analysis
- a movement trajectory that is, the above-described multidimensional vector, in a multidimensional space expressing a multidimensional vector.
- the multidimensional vector space can be compressed to a lower dimension.
- the movement trajectory can be divided into subsets in the two-dimensional space by the same processing as dividing into subsets in the image space. Is possible.
- step S206 and subsequent steps may be performed in the same manner as in the first embodiment, description thereof is omitted.
- clustering is performed based on the distance between pixels or the similarity of the movement trajectories, whereby similar movement trajectories are collected and region extraction is performed.
- a region that is close in distance and moves similarly is recognized as one region, and as a result of temporally tracking the region of the object moving in the moving image, regardless of the posture of the joint object, It is possible to detect a moving object or a part of the moving object in a moving image and extract an area of an image including the moving object.
- the region extraction that is, the detection of the moving object is performed at high speed without being affected by the posture or size of the moving object. It can be performed.
- the moving body detection apparatus has the same configuration except for the processing of the region extraction unit 107 as compared with the first embodiment, the description of the same components is omitted.
- This area extraction unit 107 dimensionally compresses a geodetic distance matrix, and clustering movement trajectories in the dimension-compressed space, thereby extracting a region. It is a point to detect a moving body.
- steps S201 to S207 are the same as those in the first embodiment, description thereof is omitted.
- step S1001 the region extraction unit 107 performs dimensional compression of the geodetic distance matrix calculated in step S207.
- Dimensional compression can be realized by obtaining Eigensystem after performing Young-Householder conversion. This is a method for efficiently projecting data distributed in a multi-dimensional space to a low-dimensional space, and noise in input data (here, an error is included in motion information when calculating the movement trajectory of Equation 2 above). Data can be expressed robustly.
- the region extraction unit 107 generates a non-linear distance matrix.
- a Young-Householder transformation is performed on the above equation 25 by multiplying 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 shown in Equation 2 above.
- the region extraction unit 107 calculates Q eigenvectors (eigenvectors) e q for ⁇ (G) and the corresponding eigen value (eigenvalues) ⁇ q for dimensional compression.
- e i q is the i-th element of the q-th eigen vector e q .
- the number Q of eigen vectors may be determined experimentally according to the scene to be used, or may be determined based on the contribution rate a q calculated from the eigen value ⁇ q as follows.
- Q is the number of eigen vectors used, that is, the number of dimensions of the compressed space.
- N is the number of all eigen vectors. Therefore, Q when the contribution rate a Q becomes a certain value or more may be set as the number of eigen vectors.
- the movement trajectory i shown in the above equation 2 can be associated with the data z i q on the non-linearly dimensionally compressed space stretched by the eigen vector e q .
- FIG. 12 shows the result of projecting the movement trajectory i when a human walking image is input into a non-linearly dimensionally compressed space.
- the horizontal and vertical axes are eigen vector e 1 and e 2 , respectively.
- Points are projected on a two-dimensional (z i 1, z i 2 ) is obtained by projection of g i.
- the point (z i 1 , z i 2 ) It can be understood that it corresponds to i.
- the number of dimensions of the compression space is two-dimensional. However, the number of dimensions is not necessarily two-dimensional, and data can be projected with higher accuracy with a higher number of dimensions.
- step S1002 the region extraction unit 107 performs detection of the moving body and region extraction by performing clustering on the data projected in the compressed space as shown in FIG. 12 in step S1001. Do.
- the segment area is expressed as follows.
- the segment area is an extracted area, that is, coincides with the detection result of the moving object.
- M is the number of segment areas and is determined empirically according to the scene to be used. However, it is different from M in Equation 23 above.
- Each segment area ⁇ m is a parameter
- I an average value of the coordinate values of the data belonging to the segment area ⁇ m in the compression space
- Z m is a covariance matrix relating to the coordinate values of the data belonging to the segment area ⁇ m .
- 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 space by a grid at equal intervals.
- C m is the number of data belonging to the segment area ⁇ m in the compressed space.
- the region extraction unit 107 obtains a segment region ⁇ m to which the data z i on the compressed space belongs using the following distance function.
- ⁇ m (z i ) indicates the distance between the data z i on the compression space corresponding to the movement locus i and each segment region ⁇ m .
- Each data belongs to the segment area ⁇ m in which ⁇ m (z i ) has the minimum value.
- ⁇ m (z i ) is the Mahalanobis distance
- ⁇ m (z i ) may be used in place of ⁇ m (z i ).
- p ( ⁇ m ) may be a constant value, or may be set in advance based on the shape and area ratio of the person part when segmenting a predetermined moving body such as a person.
- ⁇ m is a weighting factor for the segment region ⁇ m .
- the region extraction unit 107 uses the data z i belonging to the segment region ⁇ m from the calculation result of the above equation 35, and sets the parameter of the region ⁇ m as follows.
- the region extraction unit 107 can obtain the segment region ⁇ m to which each data on the compression space belongs by repeating the distance calculation and the parameter update of the above Equations 35 to 38 for a specified number of times.
- other clustering methods such as k-means and competitive learning may be used.
- the segment area ⁇ 1 is the head of the person
- the segment area ⁇ 2 is the upper body
- the segment area ⁇ 3 is the arm
- the segment area ⁇ 4 corresponds to the lower part of the body
- segment areas ⁇ 5 and ⁇ 7 correspond to the thigh
- segment areas ⁇ 6 and ⁇ 8 correspond to the lower leg.
- the segment area ⁇ 9 mainly corresponds to a part of the background.
- the segment area on the compression space corresponds not only to the area on one image but also to a plurality of temporally continuous images. It is the result of tracking. That is, as a result of temporally tracking the region of an object moving in the image by performing segmentation on the compressed space, detection of the moving object (part of the moving object) in the image, and detection of the image including the subject Extraction can be done.
- the threshold value R is set to a value larger than that for the purpose of segmentation of the part by using the formula 14 or the formula 15.
- the correspondence of the segment areas ⁇ 1 and ⁇ 2 on the compressed space on the image will be described.
- the segment area ⁇ 1 corresponds to the movement of the bicycle on the image
- the segment area ⁇ 2 corresponds to the walking of the person on the image.
- the segment area in the compressed space corresponds to not only the area on one image but also a plurality of temporally continuous images. It is a result. That is, by performing segmentation on the compression space, each moving object can be detected as a result of temporally tracking the area of the moving object.
- a motion detection error or the like is performed by performing clustering after performing dimensional compression based on the distance between pixels or the similarity of movement trajectories.
- Objects that are robust against noise and similar movement trajectories are grouped together and region extraction is performed, so that objects that are close in distance and have similar movement are recognized as one area and move in the moving image
- the region extraction that is, the detection of the moving object is performed at high speed without being affected by the posture or size of the moving object. It can be performed.
- a region extraction candidate is generated, and a candidate closest to the predetermined number of moving objects is selected from the candidates.
- An example of performing region extraction by selection will be described.
- FIG. 15 is a diagram illustrating a configuration of the moving object detection device 100a according to the third modification of the first embodiment.
- the moving object detection device 100 a includes an image input unit 101, a motion analysis unit 102, a region division unit 103, a distance calculation unit 104, a geodetic distance conversion unit 105 a, an approximate geodetic distance calculation unit 106 a, and a region An extraction unit 107a and an output unit 108 are provided.
- the region extraction unit 107a includes a region extraction candidate generation unit 1401 and a region extraction candidate selection unit 1402.
- the moving body detection apparatus 100a detects the moving body in the moving image by performing region extraction that specifies all or a part of the moving body in the moving image.
- the image input unit 101, the motion analysis unit 102, the region division unit 103, and the distance calculation unit 104 are the same as those in the first embodiment, description thereof will be omitted.
- the geodetic distance conversion unit 105a generates a plurality of determination criteria for the distance matrix in each subset calculated by the distance calculation unit 104, and uses the determination criteria for each of the generated plurality of determination criteria. It is a processing unit that generates a geodetic distance corresponding to each of a plurality of determination criteria by performing conversion.
- the approximate geodesic distance calculation unit 106a calculates an approximate geodetic distance across the subsets by integrating the geodetic distance matrices in the respective subsets calculated by the geodetic distance conversion unit for each of the plurality of determination criteria. It is a processing unit.
