JP2013077202A - Moving object tracking device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problem that a device for tracking a moving object within a monitoring space on a monitoring image does not easily perform processing for preventing mistake in the case that a target object approaches other congesting objects.SOLUTION: A hypothesis setting part 51 predicts a plurality of candidate positions at a current time from past positions for a plurality of objects. A likelihood calculation part 52 calculates an evaluation value representing the degree of movement destination likelihood of a target object on the basis of an image feature for each movement destination candidate of a target object. Further, correction for lowering the evaluation value is performed in accordance with abundance representing existence probability of other objects in the movement destination candidate. On the other hand, evaluation value compensation for weakening the correction is performed for a part of the movement destination candidate of the target object, and the concession of the other objects is reflected on the evaluation value. A movement destination position of the target object is determined from the correction and the evaluation value subjected to limited compensation.

Description

  The present invention relates to a moving object tracking apparatus that tracks a plurality of moving objects by image processing.

  As a conventional technique for tracking a moving object by image processing, there is known a technique for searching for and tracking a place where image features are similar using a template matching process. In the template matching process, if the entire image is set as the search range, the processing load increases. Therefore, the search range is around the position of the moving object of interest at the previous time or around the predicted position at the current time based on past motion or the like. Set. Then, the degree of coincidence between the input image and the template is examined within the limited search range, and it is determined that there is a moving object that is focused on the most coincident position.

  Here, when there are multiple moving objects to be tracked in the tracking area, the search range of those templates may overlap when moving objects with similar image features approach each other. May match most templates in position. In this case, mistracking may occur, such as mistaking them and starting tracking, or determining that a plurality of moving objects are present at the same position.

  As a countermeasure, after determining the position candidate of each object at the position where the respective templates of the plurality of moving objects most closely match, if the distance between the position candidates is too close, a correction is made to release and based on the degree of coincidence with the template There is a conventional technique for performing reevaluation (Patent Document 1).

JP 2009-48428 A

  However, in the prior art, for example, when the target object passes by a plurality of other objects that are closely packed in the direction of travel, the correction process that separates the target object from the other object results in a mismatch with the actual movement of the object. There is a problem that the tracking object may be mistaken.

  More specifically, for example, when a person of interest approaches a plurality of people standing side by side in front of a narrow passage, or when a person of interest approaches a place where people gather, The situation changes as a person in the direction of travel gives way, and the person of interest may pass between these persons. In this case, if the correction is performed to release each approaching object, the position of the person of interest is corrected in the direction opposite to the direction of travel, while other than the person of interest is corrected in the direction of travel of the person of interest. It becomes easy to mix up with other people. In other words, in the prior art, the search range of persons close to a plurality of dense persons is limited to the pushed-back position, so that there is a problem that the passing through is mistracked.

  On the other hand, the attention person does not pass between the dense people who maintain the movement so far, but after approaching, they join the circle of those people and stop, or the attention person avoids the group of people. The movement of the person of interest may change. In such a case, control for separating the search position is effective in order to avoid misunderstanding.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a moving object tracking device that can track with high accuracy without mistaking them even when a plurality of moving objects are densely packed.

  A moving object tracking device according to the present invention tracks a plurality of objects moving in the space using time-series images obtained by photographing a predetermined space, and each object in the past from the attention time A storage unit that stores position information, a position prediction unit that predicts and distributes the movement destination candidates of each object at the time of interest from the position information, and each of the movement destination candidates in the image at the time of interest. An evaluation value calculation unit that obtains an evaluation value of the movement destination likelihood for the movement destination candidate from the image feature of the position; an object position determination unit that determines the movement destination position of each object based on the height of the evaluation value; The evaluation value calculation unit includes one of the objects as a target object, and the evaluation value for the destination candidate of the target object predicted within the distribution range of the destination candidate of other objects. Corrects lower predetermined value content, this time for some candidates of the destination candidates in the distribution range to reflect the concession of the by reducing the predetermined value other objects in the evaluation value.

  In the moving object tracking device according to the present invention, when the evaluation value calculation unit determines that the target object is moving from the position information, the evaluation value calculation unit determines the destination candidate predicted ahead of the target object. It can be set as the structure made into a part candidate.

  In the moving object tracking device according to the present invention, the evaluation value calculation unit, when it is determined from the position information that the object of interest is moving, the movement destination candidate that is closer to the traveling direction of the object of interest. It can be set as the structure which makes a value small.

  In the moving object tracking device according to the present invention described above, the evaluation value calculation unit, when it is determined from the position information that the object of interest is moving, the movement destination candidate closer to the closest position of the object of interest. It can be set as the structure which makes a value small.

  In the moving object tracking device according to the present invention, the evaluation value calculation unit obtains the partial candidates from the movement destination candidates within the distribution range when the target object is determined to be stopped from the position information. It can be set as the structure selected at random.