- the region extraction candidate generation unit 1401 is a region obtained by clustering a plurality of movement trajectories calculated by the motion analysis unit 102 from a geodetic distance matrix integrated for each of a plurality of determination criteria by the approximate geodetic distance calculation unit 106a. It is a processing unit that performs extraction and generates a result of the region extraction as a region extraction candidate. Specifically, the region extraction candidate generation unit 1401 detects discontinuous points in the distribution of distances between moving trajectories using a threshold, and one moving trajectory separated by a geodetic distance smaller than the detected discontinuous point is one. Region extraction candidates for each of the plurality of threshold values are generated by clustering the movement trajectories distributed continuously so as to form a cluster.
- the area extraction candidate selection unit 1402 acquires an instruction about the number of classes by referring to a predetermined numerical value or accepting an instruction from the outside such as a user, and the number of areas close to the acquired number of classes.
- the region extraction candidates divided into two are selected from a plurality of region extraction candidates generated by the region extraction candidate generation unit 1401, and the selected region extraction candidates are extracted from the movement trajectory calculated by the motion analysis unit 102. It is a processing part which outputs as a result.
- the region extraction candidate selection unit 1402 selects a region extraction result closest to the instructed class number from the region extraction candidates generated by the region extraction candidate generation unit 1401 for each of a plurality of threshold values. . That is, a region extraction result with a threshold corresponding to the instructed number of classes is selected.
- the output unit 108 is the same as that in the first embodiment. Thereby, the final detection of the moving body and the region extraction result can be obtained.
- Steps S201 to S205 are the same as those in the first embodiment, and a description thereof will be omitted.
- step S206a the geodetic distance conversion unit 105a uses the K thresholds R k as a plurality of determination criteria for the linear distance f p (i, j) in the subset p calculated by the above equation 3 or 4. , And using these K threshold values R k , a non-linearization process is performed on each threshold value as follows to calculate f ′ k p (i, j).
- R k movement trajectories j are selected in ascending order of the linear distance from the movement trajectory i, the distance from the selected movement trajectory j is not changed, and the unselected movement trajectory j Change the distance to infinity.
- the linear distance is selected in ascending order, but the threshold value R k may be set as in the following equation.
- the geodetic distance conversion unit 105a uses the K thresholds R k for the linear distance matrix for each of the plurality of movement trajectories belonging to the subset p calculated by the distance calculation unit 104, and the distance is calculated. After selecting a predetermined number of movement trajectories in ascending order and making it non-linear to change the distance from the unselected movement trajectory to infinity, each of a plurality of distances may be converted into a geodetic distance. After selecting a movement trajectory whose distance is equal to or less than a predetermined threshold, and making the distance from the movement trajectory not selected to infinity, each of the plurality of distances may be converted into a geodetic distance Good.
- the distance non-linearization is not limited to the above function, and any distance can be used as long as the nonlinear transformation is performed on the distances related to the movement trajectory i and the movement trajectory j.
- the linear distance f p (i, j) is multiplied by the weight calculated using Expression 17 and Expression 18, and then weighted.
- the processing of 39 or the above equation 40 may be performed.
- a geodetic distance is calculated using the non-linearized distance f ′ k (i, j) as in the following equation.
- min (x, y) is a function that returns the smaller of the values x and y.
- s is a movement trajectory s, which is a relay point for tracing from the movement trajectory i to the movement trajectory j.
- the relay point s in f ′ k p (i, s) + f ′ k p (s, j) is not limited to one point.
- p corresponds to a subset
- k corresponds to a plurality of threshold values R k .
- step S207a the approximate geodesic distance calculation unit 106a approximates the subsets by integrating the geodetic distance matrix g k p (i, j) in each subset p for each threshold value R k. Calculate geodetic distance. Specifically, the integration process is performed as follows using the shared point set in step S204.
- c is a shared point and is a movement locus belonging to both the subset p and the subset q. That is, the movement trajectory is overlapped between a plurality of subsets.
- the shared point does not necessarily belong to two subsets, and may belong to three or more subsets.
- i and h are movement trajectories belonging to different subsets p and q, respectively.
- the shared point c is not limited to one point as in the example shown in FIG.
- it is not limited to two sets of subsets p and q, but a geodetic distance across two or more subsets can be calculated.
- calculation for connecting the common points may be performed for each threshold value.
- step S208a the region extraction candidate generation unit 1401 performs clustering by detecting discontinuous points using the integrated geodetic distance matrix g k (i, j) corresponding to each threshold value R k. Do.
- FIG. 17A shows the movement trajectories a to h
- FIG. 17B is a conceptual diagram of the multidimensional space expressing the movement trajectories a to h shown in FIG.
- movement loci is eight (movement loci a to h), in actuality, a movement locus corresponding to each pixel may be used, or a movement locus obtained in units of blocks may be used.
- one point in the multidimensional space expressing the movement locus shown in FIG. 17B corresponds to one movement locus shown in Expression 2, respectively.
- this is a result of tracking pixels not only over an area on one picture but over a plurality of pictures that differ in time.
- the distance between points corresponds to a geodetic distance, not a Euclidean distance between vectors.
- the threshold value R k is a sufficiently large value, for example, when the threshold value R k is larger than the maximum value of f p (i, j), the geodetic distance g k (i , J) does not become infinite in all combinations of i and j. That is, since there is no discontinuous point, it can be determined that there is one cluster.
- the threshold value R k is sufficiently small, specifically, when the threshold value R k is smaller than the minimum value of f p (i, j), g k p (i , J) becomes infinite. That is, the number of clusters is the same as the number of movement trajectories.
- the movement trajectories f to h and the movement trajectories a to e The geodetic distance to each movement locus is infinite because it passes through the discontinuous point g 1 (e, f).
- the set of the movement trajectory i and the movement trajectory j whose geodetic distance does not become infinite is the same cluster, and when it is infinite, it is a different cluster.
- FIG. 17D it can be separated into two clusters of ⁇ 1 and ⁇ 2 . Further, as shown in FIG.
- the geodesic distance that is infinite calculated in the approximate geodesic distance calculating step (S207a) is g It is assumed that 2 (c, d), g 2 (e, f), and g 2 (f, g). In this case, 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 the case of clustering shown in FIG.
- a group in which the geodesic distance is infinite and a group in which the geodesic distance is not infinite are rearranged, and a total of four clusters of ⁇ 1 , ⁇ 2 , ⁇ 3 , ⁇ 4 To separate.
- a set of movement trajectories where the geodetic distance is not infinite can be determined as the same cluster by making it continuous, and a set of movement trajectories where the geodetic distance is infinite is made discontinuous, Clusters can be separated based on discontinuities.
- the region extraction candidate generation unit 1401 specifies the Nth smallest distance from the movement locus for each of the plurality of movement tracks calculated by the motion analysis unit 102, and increases the order of the specified plurality of distances. May be generated as a plurality of threshold values for generating a plurality of region extraction candidates.
- the region extraction candidate selection unit 1402 has the number of clusters closest to the preset number of classes (or the number of classes instructed from outside) from the plurality of clustering results performed in step S208a. Select the clustering result.
- the clustering result at the threshold R 2 (FIG. 17E) is selected.
- the clustering result when the threshold R 1 is selected (FIG. 17 (d))
- the output unit 108 outputs the result of labeling the movement trajectory .
- the number of thresholds is not limited to two, and any number of thresholds may be prepared. If there are not the same number of clustering results as the set number of classes, the cluster with the closest number of clusters may be selected, or the closest number of clusters out of the set number of classes. The clustering result that becomes the closest cluster number may be selected from among the class numbers larger than the set class number. Furthermore, by setting a plurality of classes and performing processing, the region extracted as the upper body in FIG. 17D is as shown in FIG. 17D and the clustering example shown in FIG. In FIG. 17E, it can be extracted as a head and an arm. That is, hierarchical region extraction can be performed. This has the effect of making it possible to extract a region that reflects the structure of the subject. For example, when analyzing motion or walking in sports, for example, moving the center of gravity of the region from detailed part analysis, depending on the number of classes. In addition, it is possible to simultaneously perform coarse subject motion analysis, such as motion analysis of only the upper body.
- clustering is performed for a plurality of threshold values based on discontinuous points calculated using geodetic distances, and by selecting the clustering result closest to the specified number of classes, detection and region extraction are performed as a result. It can be performed.
- clustering is performed based on the distance between pixels or the similarity of the movement trajectory, thereby collecting similar movement trajectories and extracting regions.