  According to the present invention, by performing correction that lowers the evaluation value as the movement destination position with respect to the candidate position of the movement destination of the object of interest according to the probability of existence of other objects, the movement destination positions of a plurality of approaching objects are separated. . At that time, by partially weakening the correction for lowering the evaluation value to compensate the evaluation value, it is possible to track a moving object that passes through the crowd. By limiting the compensation of the evaluation value with respect to the correction, the effect of the correction can be ensured, and even if there is a movement other than passing through, the confusion can be prevented. That is, even when a plurality of moving objects are densely packed, they can be tracked with high accuracy without mistaking them.

It is a block block diagram of the moving object tracking apparatus which concerns on embodiment of this invention. It is a perspective view which shows an example of a three-dimensional model typically. As an example in which a plurality of moving objects exist in the monitoring space, a schematic diagram showing a state in which four persons are approaching and a presence map set for three of the moving objects other than the target moving object. is there. It is a schematic diagram explaining the method of producing an individual presence map. It is a schematic diagram explaining the setting of the relaxation value E when there are other objects that are densely packed in the traveling direction of the target object in the moving state. It is a schematic diagram explaining the setting of the relaxation value E when there is no other object in the traveling direction of the target object in the moving state. It is a schematic diagram explaining the setting of the relaxation value E when an attention object is a stop state. It is a general | schematic flowchart of the tracking process of the moving object tracking apparatus which concerns on embodiment of this invention. FIG. 9 is a schematic flowchart of likelihood correction processing in the tracking processing of FIG. 8.

  Hereinafter, a moving object tracking device 1 according to an embodiment of the present invention (hereinafter referred to as an embodiment) will be described with reference to the drawings. The moving object tracking device 1 uses a person moving in the monitoring space as a tracking target (hereinafter referred to as a moving object), and processes the monitoring image obtained by imaging the monitoring space to track the moving object. The moving object tracking device 1 is configured to be able to track a plurality of moving objects existing in the monitoring space. The monitoring space is not limited to being indoors but may be outdoors, and the moving object may be a person other than a person such as a vehicle.

[Configuration of moving object tracking device]
FIG. 1 is a block diagram of a moving object tracking device 1 according to the embodiment. The moving object tracking device 1 includes an imaging unit 2, a setting input unit 3, a storage unit 4, a control unit 5, and an output unit 6. The imaging unit 2, the setting input unit 3, the storage unit 4, and the output unit 6 are connected to the control unit 5.

  The imaging unit 2 is a monitoring camera, is installed so as to face the monitoring space, and images the monitoring space at a predetermined time interval. The captured monitoring images of the monitoring space are sequentially output to the control unit 5. In order to grasp the position and movement of a person who moves along a reference plane such as the floor surface or the ground surface, the imaging unit 2 is basically installed at a height that allows a bird's-eye photography, for example, in this embodiment, The moving object tracking device 1 is used for indoor monitoring, and the imaging unit 2 is installed on the ceiling. The time interval at which the monitoring image is captured is 1/5 second, for example. Hereinafter, the unit of time recorded at the time interval of imaging is referred to as time.

  The setting input unit 3 is an input means for the administrator to perform various settings for the control unit 5, and is a user interface device such as a touch panel display, for example.

  The storage unit 4 is a storage device such as a ROM (Read Only Memory) or a RAM (Random Access Memory). The storage unit 4 stores various programs and various data, and inputs / outputs such information to / from the control unit 5. The various data includes a three-dimensional model 40, a camera parameter 41, a hypothesis 42, an object position 43, an abundance map 44, and an object feature 45.

  The three-dimensional model 40 is data describing a state in which a moving object model that approximates a three-dimensional shape of a moving object is arranged in a virtual space that simulates a monitoring space (real space). In the present embodiment, the monitoring space and the virtual space are represented by a right-handed orthogonal coordinate system including the X, Y, and Z axes, and the vertically upward direction is set to the positive direction of the Z axis. A reference surface such as a floor surface is horizontal and is defined by an XY plane represented by Z = 0. The arrangement of the moving object model in the virtual space can be set at an arbitrary position.

  FIG. 2 is a perspective view schematically showing an example of the three-dimensional model 40. The reference plane 100, the camera position 101, and the moving object model 103 of the virtual space obtained by discretizing the monitoring space into N × M × K position coordinates. An arrangement example is shown.

  The moving object model is, for example, solid shape data having a height H obtained by stacking spheroids approximating the three-dimensional shapes of the human head torso and leg portions in the Z-axis direction and a maximum width W of the torso. is there. In this embodiment, in order to simplify the description, the height H and the width W are standard human sizes and are common to any moving object. The center of the head is set as the representative position of the moving object. Note that the moving object model may be further simplified and approximated by one spheroid. Data relating to the three-dimensional shape of the moving object model is stored in the storage unit 4 prior to the tracking process.

  The camera parameters 41 include external parameters and internal parameters of the imaging unit 2. These parameters obtained by actual measurement or the like are input in advance from the setting input unit 3 and stored in the storage unit 4. The three-dimensional model 40 can be projected onto the coordinate system of the monitoring image (imaging surface of the imaging unit 2; xy coordinate system) by a coordinate transformation formula in which the camera parameter 41 is applied to a camera model such as a known pinhole camera model.