- the distance is close and the similar moving part is recognized as one region. It is possible to detect a moving body or a part of the moving body in a moving image and extract an area of an image including the moving body.
- clustering is performed using multiple thresholds based on the similarity of moving trajectories, which is particularly problematic when there are moving objects of various sizes in the image or when some of the moving objects are hidden. There is no detection error of the person candidate area, and the area extraction failure caused by the mistake.
- the region extraction that is, the detection of the moving object is performed at high speed without being affected by the posture or size of the moving object. It can be performed.
- the output unit 108a includes an image display unit 1701 in addition to the function of the output unit 108 in the first embodiment.
- the image display unit 1701 is a processing unit that displays the region extraction result obtained by the region extraction unit 107 on the moving image (picture) received by the image input unit 101, and includes, for example, a display such as an LCD and the like. It consists of the display control unit.
- FIGS. 19A and 19B are diagrams showing display examples by the image display unit 1701. FIG.
- each moving object may be displayed in different colors, or as shown in FIG. 19 (b), it may be displayed separately for each part. That is, as can be seen from FIG. 19, the mobile body detection device 100b according to the present modification can detect a plurality of mobile bodies by the above-described region extraction (FIG. 19 (a)), or one mobile body. Can be detected, that is, a region can be extracted for one moving body (FIG. 19B).
- the movement trajectory i shown in Equation 2 corresponds to one of the regions ⁇ m except for the outlier. Therefore, if a pixel on the picture is selected based on the movement trajectory i belonging to the region ⁇ m and color coding or the like is performed based on the region label m, the extracted region can be displayed on the image.
- the moving object detection device 100b in the fourth modification example it is possible to correctly extract a region even for a moving image including a person who moves while changing its shape, as described in the first embodiment.
- by displaying the extracted area on the moving image there is an effect that it becomes easy for the user to distinguish each moving body or a part of the moving body, and also its movement.
- an alarm unit may be provided to notify the extraction by sound instead of the image display. It is also possible to combine the alarm unit and the image display unit.
- the output unit 108b includes a recording / transmission unit 1901 in addition to the function of the output unit 108 in the first embodiment.
- the recording / transmission unit 1901 identifies the region in the picture received by the image input unit 101 based on the region extraction result in the region extraction unit 107, and for each identified region, the corresponding region extraction result is It is a processing unit that records (or holds) itself in a recording medium such as a built-in or external memory or a hard disk, or transmits it to the outside via a communication interface and a transmission path. That is, this recording / transmission unit 1901 separately records and transmits the result image obtained by extracting the area according to the area label ⁇ m as in the case of image display.
- motion information it is also possible to compress the motion information by averaging the motion information belonging to each area as follows. Normally, it is necessary to hold motion information for each pixel, but if the following processing is performed, it is only necessary to hold one motion for one region. For example, when the motion vector (u i t , v i t ) of the pixel i is used as the motion information, the motion information averaged for each extracted region can be calculated.
- C m is the number of pixels belonging to the region ⁇ m or the number of movement trajectories.
- FIG. 21 shows an example of the data structure in the case where the region extraction is performed with respect to the processing result of FIG. 19A by inputting t images from time T, and the segment region is recorded and transmitted using the result. And each region labeled theta m as an identifier, the pixel position and the pixel value of the picture at the time T that belong to theta 3 from each of the regions labeled theta 1, Then, the motion vector u m T from the time T corresponding to each region label to time T + t , V m T , u m T + 1 , v m T + 1 ,. . .
- each region label may be attached to one picture at time T and transmitted.
- an average value of pixel positions moved based on the affine motion may be calculated instead of the above equations 43 and 44.
- the moving object detection device 100c in the fifth modification it is possible to record and transmit an area while compressing pixel motion information.
- one piece of motion information is used for each region.
- the picture at time T is restored from the pixel position and pixel value corresponding to each region label shown in FIGS. 19 (a) and 19 (b). Furthermore, it is possible to restore the picture from time T + 1 to T + t by moving each pixel at time T using the pixel movement trajectory information.
- the restored image is the background image. May be overwritten. Accordingly, there is an effect that the information can be restored as a picture with a low calculation amount by using the information transmitted and recorded efficiently.
- Embodiment 2 Next, a moving body detection apparatus and a moving body detection method according to Embodiment 2 of the present invention will be described.
- a mobile body detection apparatus to which a function for predicting the movement of a mobile body is added based on the result of detection and area extraction of the mobile body using the mobile body detection method described in Embodiment 1. explain.
- FIG. 22 is a diagram illustrating a configuration of the moving object detection device 100d according to the second embodiment.
- the moving body detection apparatus 100d in the present embodiment includes an image input unit 101, a motion analysis unit 102, a region division unit 103, a distance calculation unit 104, a geodetic distance conversion unit 105, and an approximate geodetic distance calculation unit.
- 106 by adding a motion prediction unit 2101 to the region extraction unit 107 and the output unit 108c, it has a function of predicting the motion of the moving object.
- the motion prediction unit 2101 receives a region extraction result, calculates a representative trajectory from the movement trajectory of the pixels included in each region, and predicts the motion of the moving object based on the representative trajectory.
- the output unit 108c outputs information related to the position of the moving body and the position of the moving body part predicted by the motion prediction unit 2101 in addition to the function of the output unit 108 in the first embodiment.
- FIG. 23 shows a flowchart of processing according to the second embodiment. Since steps S201 to S208 are the same as those in the first embodiment, the description thereof is omitted.
- step S2201 the motion prediction unit 2101 uses the region extraction result performed in step S208 to obtain a cluster representative point and its representative locus as follows.
- the movement trajectory of the pixel belonging to the region ⁇ m is
- the motion prediction unit 2101 obtains a representative movement locus for each cluster region ⁇ m .
- an average movement locus is calculated as a representative movement locus.
- Weighting or the like may be performed every time, or a pixel movement locus corresponding to the center of gravity of the cluster on the image may be used as a representative movement locus.
- C m is the number of pixels belonging to the region ⁇ m or the number of pixel trajectories.
- FIG. 24 shows an example in which a representative movement trajectory is obtained for each cluster region ⁇ m based on Expression 45 above. However, for ease of viewing, only the representative movement trajectories relating to the cluster region ⁇ 1 corresponding to the head and the cluster region ⁇ 8 corresponding to the legs are shown in the figure. Each x in the figure corresponds to time t.
- the pixel position is shown. Furthermore, as shown in Equation 3 and Equation 4 above, region extraction by clustering in a non-linear space is performed in consideration of the similarity of pixel motion. Compared with the method for obtaining the above, since it is possible to calculate using only the movement trajectory of the pixel having similar motion, the representative movement trajectory can be obtained with higher accuracy. In this way, by obtaining a representative movement trajectory for each cluster region, it is possible to accurately and easily represent the movement of each part.
- step S2202 the motion prediction unit 2101 predicts the position of the moving body at a time earlier than time T from the representative movement trajectory calculated in step S2201.
- the representative movement trajectory is as follows:
- u m t is a motion vector and can be expressed as follows.
- the motion predicting unit 2101 for each moving body part, indicates the moving body part position pos m (at time T + t ′), as indicated by the dashed arrow and ⁇ in FIG. T + t ′) can be predicted as follows:
- the part of the moving body is taken as an example, but it is also possible to predict the position for each moving body from the detection example shown in FIG.
- the output unit 108c outputs the position of the moving body and the position of the moving body part predicted in step S2202. Thereby, prediction in consideration of acceleration is possible. There is an effect that the position of the moving body can be predicted by reflecting the acceleration when the movement is rapidly accelerated or stopped suddenly. Further, an affine parameter may be used instead of the motion vector. Since the affine parameter can express a motion including a rotational motion and is suitable for a rotational motion of an arm or a leg, the position of the joint object can be predicted more accurately.
- the moving object detection apparatus 100d in the present embodiment since the movement trajectories of pixels with similar motion can be calculated as the same cluster, the representative movement trajectory can be obtained with high accuracy. In particular, it is possible to express the movement of each part with respect to a joint object, etc., and it is possible to predict the position of a moving body part with high accuracy without setting a human candidate area as preprocessing. There is.
- a moving body detection apparatus will be described in which the moving body detection method described in Embodiment 1 is extended when a moving body is detected and a region is extracted from a plurality of camera images.
- FIG. 25 is a diagram illustrating a configuration of the moving object detection device 100e according to the third embodiment.