  Hypothesis 42 (movement destination candidate) is information regarding the predicted value of the position of the movement destination of the moving object at each time. In order to determine the position (object position) of a moving object stochastically, a large number of hypotheses are set for each moving object (the number is represented by α, for example, 200 per moving object). Specifically, the hypothesis 42 is time-series data in which the identifier of the moving object is associated with the predicted value index (0 to α-1) and the position coordinates (XYZ coordinate system) at each time.

  The object position 43 is information regarding the position of the moving object at each time, and specifically is time-series data in which the identifier of the moving object is associated with the position coordinates (XYZ coordinate system). That is, the position of each moving object determined at each time is sequentially added to the object position 43, and information on the past position of each moving object is held.

  Note that it is possible to speed up the processing as a configuration in which the hypothesis 42 and the object position 43 are specified by the horizontal coordinate (XY coordinate system) of the monitoring space. In the present embodiment, the configuration will be described as an example for easy understanding.

  The presence map 44 is data in which a presence indicating the degree (probability) that each moving object may exist at each position in the monitoring space is associated with the position and the identification code of the moving object. The degree of presence is set with the hypothesis 42 or the object position 43 as the center position. Corresponding to the fact that the moving object occupies a part of the monitoring space and the occupied space of multiple moving objects exists exclusively, the presence of other moving objects when determining the object position of any moving object Is used as a penalty value to achieve an exclusive action between objects.

  The object feature 45 is information for individually characterizing the moving object to be tracked, and includes a feature amount (for example, a color histogram) of each moving object extracted from the monitoring image. The feature quantity of the moving object is extracted from the image area corresponding to the detected position and stored in the storage unit 4 when a moving object to be tracked is newly detected in the input image. It should be noted that data relating to the shape of the moving object model can be used as reference information when extracting the feature amount of the moving object.

  The object feature 45 also includes information representing a non-hidden area of the moving object. Here, the non-hiding area is an area where the moving object appears as an image on the monitoring image without being hidden by other moving objects. The non-hiding region can be obtained by projecting the three-dimensional model 40 in which the moving object model is virtually arranged on the imaging surface of the imaging unit 2 using the camera parameter 41. For each moving object, the non-hidden area when the moving object model is arranged at the position of the moving object is stored in the storage unit 4 as the object feature 45.

  The control unit 5 is configured using an arithmetic device such as a CPU (Central Processing Unit), a DSP (Digital Signal Processor), or an MCU (Micro Control Unit), and reads and executes a program from the storage unit 4 to change a pixel extraction unit. 50, functions as a hypothesis setting unit (position prediction unit) 51, likelihood calculation unit (evaluation value calculation unit) 52, object position determination unit 53, and abnormality determination unit 54. The likelihood calculating unit 52 includes a penalty setting unit 520 and a relaxation value setting unit 521.

  The change pixel extraction unit 50 extracts change pixels from the monitoring image newly input from the imaging unit 2, and outputs information about the extracted change pixels to the likelihood calculation unit 52 and the object position determination unit 53. Therefore, the change pixel extraction unit 50 performs a difference process between the newly input monitoring image and the background image, and extracts a pixel whose difference is equal to or greater than a preset difference threshold as a change pixel. The change pixel is an image feature of the moving object extracted by comparing with the monitoring image using the background image as reference information. Note that the changed pixels may be extracted by a correlation process between the newly input monitoring image and the background image instead of the difference process, or the background model is learned instead of the background image and the difference process with the background model is performed. The change pixel may be extracted by.

  The hypothesis setting unit 51 performs motion prediction of the moving object using past position information (object position or hypothesis) of the moving object, and obtains and distributes a plurality of hypotheses that are movement destination candidate positions at the current time of the moving object. Let

  First, the hypothesis setting unit 51 performs motion prediction from the object position of each moving object determined in the past or the hypothesis of each moving object set in the past, and the time when a monitoring image is newly input (attention time) A plurality of moving object candidates for the moving object is predicted, a hypothesis including each moving destination candidate and the identifier of the moving object is set, and the set hypothesis is output to the likelihood calculating unit 52. As described above, α hypotheses are set for each moving object. A method of determining a position (object position) of a moving object in a probabilistic manner by sequentially setting a number of hypotheses in this way is called a particle filter or the like. The hypothesis can be set in the xy coordinate system of the monitoring image, but in the present embodiment, it is set in the XYZ coordinate system of the monitoring space. Motion prediction is performed by applying a predetermined motion model or applying a predetermined prediction filter to past position data.

  The likelihood calculating unit 52 extracts an image feature of the moving object from a portion (non-hidden region) corresponding to each hypothesis of each moving object in the monitoring image, and the object position of the hypothesis corresponding to the degree of image feature extraction The likelihood (evaluation value) is calculated and output to the object position determination unit 53. At this time, the likelihood calculating unit 52 inputs the hypothesis of the target object to the penalty setting unit 520, calculates the presence of other objects other than the target object in the destination candidate indicated by the hypothesis, A correction that lowers the likelihood of the hypothesis is performed, and at the same time, the hypothesis of the object of interest is input to the relaxation value setting unit 521 and an adjustment (evaluation value compensation) is performed to reduce the correction for some hypotheses.