- the moving body detection apparatus 100e in the present embodiment includes an image input unit 101a having a multi-camera image input unit 2401, a motion analysis unit 102a, a region division unit 103a, a distance calculation unit 104a, and a geodetic distance conversion.
- Part 105b approximate geodesic distance calculation part 106b, area extraction part 107b, and output part 108d.
- the moving body detection apparatus 100e in the present embodiment basically has the same function as in the first embodiment.
- a description will be given focusing on differences from the first embodiment.
- the image input unit 101a includes a multi-camera image input unit 2401 in addition to the function of the image input unit 101 in the first embodiment.
- the multi-camera image input unit 2401 accepts input of video from a plurality of cameras and partially overlapping imaging areas.
- the shooting conditions of the video input to the multi-camera image input unit 2401 are not limited to the camera arrangement and number as shown in FIGS. 26 (a) and 26 (b). What is necessary is just to arrange
- the motion analysis unit 102a is a processing unit that calculates a movement trajectory for each block including one or more pixels constituting a picture obtained from each camera received by the multi-camera image input unit 2401, as in the first embodiment. It is. In addition, a plurality of camera images may be integrated based on the overlapping shooting area to form one image. In that case, since the integrated image can be processed as one image by the same method as in the first embodiment, the following description is omitted.
- the region dividing unit 103a is a processing unit that sets at least one of the movement trajectories in the overlapping photographing region as a shared point.
- the region dividing unit 103a divides a plurality of movement trajectories calculated for each of the moving images of the plurality of cameras so that the movement trajectory obtained from the image obtained from the camera p corresponds to the subset p.
- Each camera image may be divided into subsets as in the first embodiment. Since the processing in this case is the same as that of the first embodiment, the following description is omitted.
- the distance calculation unit 104a is a processing unit that calculates a distance representing the similarity between a plurality of movement trajectories for the movement trajectories calculated from images obtained from the respective cameras.
- a block of a block i included in the subset calculated by the motion analysis unit 102a and a block of a block included in the subset other than i are used. It is a process part which calculates the distance showing the similarity of a motion of. For example, when the movement trajectory of N blocks is used, the calculated distance is an N ⁇ N distance matrix.
- the distance between the blocks changes depending on the movement, in particular, an object such as a person that moves while changing its shape like a joint object. It is possible to express movement as a distance matrix.
- the geodetic distance conversion unit 105b is a processing unit that performs geodetic distance conversion on the distance matrix in the subset corresponding to each camera image calculated by the distance calculation unit 104a.
- the approximate geodetic distance calculation unit 106b integrates the geodetic distance matrix in each subset calculated by the geodetic distance conversion unit 105b, that is, the subset corresponding to each camera image, using the common points, that is, between the subsets. It is a processing unit that calculates an approximate geodetic distance that spans between camera images.
- the region extraction unit 107b is a processing unit that performs region extraction by specifying a region composed of blocks having similar movement trajectories based on the approximate geodetic distance calculated by the approximate geodetic distance calculation unit 106b. is there.
- the output unit 108d outputs a result of region extraction, that is, subject detection by integrating a plurality of camera images performed by the region extraction unit 107.
- step S201b the multi-camera image input unit 2401 accepts a plurality of pictures from a plurality of cameras as shown in FIGS. 26 (a) and 26 (b).
- step S202b the motion analysis unit 102a calculates a block motion from at least two pictures obtained from the respective cameras.
- the processing here is not described because it is only necessary to perform the processing in step S202 of the first embodiment for each picture obtained from each camera.
- step S203b the motion analysis unit 102a performs the same processing as that in step S203 described in the first embodiment using the motion information calculated in step S202b.
- the movement trajectory is calculated.
- the difference from the first embodiment is that the process of step S203 is performed for each of a plurality of camera images, and thus the following description is omitted.
- Equation 2 is rewritten as Equation 49 below to distinguish the movement trajectory obtained from each camera image.
- p_i is the movement trajectory i of the image obtained from each camera p.
- step S204b the area dividing unit 103a sets P subsets according to the camera image as shown in FIGS. 28 (a) and 28 (b).
- the subset is a set of movement trajectories obtained from each camera image.
- FIG. 28 (a) corresponds to FIG. 26 (a)
- FIG. 28 (b) corresponds to FIG. 26 (b).
- the area dividing unit 103a sets a shared point 2704 on the overlapping shooting area 2703 in the camera image 2701 and the camera image 2702.
- the number of cameras is not limited to two.
- the shared points 2704 are set on the respective overlapping photographing regions 2703. May be.
- the overlapping photographing area is an area on the image when the same place is photographed at least partially from a plurality of cameras placed at different positions. Further, at least one shared point may be set on the overlapping photographing area.
- the following processing may be performed by regarding the images obtained from the plurality of cameras p and the subset p obtained by dividing the corresponding movement trajectory.
- the same processing as that in step S204 of the first embodiment may be performed, and thus the description thereof is omitted.
- step S205b the distance calculation unit 104a calculates a distance matrix by regarding the movement locus calculated for each image obtained by the camera p in step S204b as a movement locus belonging to the subset p.
- the subsequent processing in this step may be performed considering the images obtained from the plurality of cameras p and their movement trajectories as the subset p divided in the first embodiment.
- step S205 of Embodiment 1 since the same processing as step S205 of Embodiment 1 may be performed, description thereof is omitted.
- step S206b the geodetic distance conversion unit 105b considers the movement trajectory calculated for each image obtained by the camera p in step S205b as a subset p and calculates a geodetic distance. Since the subsequent processing in this step is the same as that in step S206 in the first embodiment, description thereof is omitted.
- step S207b the approximate geodesic distance calculation unit 106b integrates the geodetic distance matrix calculated for each image obtained by the camera p in step S206b.
- the same processing as that in step S207 of the first embodiment may be performed using the shared points on the overlapping imaging region. Therefore, explanation is omitted.
- step S208b the region extraction unit 107b uses the geodetic distance integrated by the approximate geodetic distance calculation unit 106b to detect a region composed of blocks having similar movement trajectories by detecting discontinuous points. Region extraction is performed by specifying. Since the processing here is the same as that in step S208 of the first embodiment, description thereof is omitted.
- an image including a moving object such as a person who moves while changing a shape by inputting a plurality of camera images, or between different camera images. Even for a moving body that moves so as to straddle, it is possible to extract a region at high speed and correctly, that is, to detect the moving body, without being affected by the posture or size of the moving body.
- FIG. 29 is a diagram showing a configuration of a data classification device 2800 in the fourth embodiment.
- the data classification device 2800 includes a vector data input unit 2801, a region division unit 103, a distance calculation unit 104, a geodetic distance conversion unit 105, an approximate geodetic distance calculation unit 106, a data classification unit 2802, and an output. Part 2803.
- the multi-dimensional vector data such as the movement trajectory described above is input and the data is classified.
- the vector data input unit 2801 is a processing unit that receives vector data.
- the vector data may be, for example, data indicating a three-dimensional position obtained from a stereo camera, a CG, or the like, or may be data indicating the change over time.
- the area dividing unit 103 divides the multidimensional vector obtained by the vector data input unit 2801 into P subsets in the space of the multidimensional vector, as in the first modification of the first embodiment.
- a part of vector data included in one of the divided subsets (for example, at least one of the subsets with respect to the adjacent subset) is set as a common point.
- the configurations of the distance calculation unit 104, the geodetic distance conversion unit 105, and the approximate geodetic distance calculation unit 106 are the same as those of the first embodiment and the first modification of the first embodiment, the description thereof is omitted. That is, in these embodiments and modifications, the data to be processed is a “movement trajectory”, but the present embodiment is different only in that the data to be processed is “vector data”. The processing contents in the processing unit are the same.
- the data classification unit 2802 classifies data by clustering similar multi-dimensional vector data, similarly to the region extraction unit 107 in the first embodiment.
- the distance distribution between the multidimensional vectors is calculated based on the approximate geodetic distance obtained by the geodetic distance conversion unit 105 and the approximate geodetic distance calculation unit 106. Detect continuity and cluster multidimensional vectors that are continuously distributed based on discontinuous points so that multidimensional vectors separated by a distance smaller than the detected discontinuous points form one cluster. Thus, the multidimensional vector data is classified.
- the result of the data classification unit 2802 is written to a memory, a hard disk, or the like by the output unit 2803 or displayed on a display panel or the like.
- the constituent elements other than the input / output device such as the display device among the constituent elements are executed by the computer 1002 shown in FIG.
- Program and data that is, software
- hardware such as an electronic circuit, a memory, and a recording medium, or a mixture thereof.
- the vector data input unit 2801 receives a plurality of vector data.
- the vector data may be anything as long as it is a multidimensional vector composed of a plurality of elements as shown in Equation 2 above.
- the movement trajectory may be processed as multidimensional vector data in the first modification of the first embodiment.
- the data classification device 2800 in the fourth embodiment if not only the movement trajectory but also data representing a three-dimensional position obtained from a range finder or a stereo camera, for example, is input as a multidimensional vector, 3 A three-dimensional subject can also be classified based on the three-dimensional position.
- the input multidimensional vector data can be any vector data that can be analyzed by the multidimensional scaling method, and the geodetic distance conversion enables more accurate classification even for highly nonlinear data. And there is an effect that processing can be performed at high speed.
- FIGS. 31A to 31C show examples of figures that can be separated when the method of the first embodiment is used.
- the method of the first embodiment it is possible to extract regions into the graphic ⁇ 1 and the graphic ⁇ 2 , respectively.
- the condition is that all the pixels belonging to the graphic ⁇ 1 have the same movement, and all the pixels belonging to the graphic ⁇ 2 have the same movement.
- steps S205 and S207 in the first embodiment geodetic distance conversion having the characteristics shown in FIG. 6 is performed.
- a distance obtained by tracing (tracing) the movement locus from the movement locus can be calculated. Therefore, the distance along the shape can be calculated for the curved shape so as to be common to FIGS. 31 (a) to 31 (c).
- step S208 clustering is performed using a discontinuous point between the movement locus where the geodetic distance is infinite. For this reason, when there is a distance equal to or greater than the threshold between the movement trajectory, it is possible to extract the areas of the graphic ⁇ 1 and the graphic ⁇ 2 based on the discontinuous points.
- this embodiment it is possible to extract a region based on a discontinuous point when there is a moving object including a curved shape and there are certain discontinuous points.
- step S205 in order not to perform the process of tracing (tracing) data unless geodetic distance conversion is performed, it is determined whether or not the distance between adjacent movement trajectories is discontinuous or continuous. I can't. For this reason, in the example shown in FIG. 31, it is difficult to extract a region based on a discontinuous point when the mobile object includes a curved shape and has a certain discontinuous point.
- clustering considering the continuity regarding the similarity between the movement trajectories is performed by clustering using the geodetic distance as compared with the clustering using the Euclidean distance that is a linear distance. Even if the regions have complicated relations, it is reliably discriminated whether they belong to the same object (or part) or separate objects (or parts).
- the mobile body detection apparatus and method and the data classification apparatus and method according to the present invention have been described based on the embodiments and the modified examples thereof, the present invention is limited to these embodiments and modified examples. Is not to be done. Within the scope of the present invention, the embodiments realized by various modifications conceived by those skilled in the art for each embodiment, and any combination of the components in each embodiment and modification are realized. Forms are also included in the present invention.
- a moving body detection apparatus configured by adding a recording / transmission unit 1901 in the fifth modification and a motion prediction unit 2101 in the second embodiment to the fourth modification of the first embodiment is also included in the present invention. include.
- the mobile body detection device may include an area division unit 103b included in the mobile body detection device 103f shown in FIG. 32, instead of the area division unit in the above embodiment.
- the area dividing unit 103b includes two processing units (manually supporting the process of dividing the movement trajectory calculated by the motion analyzing unit 102 into a plurality of subsets.
- the manual setting unit 1031 divides the movement trajectory calculated by the motion analysis unit 102 into a plurality of subsets according to the space division specified by the user for the picture received by the image input unit 101. For example, when the user designates an area where a moving object is expected to exist (or moves) on a picture using a mouse or the like, the manual setting unit 1031 selects the area as one subset (that is, The movement trajectory corresponding to the block included in the area is divided into one subset). For example, when the flow line of the moving body is along a passage, the passage may be a single region.
- the automatic setting unit 1032 automatically divides the movement trajectory calculated by the motion analysis unit 102 into a plurality of subsets without the user's input.
- the automatic setting unit 1032 refers to the motion detected by the motion analysis unit 102 so that a region having a motion with a magnitude exceeding a preset threshold is a single region, and the other regions are a plurality of regions.
- the movement trajectory is automatically divided into a plurality of subsets by dividing into subsets.
- the motion detected by the motion analysis unit 102 may be accumulated over time so that a region with a large accumulated value becomes one region.