The hypothesis XY coordinates are (X, Y), the likelihood before correction for the hypothesis is L0 (X, Y), the presence of other objects in the hypothesis is P (X, Y), and the relaxation value for the hypothesis is Assuming that E, the likelihood L (X, Y) after correction for the hypothesis is calculated by subtracting the presence adjusted by the relaxation value from the likelihood before correction as shown in the following equation (1). However, E is a value of 0 or more and 1 or less.
L (X, Y) = L0 (X, Y) -P (X, Y) × (1-E) (1)

  That is, the presence of other objects acts as a penalty value that reduces the likelihood, and the relaxation value acts to weaken the penalty value.

  Hereinafter, calculation of the likelihood L0, the presence P, and the relaxation value E before correction will be described.

The likelihood L0 before correction is calculated as follows. The likelihood calculating unit 52 generates a three-dimensional model 40 in which a moving object model is virtually arranged on a destination candidate indicated by a hypothesis. Then, the camera parameter 41 is used to project the three-dimensional model 40 onto the imaging surface of the imaging unit 2 to obtain a non-hidden area of the moving object model, and the similarity, inclusion degree, and degree are determined by the following methods (1) to (3). The degree of edge extraction is calculated, and the likelihood corresponding to these weighted addition values is calculated. In addition, by including a model of an installation such as a fixture in the virtual space, a non-hidden area that further considers concealment by the set object can be obtained.
(1) The feature amount extracted from the non-hidden area at the past object position of each moving object is stored in the storage unit 4 as the object feature 45 of the moving object. As the hypothesis is closer to the position where the object of interest actually exists, the feature quantity of the background and other objects will not be mixed, so the similarity between the feature quantity extracted from the non-hidden region and the object feature 45 increases, while the distance increases. Similarity tends to be low. Therefore, the feature quantity in the non-hidden region is extracted from the monitoring image, and the similarity between the extracted feature quantity and the object feature 45 is calculated. As the feature amount here, for example, various image feature amounts such as an edge distribution, a color histogram, or both of them can be used.
(2) The non-obscured region is overlapped with the changed pixel extracted by the changed pixel extraction unit 50, and the ratio (inclusion degree) of the changed pixel included in the non-hidden region is obtained. Inclusion is a feature quantity of motion, and it tends to increase as the hypothesis is closer to the position where the moving object actually exists, and to decrease as the distance increases.
(3) An edge is extracted from a portion corresponding to the contour of the non-hidden region in the monitoring image. The extracted edge is a feature amount of the shape. The closer the hypothesis is to the position where the moving object actually exists, the more the edge is extracted (for example, the sum of the extracted edge strengths) because the contour of the non-hidden region matches the edge position. Tends to be low.

  The likelihood corresponding to any one of the above-described similarity, inclusion, and edge extraction may be calculated, or the likelihood may be calculated according to the weighted addition value of two of these. It may be calculated.

  Instead of using the likelihood as an evaluation value, the above-described similarity, inclusion, edge extraction, or a combination of two or more of these can be used as the evaluation value.

  Next, calculation of the presence level P will be described. FIG. 3 shows a schematic perspective view 200 of a monitoring space where four persons (moving objects # 1 to # 4) are present as a plurality of moving objects, and moving objects # 2 to # 4 in the perspective view 200. It is the schematic diagram 210 explaining the example of the presence map 44 set with respect to it. The abundance map 44 has an X axis and a Y axis that represent positions corresponding to the reference plane, and a P axis that represents the abundance perpendicular to the X and Y axes. The presence map 44 is generated by appropriately combining individual presence maps for each moving object. In the schematic diagram 210, an individual abundance map 212 of the moving object # 2, an individual abundance map 213 of the moving object # 3, and an individual abundance map 214 of the moving object # 4 are plotted. The presence map is synthesized by adding and normalizing these for each corresponding position. The presence map obtained by combining the individual presence maps 212, 213, and 214 of the moving objects # 2, # 3, and # 4 is the presence of other objects viewed from the moving object # 1 at each position in the monitoring space, that is, This represents the degree to which the moving object # 1 cannot exist at the position, and this presence map is used as a penalty value when determining the object position of the moving object # 1.

Each individual abundance map is an output value of a probability density function of one two-dimensional normal distribution having XY coordinates as a variable, or a function obtained by adding and combining probability density functions of a plurality of two-dimensional normal distributions having XY coordinates as variables. Expressed as an output value. In this embodiment, since the reference plane is discretized into N × M, the individual abundance map is actually XY in the range of X = [0, N−1] and Y = [0, M−1]. The output values of the above functions at the coordinates are arranged as individual abundances, and the collection codes of the individual abundances are associated with the moving object identification codes. When the individual presence map is represented by one two-dimensional normal distribution, the average value μ and the variance value σ ^{2 of the} normal distribution can be data associated with the identification code of the moving object, In addition, in the case of being expressed by addition synthesis of a plurality of two-dimensional normal distributions, a set of the average value μ and variance value σ ^{2 of} each normal distribution and the weight ω of each distribution is associated with the identification code of the moving object. It can also be data.