- the shooting area is known in advance. For this reason, it is often possible to specify in advance the region through which the moving object passes on the image. In such a case, there is an effect that a decrease in detection accuracy can be prevented.
- the region where the moving object exists is divided into a plurality of subsets by such an area dividing unit 103b, it is possible to avoid a risk that the accuracy of geodesic distance calculation is lowered. Can be prevented.
- the present invention is an apparatus for detecting all or part of a moving object in a moving image, that is, a person who moves while changing its shape based on the movement of a block composed of one or more pixels in a plurality of images.
- a moving body detection device that detects a moving body in an image by extracting an area including an image of the moving body
- a moving body detection device incorporated in an AV device such as a motion analysis device, a monitoring device, a video camera, or a TV Etc.
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Image Analysis (AREA)
Abstract
Description
まず、本発明の実施の形態1に係る移動体検出装置及び移動体検出方法について説明する。
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
次に、ステップS206で、測地距離変換部105は、上記式3もしくは上記式4で算出した部分集合pにおける線形距離fp(i,j)に対して閾値Rを用いて、以下のように非線形化処理を行いf'p(i,j)を算出する。
E.W.Dijkstra, "A note on two problems in connexion with graphs", Numerische Mathematik, pp.269-271, 1959
次に、本発明の実施の形態1の第1変形例におけ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1の第2変形例におけ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1の第3変形例におけ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1の第4変形例における移動体検出装置について説明する。ここでは、実施の形態1において移動体を検出及び部位を領域抽出した結果を画像として表示する機能を付加した例について説明する。このような第4変形例に係る移動体検出装置100bは、図18に示す構成図のように、基本的には実施の形態1と同じ構成を備えるが、実施の形態1における出力部108に代えて、画像表示部1701を有する出力部108aを備え、この特徴により、領域抽出部107によって抽出した領域を画像としてモニタ等に表示することが可能である。
次に、本発明の実施の形態1の第5変形例における移動体検出装置について説明する。ここでは、実施の形態1において移動体を検出及び部位を領域抽出した結果をそれぞれ抽出した領域ごとに記録・送信する機能を付加した例について説明する。このような第5変形例に係る移動体検出装置100cは、図20に示す構成図のように、基本的には実施の形態1と同じ構成を備えるが、実施の形態1における出力部108に代えて、記録・送信部1901を有する出力部108bを備える。
次に、本発明の実施の形態2に係る移動体検出装置及び移動体検出方法について説明する。
次に、本発明の実施の形態3に係る移動体検出装置及び移動体検出方法について説明する。
次に、本発明の実施の形態4に係る移動体検出装置及びその方法の応用例であるデータ分類装置及びその方法について説明する。
ここでは、実施の形態1で説明した方法を用いた場合に、領域抽出可能な移動体の性質について補足する。図31(a)~図31(c)に実施の形態1の方法を用いた場合に分離可能な図形の一例を示す。実施の形態1の方法を用いることで、それぞれ図形θ1と図形θ2とに領域抽出することが可能である。ここでは、図31(a)から図31(c)に示したそれぞれの図形θ1と図形θ2の移動方向は、同一であっても、異なっていてもかまわない。ただし、図形θ1に属する画素はすべて同じ動きであり、かつ図形θ2に属する画素はすべて同じ動きであることが条件である。
101、101a 画像入力部
102、102a 動き解析部
103、103a、103b 領域分割部
104、104a 距離計算部
105、105a、105b 測地距離変換部
106、106a、106b 近似測地距離算出部
107、107a、107b 領域抽出部
108、108a~108d、2803 出力部
1031 手動設定部
1032 自動設定部
1401 領域抽出候補生成部
1402 領域抽出候補選択部
1701 画像表示部
1901 記録・送信部
2101 動き予測部
2401 複数カメラ画像入力部
2800 データ分類装置
2801 ベクトルデータ入力部
2802 データ分類部
Claims (21)
- 動画像中の移動体の全部又は一部の領域を抽出することによって動画像中の移動体を検出する移動体検出装置であって、
動画像を構成する複数枚のピクチャを受け付ける画像入力部と、
受け付けられた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ごとに、異なる2枚のピクチャ間での画像の動きを検出し、検出した動きを前記複数枚のピクチャについて連結した移動軌跡を複数算出する動き解析部と、
算出された前記複数の移動軌跡を複数の部分集合に分割するとともに、分割された複数の部分集合において一部の移動軌跡を共有点として設定する領域分割部と、
分割された前記複数の部分集合のそれぞれについて、複数の移動軌跡間の類似性を表す距離を算出する距離計算部と、
算出された前記距離を、測地距離に変換する測地距離変換部と、
変換された前記測地距離のうち、前記共有点を共有する測地距離を統合することで、前記部分集合にまたがる近似的な測地距離を算出する近似測地距離算出部と、
算出された近似的な測地距離を用いて、類似する移動軌跡をもつブロックどうしを同一の領域として特定するクラスタリングを行うことによって、前記動画像に対して少なくとも一つの領域を抽出する領域抽出部と
を備える移動体検出装置。 - 前記画像入力部は、複数のカメラによる複数の動画像のそれぞれについて、前記複数枚のピクチャを受け付け、
前記動き解析部は、前記複数の動画像のそれぞれについて、前記複数の移動軌跡を算出し、
前記領域分割部は、前記複数の動画像のそれぞれについて算出された複数の移動軌跡を、分割された前記複数の部分集合として保持し、
前記領域抽出部は、前記クラスタリングを行うことで、前記複数の動画像に対して少なくとも一つの領域を抽出する
請求項1記載の移動体検出装置。 - 前記領域分割部は、移動軌跡に対応する前記ピクチャ上でのブロックの画像上での位置に関する空間上で、前記複数の移動軌跡を前記複数の部分集合に分割する
請求項1記載の移動体検出装置。 - 前記領域分割部は、前記ピクチャに対してユーザが指定した空間分割に従って、前記複数の移動軌跡を前記複数の部分集合に分割する
請求項3記載の移動体検出装置。 - 前記移動軌跡は、多次元ベクトルで表現され、
前記領域分割部は、前記多次元ベクトルを表現する多次元ベクトル空間上で、前記複数の移動軌跡を前記複数の部分集合に分割する
請求項1記載の移動体検出装置。 - 前記領域分割部は、近傍の部分集合の一部が重複するように前記分割をするとともに、重複した領域に含まれる移動軌跡を前記共有点として設定する
請求項3~5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、前記複数の部分集合のそれぞれについて、当該部分集合に属する移動軌跡のうち、他の部分集合との境界に近接する移動軌跡を、前記共有点として設定する
請求項3~5のいずれか1項に記載の移動体検出装置。 - 前記測地距離変換部は、前記距離計算部で算出された距離のうち、予め定められた条件を満たす小さい距離を連結することで、前記動き解析部で算出された一の移動軌跡から他の一の移動軌跡にたどりつく最短経路を求めることにより、前記距離計算部で算出された距離のそれぞれを測地距離に変換する
請求項1記載の移動体検出装置。 - 前記測地距離変換部は、前記複数の部分集合に含まれる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、小さい順に予め定められた個数の距離を選択し、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、前記最短経路を求めることにより、前記距離計算部で算出された距離のそれぞれを測地距離に変換する
請求項8記載の移動体検出装置。 - 前記測地距離変換部は、前記複数の部分集合に含まれる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、予め定められた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、前記最短経路を求めることにより、前記距離計算部で算出された距離のそれぞれを測地距離に変換する
請求項8記載の移動体検出装置。 - 前記領域抽出部は、前記近似測地距離算出部で算出された測地距離の分布における少なくとも一つ以上の不連続点を検出し、検出した不連続点よりも小さい測地距離だけ離れた移動軌跡どうしが一つのクラスタとなるように、前記クラスタリングを行う
請求項1記載の移動体検出装置。 - 前記領域抽出部は、前記近似測地距離算出部で算出された測地距離に対して、固有値および固有ベクトルを求めることによって次元圧縮を行い、次元圧縮した空間上で前記クラスタリングを行う
請求項1記載の移動体検出装置 - 前記測地距離変換部は、前記距離を前記測地距離に変換するための複数の判断基準を生成し、生成した複数の判断基準のそれぞれについて、当該判断基準を用いて前記距離を前記測地距離に変換することで、前記複数の判断基準のそれぞれに対応する測地距離を生成し、
前記近似測地距離算出部は、前記複数の判断基準のそれぞれに対応する測地距離に対して前記統合を行い、
前記領域抽出部は、
前記複数の判断基準のそれぞれに対応する統合された近似的な測地距離に対して前記クラスタリングを行って領域を抽出することで、前記複数の判断基準のそれぞれに対応させて、前記領域抽出の結果を領域抽出候補として生成する領域抽出候補生成部と、
クラス数についての指示を取得し、取得したクラス数に近い個数の領域が抽出された領域抽出候補を、前記領域抽出候補生成部で生成された複数の領域抽出候補から選択し、選択した領域抽出候補を、当該領域抽出部による領域抽出の結果として出力する領域抽出候補選択部とを有する
請求項1に記載の移動体検出装置。 - 前記測地距離変換部は、前記複数の判断基準として、複数の閾値を生成し、生成した複数の閾値のそれぞれについて、前記距離計算部で算出された距離のうち、当該閾値よりも小さい距離を連結することで、前記複数の判断基準のそれぞれに対応する測地距離を生成する
請求項13記載の移動体検出装置。 - 前記移動体検出装置はさらに、前記領域抽出部で得られた領域抽出の結果を、前記画像入力部で受け付けたピクチャに重ねて表示する画像表示部を備える
請求項1~14のいずれか1項に記載の移動体検出装置。 - 前記移動体検出装置はさらに、前記領域抽出部での領域抽出の結果に対応させて、前記画像入力部で受け付けたピクチャにおける領域を特定し、特定した領域ごとに、対応する領域抽出の結果を、記録して保持する、又は、伝送路を介して外部に送信する記録・送信部を備える
請求項1~15のいずれか1項に記載の移動体検出装置。 - 前記画像入力部は、2つ以上の移動体が含まれる動画像を受け付け、
前記領域抽出部では、前記2以上の移動体について前記領域抽出をすることで、2以上の移動体を検出する
請求項1~16のいずれか1項に記載の移動体検出装置。 - 前記移動体検出装置はさらに、前記領域抽出部で抽出された領域に含まれるブロックの移動軌跡から、当該領域を代表する移動軌跡を算出し、算出した移動軌跡に従って当該領域が移動すると予測することで、前記移動体の動きを予測する動き予測部を備える
請求項1~17のいずれか1項に記載の移動体検出装置。 - 複数のベクトルデータを、類似するベクトルデータの集まりであるクラスに分類するベクトルデータ分類装置であって、
複数のベクトルデータを受け付けるベクトルデータ入力部と、
受け付けられた前記複数のベクトルデータを、測地距離の計算に利用される複数の部分集合に分割するとともに、分割された複数の部分集合の少なくとも一つに含まれる一部のベクトルデータを共有点として設定する領域分割部と、
分割された前記複数の部分集合のそれぞれについて、複数のベクトルデータ間の類似性を表す距離を算出する距離計算部と
算出された前記距離を、中継点としての移動軌跡をたどりながら一の移動軌跡から他の一の移動軌跡にたどりつく経路の距離である測地距離に変換する測地距離変換部と
変換された前記測地距離のうち、前記共有点を共有する測地距離を統合することで、前記部分集合にまたがる近似的な測地距離を算出する近似測地距離算出部と、
算出された近似的な測地距離を用いて、類似するベクトルデータをもつブロックどうしを一つの領域として特定するクラスタリングを行うことによって、前記動画像に対して少なくとも一つのクラスを生成するデータ分類部と
を備えるベクトルデータ分類装置。 - 動画像中の移動体の全部又は一部の領域を抽出することによって動画像中の移動体を検出する移動体検出方法であって、
動画像を構成する複数枚のピクチャを受け付ける画像入力ステップと、
受け付けられた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ごとに、異なる2枚のピクチャ間での画像の動きを検出し、検出した動きを前記複数枚のピクチャについて連結した移動軌跡を複数算出する動き解析ステップと、
算出された前記複数の移動軌跡を複数の部分集合に分割するとともに、分割された複数の部分集合において一部の移動軌跡を共有点として設定する領域分割ステップと、
分割された前記複数の部分集合のそれぞれについて、複数の移動軌跡間の類似性を表す距離を算出する距離計算ステップと、
算出された前記距離を、中継点としての移動軌跡をたどりながら一の移動軌跡から他の一の移動軌跡にたどりつく経路の距離である測地距離に変換する測地距離変換ステップと、
変換された前記測地距離のうち、前記共有点を共有する測地距離を統合することで、前記部分集合にまたがる近似的な測地距離を算出する近似測地距離算出ステップと、
算出された近似的な測地距離を用いて、類似する移動軌跡をもつブロックどうしを同一の領域として特定するクラスタリングを行うことによって、前記動画像に対して少なくとも一つの領域を抽出する領域抽出ステップと
を含む移動体検出方法。 - 動画像中の移動体の全部又は一部の領域を抽出することによって動画像中の移動体を検出するためのプログラムであって、
請求項20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
プログラム。
Priority Applications (4)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
EP10804052.8A EP2461292B1 (en) | 2009-07-31 | 2010-07-05 | Mobile body detection apparatus and mobile body detection method |
JP2010536277A JP4643766B1 (ja) | 2009-07-31 | 2010-07-05 | 移動体検出装置及び移動体検出方法 |
CN201080003380.4A CN102227750B (zh) | 2009-07-31 | 2010-07-05 | 移动体检测装置及移动体检测方法 |
US12/976,169 US8300892B2 (en) | 2009-07-31 | 2010-12-22 | Moving object detection apparatus and moving object detection method |
Applications Claiming Priority (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
JP2009-178788 | 2009-07-31 | ||
JP2009178788 | 2009-07-31 |
Related Child Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
US12/976,169 Continuation US8300892B2 (en) | 2009-07-31 | 2010-12-22 | Moving object detection apparatus and moving object detection method |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2011013299A1 true WO2011013299A1 (ja) | 2011-02-03 |
Family
ID=43528973
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/JP2010/004378 WO2011013299A1 (ja) | 2009-07-31 | 2010-07-05 | 移動体検出装置及び移動体検出方法 |
Country Status (5)
Country | Link |
---|---|
US (1) | US8300892B2 (ja) |
EP (1) | EP2461292B1 (ja) |
JP (1) | JP4643766B1 (ja) |
CN (1) | CN102227750B (ja) |
WO (1) | WO2011013299A1 (ja) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102739957A (zh) * | 2011-03-31 | 2012-10-17 | 卡西欧计算机株式会社 | 能够确定被摄体的运动的图像处理装置及图像处理方法 |
CN106354333A (zh) * | 2016-09-30 | 2017-01-25 | 南京仁光电子科技有限公司 | 一种提高触控系统跟踪精确度的方法 |
JP2018124689A (ja) * | 2017-01-31 | 2018-08-09 | 株式会社日立製作所 | 移動物体検出装置、移動物体検出システム、及び移動物体検出方法 |
KR20210146109A (ko) * | 2020-05-26 | 2021-12-03 | 동명대학교산학협력단 | 다중 영상데이터의 학습을 통한 연속적 객체 위치 추적 장치 및 그 방법 |
CN114004645A (zh) * | 2021-10-29 | 2022-02-01 | 浙江省民营经济发展中心(浙江省广告监测中心) | 融媒体广告智慧监测平台和电子设备 |
WO2022102038A1 (ja) | 2020-11-12 | 2022-05-19 | 富士通株式会社 | 情報処理装置、生成方法、および生成プログラム |
KR20230122297A (ko) * | 2022-02-14 | 2023-08-22 | 이정호 | 인공지능 기반의 수중 위치 보정 방법 |
Families Citing this family (28)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102378992B (zh) * | 2009-12-28 | 2014-12-24 | 松下电器产业株式会社 | 关节状区域检测装置及其方法 |
EP2447882B1 (en) * | 2010-10-29 | 2013-05-15 | Siemens Aktiengesellschaft | Method and device for assigning sources and sinks to routes of individuals |
US8855361B2 (en) * | 2010-12-30 | 2014-10-07 | Pelco, Inc. | Scene activity analysis using statistical and semantic features learnt from object trajectory data |
AU2011265430B2 (en) * | 2011-12-21 | 2015-03-19 | Canon Kabushiki Kaisha | 3D reconstruction of partially unobserved trajectory |
US10095954B1 (en) * | 2012-01-17 | 2018-10-09 | Verint Systems Ltd. | Trajectory matching across disjointed video views |
TWI588778B (zh) * | 2012-01-17 | 2017-06-21 | 國立臺灣科技大學 | 動作辨識方法 |
US9336302B1 (en) | 2012-07-20 | 2016-05-10 | Zuci Realty Llc | Insight and algorithmic clustering for automated synthesis |
TWI482468B (zh) * | 2012-11-23 | 2015-04-21 | Inst Information Industry | 物體偵測裝置、方法及其電腦可讀取紀錄媒體 |
CN103841312B (zh) * | 2012-11-27 | 2017-03-01 | 财团法人资讯工业策进会 | 物体侦测装置及方法 |
JP6381918B2 (ja) * | 2013-01-23 | 2018-08-29 | キヤノンメディカルシステムズ株式会社 | 動作情報処理装置 |
JP6072892B2 (ja) * | 2013-02-27 | 2017-02-01 | 日立オートモティブシステムズ株式会社 | 撮像装置及び車両制御装置 |
JP6110174B2 (ja) * | 2013-03-26 | 2017-04-05 | 株式会社メガチップス | 画像検出装置及び制御プログラム並びに画像検出方法 |
US10235605B2 (en) * | 2013-04-10 | 2019-03-19 | Microsoft Technology Licensing, Llc | Image labeling using geodesic features |
US10191956B2 (en) * | 2014-08-19 | 2019-01-29 | New England Complex Systems Institute, Inc. | Event detection and characterization in big data streams |
GB2523253B (en) * | 2015-01-23 | 2017-04-12 | Visidon Oy | Image processing method |
US9990857B2 (en) * | 2015-07-01 | 2018-06-05 | Grafty, Inc. | Method and system for visual pedometry |
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 |
US11205103B2 (en) | 2016-12-09 | 2021-12-21 | The Research Foundation for the State University | Semisupervised autoencoder for sentiment analysis |
CN107015017B (zh) * | 2017-03-02 | 2020-08-11 | 惠州Tcl移动通信有限公司 | 一种基于速度计算的速度提示方法及系统 |
US10733755B2 (en) * | 2017-07-18 | 2020-08-04 | Qualcomm Incorporated | Learning geometric differentials for matching 3D models to objects in a 2D image |
JP6943294B2 (ja) * | 2017-12-14 | 2021-09-29 | 富士通株式会社 | 技認識プログラム、技認識方法および技認識システム |
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 |
CN109063714A (zh) * | 2018-08-06 | 2018-12-21 | 浙江大学 | 基于深度神经网络的帕金森病运动迟缓视频检测模型的构建方法 |
US11488374B1 (en) * | 2018-09-28 | 2022-11-01 | Apple Inc. | Motion trajectory tracking for action detection |
US10937169B2 (en) * | 2018-12-18 | 2021-03-02 | Qualcomm Incorporated | Motion-assisted image segmentation and object detection |
CN110426178B (zh) * | 2019-07-23 | 2020-09-29 | 中国科学院软件研究所 | 一种基于尾流示踪的风场测量方法及系统 |
CN116648596A (zh) * | 2020-12-22 | 2023-08-25 | 松下知识产权经营株式会社 | 成像装置 |
CN115082545B (zh) * | 2022-06-08 | 2023-02-07 | 国网黑龙江省电力有限公司大庆供电公司 | 应用于电力现场的安全系统 |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH08214289A (ja) | 1994-12-06 | 1996-08-20 | Olympus Optical Co Ltd | 時系列画像解析装置及びその解析方法 |