The composite presence map is obtained by superimposing a plurality of individual presence maps, and is data in which composite presences are arranged in the same XY coordinate range as the individual presence map. The composite abundance map may be data composed of a set of the average value μ and variance value σ ^{2 of} each added two-dimensional normal distribution and the weight ω of the distribution.

  The penalty setting unit 520 creates the presence map 44 with reference to the hypothesis 42 set by the hypothesis setting unit 51 and / or the object position 43 determined by the object position determination unit 53, and the created presence map 44 Is stored in the storage unit 4. When the penalty setting unit 520 receives the hypothesis 42 of the target object, the penalty setting unit 520 outputs the presence level of the other object at the position of the hypothesis 42 with respect to the target object from the presence map 44.

FIG. 4 is a schematic diagram for explaining a method of creating an individual presence map. FIG. 4A is a schematic diagram showing an example of the arrangement of the hypothesis 42 and the object position 43 on the reference plane for a moving object to be created as an individual abundance map. (Circle) represents the position of the hypothesis, and the triangle (Δ) represents the current object position. FIGS. 4B to 4D show three types of methods (A1) to (A3) for creating individual presence maps for moving objects whose hypotheses and the like are illustrated in FIG. 4A.
(A1) This method is a method of creating an individual presence map from the object position 43. In this method, a probability density function of a two-dimensional normal distribution is created with the object position 43 of the moving object to be created as an average value μ. FIG. 4B shows a two-dimensional normal distribution (average value μ, variance value σ ² ) of the individual abundance map set by this method, and a mountain shape having a peak in the P-axis direction at the object position. And a circle on the XY plane showing a range of 3σ. The variance value σ ² is determined from the hypothesized distribution width.
(A2) This method is a method of creating an individual presence map from the hypothesis 42. In this method, α hypotheses of the moving object to be created are read from the storage unit 4 and the individual existence degree of the moving object is calculated. The average value and variance of the hypothesis 42 of the moving object are calculated, and the hypothesis distribution is approximated by a two-dimensional normal distribution probability density function in which the calculated average value and variance are set. Then, the value of the function at each value in the monitoring space is set as the individual presence degree of the object at each position. FIG. 4C is a schematic diagram showing a hypothesis distribution approximated by a two-dimensional normal distribution. The average position μ of a plurality of hypotheses (◯) is indicated by a black small circle (●), and the two-dimensional normal distribution is shown. (Average value μ, dispersion value σ ² ) is represented by a mountain shape having a peak in the P-axis direction at the average position μ and a circle on the XY plane indicating a range of 3σ. This method is superior to (A3) because of the small amount of calculation and memory required.
(A3) This method is another method for creating an individual presence map from the hypothesis 42. This method can be used in a state where the likelihood has not yet been calculated for each hypothesis. FIG. 4D is a schematic diagram illustrating an example of creating an individual presence map by this method. For each hypothesis (◯), a two-dimensional normal distribution (average value μ, variance value σ ² ) illustrated as a mountain function is set. The position of the hypothesis is the average value μ of the normal distribution, and the variance value σ ² is set so that 3σ matches one half (W / 2) of the width of the moving object. Then, normalization is performed by dividing the output value at each XY coordinate of the function obtained by adding and synthesizing the two-dimensional normal distribution for each hypothesis with a weight of 1.0, and the respective XY after the normalization. A value in coordinates is defined as an individual abundance map of a moving object to be created. Compared with (A2), this method can express a complicated shape, and high accuracy can be obtained.

  The individual presence maps 212 to 214 of the moving objects # 2 to # 4 shown in FIG. 3 are created by the method (A2). The composite presence degree map of other objects with respect to the target object # 1 is calculated by appropriately adding and combining the individual presence degree maps 212 to 214 as described above. One method is to add values at positions corresponding to each other in the individual abundance maps of all objects # 1 to # 4, and divide the added value at each position by the maximum value in the XY coordinates of the addition result. A composite abundance map of all objects is created by normalization. Next, by subtracting the individual abundance of the position in the individual abundance map of the object of interest # 1 from the synthesis abundance at each position in the synthesis abundance map of all objects, the existence of synthesis for other objects # 2 to # 4 Create a degree map. When subtracting the individual abundance map of the object of interest from the composite abundance map, normalization is performed on the individual abundance map in the same manner as when the composite abundance map is created.

  It should be noted that the value range of the abundance is set to [0, 1] by the normalization described above, and matches the likelihood value range [0, 1]. When the range of abundance is not normalized to [0, 1], the subtraction result on the right side of equation (1) can be less than 0, but L (X, Y) in that case is defined as 0.

  The above-mentioned composite presence map for determining the position of the target object is created based on the hypothesis of the other object excluding only the target object among a plurality of moving objects. It corresponds to the abundance function given at each position in space. When the penalty setting unit 520 receives the identification code of the target object and its hypothesis, the penalty setting unit 520 determines the presence of other objects at the position of the hypothesis based on the combined presence map for object position determination corresponding to the target object. Is output.

  It should be noted that the composite presence 1.0 can be overwritten at a position where it is known in advance that a fixture or the like is installed and no moving object can exist.

  Next, the relaxation value E will be described. The penalty value based on the presence level P described above has an effect of making it difficult for the target person to be tracked in destination candidates other than the target person. When others with similar image features approach and increase the chances of high likelihood being calculated by the other person, they are mistracked as if they had passed each other, even though they were not actually passing each other Or two people approaching each other are confused and tracked as if there were two people at one person's location, while a third person appeared at a location where the other person actually exists Tracking errors such as being judged are likely to occur. The correction that reduces the likelihood by using the presence degree as a penalty acts to separate persons who are close to each other in such a case, thereby making it possible to prevent a tracking error.

  However, the correction process for separating the persons close to each other may cause the tracking object to be misunderstood in some cases because it does not match the actual movement of the person as described in the prior art. This problem has already been described in the case where the person of interest approaches a place where people gather in front of the person as shown in FIG. In other words, in this case, the person in the direction of travel makes a concession and clears the path, and the attention person may pass between these persons, but if the penalty value acts in this case, the object position of the attention person will be It becomes easy to make a determination in the direction opposite to the traveling direction, and it becomes easy for the person of interest to be mistaken for another person. On the other hand, there are cases where the movement of the person of interest changes, such as when the person of interest joins the ring and stops, or the person of interest passes away. In this case, if the penalty value does not act, a mistake is likely to occur.

  In addition, for example, even when a person of interest in a crowded person in a stopped state starts to move out of the crowd, a mistake may occur due to the action of the penalty value. In other words, the penalty value acts so as to be pushed inward for a person who is inside the crowd, and pushes outward for a person who is outside. Therefore, even if an inner person actually escapes, the person is likely to be tracked as an outer person. On the other hand, when the outside person actually gets out of the crowd, if the penalty value does not act, a mistake is likely to occur.

  Relaxed value E is set to simulate a human-like behavior that makes use of the exclusive action that can be obtained with a penalty as the presence level, and when people pass each other in a narrow passage or in a dense space, opens up a path. Is the value to be

  The mitigation value setting unit 521 sets mitigation values for some of the hypotheses at positions where the degree of presence of other objects is set among the hypotheses of the target object. The relaxation value setting unit 521 switches the hypothesis selection method for setting a relaxation value depending on whether the object of interest is in a moving state or a stopped state. That is, if it is in the moving state, the relaxation value is set in the direction of travel, and if it is in the stopped state, the relaxation value is set in a predetermined number of hypotheses without bias.

  Note that the installed object does not leave a path with respect to the object of interest unlike a moving object, so the hypothesis of the position where the presence of the installed object is set is not set as a relaxation value setting target.

  The movement state can be determined by calculating the movement vector from the past object position of the object of interest and comparing the magnitude (movement distance) of the movement vector with a preset stop determination threshold value Tv. Specifically, the relaxation value setting unit 521 calculates the distance between the object positions at the last two times of the target object, and if the distance is larger than the threshold value Tv set as the upper limit of the fine movement and detection error of the object. If the moving state and the distance are equal to or smaller than the stop determination threshold value Tv, the stop state is determined.

  Note that the moving distance can also be calculated from the average speed using the object positions at the last three hours or more, or the predicted speed calculated by applying a Kalman filter or the like to the object positions at the last three times or more.

  When the target object is in a moving state, the relaxation value setting unit 521 sets a relaxation value only for a hypothesis that is ahead in the traveling direction of the target object (direction of the movement vector), for example. At this time, for example, a larger relaxation value is set for a hypothesis closer to the traveling direction of the target object, and a smaller relaxation value is set for a hypothesis farther from the traveling direction of the target object. As a result, it is possible to cope with a situation change in which a path is opened with respect to an object of interest coming toward the other object.

  It is also preferable to set a larger relaxation value for a hypothesis closer to the object position one hour before the target object, and to set a smaller relaxation value for a hypothesis farther from the object position one time before the target object. When passing between other objects, the speed of the object of interest generally decreases, so by setting a larger relaxation value for a hypothesis with a smaller moving distance and setting a smaller relaxation value for a hypothesis with a larger moving distance, Furthermore, it is possible to set a relaxation value close to the actual situation.

Specifically, the relaxation value setting unit 521 calculates the movement vector v from the object position one hour before the target object to the hypothesis for which the relaxation value is set, and the movement distance d1 that is the magnitude of the movement vector v and the movement The traveling direction component d2 of the object of interest at the distance d1 is obtained, and the relaxation value E is increased or decreased in proportion to the ratio value d2 / d1 of the traveling direction component d2 to the moving distance d1 up to the hypothesis. Further, with D being the maximum value of d1 (the hypothetical maximum movement distance), the relaxation value E is calculated to be larger as the ratio value d1 / D of the movement distance d1 to D is smaller. As a result, the relaxation value set in the hypothesis of the target object becomes larger as the angle θ formed by the movement direction to the hypothesis with respect to the latest traveling direction of the target object becomes smaller, and becomes larger as the movement distance d1 to the hypothesis becomes smaller. Conversely, the larger the θ and d1, the smaller the setting. Such a relaxation value E can be defined by the following equation (2), for example.