JPH1166319A (ja) * | 1997-08-21 | 1999-03-09 | Omron Corp | 移動体検出方法及び装置並びに移動体認識方法及び装置並びに人間検出方法及び装置 |
JP4456181B1 (ja) * | 2008-10-27 | 2010-04-28 | 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 |
Family Cites Families (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP3612360B2 (ja) * | 1995-04-10 | 2005-01-19 | 株式会社大宇エレクトロニクス | 移動物体分割法を用いた動画像の動き推定方法 |
US6643387B1 (en) * | 1999-01-28 | 2003-11-04 | Sarnoff Corporation | Apparatus and method for context-based indexing and retrieval of image sequences |
JP4226730B2 (ja) * | 1999-01-28 | 2009-02-18 | 株式会社東芝 | 物体領域情報生成方法及び物体領域情報生成装置並びに映像情報処理方法及び情報処理装置 |
JP2000222584A (ja) * | 1999-01-29 | 2000-08-11 | Toshiba Corp | 映像情報記述方法、映像検索方法及び映像検索装置 |
US6859554B2 (en) * | 2001-04-04 | 2005-02-22 | Mitsubishi Electric Research Laboratories, Inc. | Method for segmenting multi-resolution video objects |
US20030123704A1 (en) * | 2001-05-30 | 2003-07-03 | Eaton Corporation | Motion-based image segmentor for occupant tracking |
US8614741B2 (en) * | 2003-03-31 | 2013-12-24 | Alcatel Lucent | Method and apparatus for intelligent and automatic sensor control using multimedia database system |
CN101334845B (zh) * | 2007-06-27 | 2010-12-22 | 中国科学院自动化研究所 | 一种基于轨迹序列分析和规则归纳的视频行为识别方法 |
US8212210B2 (en) * | 2008-02-04 | 2012-07-03 | Flir Systems Ab | IR camera and method for presenting IR information |
-
2010
- 2010-07-05 EP EP10804052.8A patent/EP2461292B1/en not_active Not-in-force
- 2010-07-05 CN CN201080003380.4A patent/CN102227750B/zh not_active Expired - Fee Related
- 2010-07-05 WO PCT/JP2010/004378 patent/WO2011013299A1/ja active Application Filing
- 2010-07-05 JP JP2010536277A patent/JP4643766B1/ja not_active Expired - Fee Related
- 2010-12-22 US US12/976,169 patent/US8300892B2/en not_active Expired - Fee Related
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JPH08214289A (ja) | 1994-12-06 | 1996-08-20 | Olympus Optical Co Ltd | 時系列画像解析装置及びその解析方法 |
JPH1166319A (ja) * | 1997-08-21 | 1999-03-09 | Omron Corp | 移動体検出方法及び装置並びに移動体認識方法及び装置並びに人間検出方法及び装置 |
JP4456181B1 (ja) * | 2008-10-27 | 2010-04-28 | 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 |
Non-Patent Citations (8)
Title |
---|
E. W. DIJKSTRA: "A note on two problems in connexion with graphs", NUMERISCHE MATHEMATIK, 1959, pages 269 - 271, XP000837527, DOI: doi:10.1007/BF01386390 |
JIANBO SHI, CARLO TOMASI: "Good Features to Track", IEEE CONFERENCE ON COMPUTER VISION AND PATTERN RECOGNITION, 1994, pages 593 - 600, XP010099247, DOI: doi:10.1109/CVPR.1994.323794 |
JOSHUA TENENBAUM, VIN DE SILVA, JOHN LANGFORD: "A Global Geometric Framework for Nonlinear Dimensionality Reduction", SCIENCE, vol. 290, 22 December 2000 (2000-12-22), pages 2319 - 2322 |
P. ANANDAN: "A Computational Framework and an Algorithm for the Measurement of Visual Motion", INTERNATIONAL JOURNAL OF COMPUTER VISION, vol. 2, 1989, pages 283 - 310, XP008055537, DOI: doi:10.1007/BF00158167 |
See also references of EP2461292A4 |
VIN DE SILVA, JOSHUA B, TENENBAUM: "Global Versus Local Methods in Nonlinear Dimensionality Reduction", NEURAL INFORMATION PROCESSING SYSTEMS, vol. 15, 2002, pages 705 - 712 |
VIN DE SILVA, JOSHUA B. TENENBAUM: "Sparse Multidimensional Scaling using Landmark Points", TECHNICAL REPORT, STANFORD UNIVERSITY, June 2004 (2004-06-01) |
VLADIMIR KOLMOGOROV, RAMIN ZABIH: "Computing Visual Correspondence with Occlusions via Graph Cuts", INTERNATIONAL CONFERENCE ON COMPUTER VISION, 2001 |
Cited By (9)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102739957A (zh) * | 2011-03-31 | 2012-10-17 | 卡西欧计算机株式会社 | 能够确定被摄体的运动的图像处理装置及图像处理方法 |
CN106354333A (zh) * | 2016-09-30 | 2017-01-25 | 南京仁光电子科技有限公司 | 一种提高触控系统跟踪精确度的方法 |
JP2018124689A (ja) * | 2017-01-31 | 2018-08-09 | 株式会社日立製作所 | 移動物体検出装置、移動物体検出システム、及び移動物体検出方法 |
KR20210146109A (ko) * | 2020-05-26 | 2021-12-03 | 동명대학교산학협력단 | 다중 영상데이터의 학습을 통한 연속적 객체 위치 추적 장치 및 그 방법 |
KR102425213B1 (ko) | 2020-05-26 | 2022-07-25 | 동명대학교 산학협력단 | 다중 영상데이터의 학습을 통한 연속적 객체 위치 추적 장치 및 그 방법 |
WO2022102038A1 (ja) | 2020-11-12 | 2022-05-19 | 富士通株式会社 | 情報処理装置、生成方法、および生成プログラム |
CN114004645A (zh) * | 2021-10-29 | 2022-02-01 | 浙江省民营经济发展中心(浙江省广告监测中心) | 融媒体广告智慧监测平台和电子设备 |
KR20230122297A (ko) * | 2022-02-14 | 2023-08-22 | 이정호 | 인공지능 기반의 수중 위치 보정 방법 |
KR102601155B1 (ko) * | 2022-02-14 | 2023-11-10 | 이정호 | 인공지능 기반의 수중 위치 보정 방법 |
Also Published As
Publication number | Publication date |
---|---|
US20110091073A1 (en) | 2011-04-21 |
JP4643766B1 (ja) | 2011-03-02 |
CN102227750B (zh) | 2014-06-25 |
JPWO2011013299A1 (ja) | 2013-01-07 |
EP2461292B1 (en) | 2018-09-26 |
CN102227750A (zh) | 2011-10-26 |
US8300892B2 (en) | 2012-10-30 |
EP2461292A4 (en) | 2013-07-24 |
EP2461292A1 (en) | 2012-06-06 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
JP4643766B1 (ja) | 移動体検出装置及び移動体検出方法 | |
JP4456181B1 (ja) | 移動体検出方法及び移動体検出装置 | |
JP4782901B2 (ja) | 移動体検出装置および移動体検出方法 | |
JP4542207B1 (ja) | 移動体検出方法および移動体検出装置 | |
JP5054206B2 (ja) | 画像処理方法及び画像処理装置 | |
US8705861B2 (en) | Context processor for video analysis system | |
JP5404918B2 (ja) | 移動軌跡算出方法およびその装置、ならびに領域分割方法 | |
JP5001260B2 (ja) | オブジェクト追跡方法及びオブジェクト追跡装置 | |
JP4668360B2 (ja) | 移動体検出方法及び移動体検出装置 | |
Benedek et al. | An integrated 4D vision and visualisation system | |
KR102629213B1 (ko) | 경사 촬영 동영상 이동물체 탐지 방법 및 장치 | |
JP4966946B2 (ja) | 時空間画像分離装置、時空間画像分離方法および時空間画像分離プログラム | |
JP4660736B2 (ja) | 動画像時間分割処理方法および装置 | |
Khazaeli et al. | Kinect applications in construction: From tracking to reconstruction | |
Benedek et al. | 4D Scene Reconstruction in Multi-Target Scenarios |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
WWE | Wipo information: entry into national phase |
Ref document number: 201080003380.4 Country of ref document: CN |
|
ENP | Entry into the national phase |
Ref document number: 2010536277 Country of ref document: JP Kind code of ref document: A |
|
121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 10804052 Country of ref document: EP Kind code of ref document: A1 |
|
WWE | Wipo information: entry into national phase |
Ref document number: 2010804052 Country of ref document: EP |
|
NENP | Non-entry into the national phase |
Ref country code: DE |