  Equation (2) is a definition equation in the configuration in which the relaxation value E is set only for the hypothesis ahead in the traveling direction, and when d2 ≦ 0, that is, when θ is 90 degrees or more, the relaxation value is set to 0. To do.

5 and 6 are schematic diagrams for explaining the setting of the relaxation value E when the object of interest is in a moving state, and FIG. 5 is an example of the case where the object of interest moves toward another object that is densely packed. 6 is an example when there is no other object in the traveling direction of the target object. This figure schematically shows the object position, hypothetical position, and relaxation value on the XY plane. The circle with a large dotted line is the range in which the presence of other objects is set, and the triangle (△) is the past. The position of the object and the solid circle (○) represent the hypothesis. A square (□) indicates the size of the relaxation value set for the hypothesis (◯) indicated at the same position. The current time (attention time) is represented by t, and the object position one hour before is P _t-1 , and the object position two hours ago is P _t-2 . The most recent travel direction of the target object defined movement vector V _t-1 the direction from P _t-1 to P _t-2, tentatively a movement vector V _t of the current time to a start point P _t-1 It is shown along the traveling direction. The movement vector from P _t-1 to the hypothesis is represented by v. The magnitude of this vector v corresponds to the moving distance d1 to the hypothesis, the magnitude of v projected onto V _t is d2, and the angle between v and V _t is θ. As described above, the smaller the θ, the larger the relaxation value E, and the smaller the d1, the larger the relaxation value E. As a result, in the case shown in FIG. 5, the penalty is eased around the latest traveling direction in which the target object is likely to travel from P _t−1, and the target object passing between the other objects that are densely tracked is tracked. It becomes possible.

  On the other hand, when the traveling direction shown in FIG. 6 deviates from another object having a high density, the penalty is hardly alleviated. In view of this, it is possible to explicitly determine that the traveling direction has deviated from another object that is dense and exclude all hypotheses of the object of interest that are deviating from the relaxation value setting target. The determination of whether or not there is a deviation can be made based on whether or not the presence of another object is set in the traveling direction. If the presence is not set in the traveling direction, it is determined that there is a deviation. Further, it may be determined that a case where the number of the hypotheses of the target object set at the position where the presence degree is set is less than a predetermined ratio, for example, less than 10%, is deviated. If the traveling direction of the object of interest deviates from another object that is dense, the object of interest is likely to move avoiding the crowding.In this case, the effect of preventing misunderstanding can be achieved by maximizing exclusive control. Can be increased.

  The case where the target object is in the moving state has been described above. Next, setting of the relaxation value when the object of interest is in a stopped state will be described. When the target object is in a stopped state, the target object may start moving in any direction. Therefore, it is not possible to set a relaxation value centering on the direction in which the object of interest travels as in the case of the moving state described above. In addition, if a relaxation value is set in the entire circumference area that may progress, the exclusive action due to the presence degree is impaired, and the possibility of misunderstanding increases.

  In response to such a problem, the mitigation value setting unit 521 randomly selects a setting target at a constant rate from the hypothesis of the target object located in the region where the degree of presence of other objects is set. At this time, the relaxation value setting unit 521 also randomly sets the relaxation value E.

Specifically, the relaxation value setting unit 521 generates random numbers r1 and r2 that are not correlated with each other in the range [0, 1] by using random number generation means for each hypothesis of the position where the presence of other objects is set. Generate. Then, as expressed by the following equation (3), the relaxation value E is set to the random number r2 for the hypothesis in which the random number r1 equal to or less than Rth is generated, while the hypothesis in which the random number r1 greater than Rth is generated On the other hand, the relaxation value E is set to zero. For example, if Rth is set to 0.2, 20% of hypotheses of positions where the degree of presence of other objects is set are targeted for setting relaxation values. A value that takes into account the trade-off between the effect of the exclusive action by the penalty value and the effect of the concession reflecting action by the relaxation value is set in advance as Rth.

  FIG. 7 is a schematic diagram for explaining the setting of the relaxation value E when the object of interest is in a stopped state, and is expressed in the same format as in FIGS. 5 and 6. Incidentally, since the object of interest is in a stopped state, the triangle (Δ) representing the object position remains at one point, and the movement vector is not displayed because its size is zero. The relaxation value is randomly set to a hypothesis in the region where the presence level is set. That is, the relaxation value is different from that in the moving state described above in that the relaxation value is set without concentrating in a specific direction when viewed from the object position one time ago. Moreover, the magnitude of the relaxation value is set at random. That is, the size is different from that in the moving state in that the size is set regardless of the distance from the object position.

  In the above-described embodiment for the stop state, both the selection of the hypothesis for setting the relaxation value and the size of the relaxation value are determined using random numbers, but only one of them is determined using random numbers. May be. For example, without using the random number r2 that changes the magnitude of the relaxation value E described above, whether the relaxation value E is set to 0 or 1 depending on the magnitude relationship between the random numbers r1 and Rth, that is, whether the presence level is used as a penalty as it is You may choose to not use it. For example, when Rth is set to 0.2 and the relaxation value is set to 20% of the hypothesis, the relaxation value is set to 1 for the hypothesis in which a random number r1 equal to or less than Rth is generated, and a random number larger than Rth The relaxation value is set to 0 for the hypothesis for which r1 is generated.

  Further, the relaxation value E is calculated using random numbers having bias without performing hypothesis selection using the random number r1 and the threshold value Rth. The relaxation value E corresponding to the biased random number is calculated by E = g (r2) using, for example, a function g (·) in which a uniform random number r2 is generated and the expected value of the output value is adjusted to 0.2. it can.

  Likelihood calculation unit 52 calculates likelihood L0, presence P, and relaxation value E before correction for each hypothesis of the target object, as described above, and calculates corrected likelihood L from equation (1) using them. To do. That is, the likelihood L which performed the correction based on the correction which lowers likelihood according to the other object's presence degree P and the other object's concession with respect to the fall of the likelihood by the said correction is calculated.

The object position determination unit 53 determines the position (object position) of the moving object from each hypothesis of the moving object and the likelihood L calculated for each hypothesis, and the determination result is stored in the storage unit 4 for each moving object in time series. To accumulate. When all the likelihoods L are less than a predetermined lower limit (likelihood lower limit), it is determined that there is no object position, that is, disappeared. The following (1) to (3) are examples of the object position calculation method.
(1) For each moving object, a weighted average value of a hypothesis with likelihood as a weight is calculated, and this is set as the object position of the moving object.
(2) A hypothesis in which the maximum likelihood is calculated is obtained for each moving object, and this is set as the object position.
(3) For each moving object, an average value of a hypothesis in which a likelihood equal to or greater than a preset likelihood threshold is calculated, and this is set as an object position. Here, the likelihood threshold> the likelihood lower limit value.

  The abnormality determination unit 54 refers to the time-series object positions accumulated in the storage unit 4, determines that a suspicious movement that stays for a long time or a suspicious movement that deviates from the normal flow line is abnormal, and determines the abnormality. And output an abnormal signal to the output unit 6.

  The output unit 6 includes a sound output means for outputting a warning sound, a display means for displaying a monitoring image determined to be abnormal, or a communication means for transmitting to a security company center device via a communication line. When an abnormality signal is input from the unit 54, the fact that an abnormality has occurred is output to the outside.

[Operation of moving object tracking device]
Next, the tracking operation of the moving object tracking device 1 will be described. FIG. 8 is a schematic flowchart of the tracking process of the moving object tracking device 1.

  When the tracking process is started, the control unit 5 receives a monitoring image every time the imaging unit 2 images the monitoring space (S1). Hereinafter, the time when the latest monitoring image is input is called the current time, and the latest monitoring image is called the current image.

  The current image is compared with the background image by the change pixel extraction unit 50, and the change pixel extraction unit 50 extracts the change pixel (S2). Here, the isolated change pixel is excluded from the extraction result as being caused by noise. Immediately after the start of the operation without a background image, the current image is stored in the storage unit 4 as a background image, and there is no change pixel for convenience.

  The hypothesis setting unit 51 sets α hypotheses based on the motion prediction for each moving object being tracked (S3). Note that a hypothesis of a moving object that is determined to be a new appearance in step S16, which will be described later, is unpredictable, so α hypotheses are set in a wider range centering on the appearance position. Further, the hypothesis of the moving object determined to be lost in step S16 described later is deleted.

  If the change pixel is not extracted in step S2 and the hypothesis is not set in step S3 (there is no moving object being tracked) (if “YES” in S4), the control unit 5 performs step S1. Return to, and wait for input of the next monitoring image.

  On the other hand, if “NO” in the step S4, the processes of the steps S5 to S16 are performed. The penalty setting unit 520 creates an individual presence map of the moving object by the method (A2) described above based on the hypothesis of each moving object, and overwrites and saves it in the storage unit 4 (S5).

  The control unit 5 determines the front-rear relationship of the moving object (S6). Specifically, the distance between the hypothetical centroid (average value) and the camera position is calculated for each moving object, and a contextual list in which the identifiers of the objects are arranged in ascending order of the distance is created. Then, the control unit 5 sequentially sets each moving object being tracked as a target object from the previous one based on the context list (S7), and sets the presence map used for the tracking process of the target object for other objects. The individual abundance maps are synthesized and created (S8). Subsequently, the control unit 5 sequentially sets each hypothesis of the target object as the target hypothesis (S9). However, hypotheses that are outside the field of view of the monitoring image are excluded from the target of setting the hypothesis of interest, the object region in the hypothesis is not estimated, and the likelihood is set to zero.

  The likelihood calculating unit 52 arranges the moving object model at the position of the target hypothesis in the virtual space, and generates the three-dimensional model 40 on which the moving object model is arranged. Then, the camera parameter 41 is used to project the three-dimensional model 40 onto the imaging surface of the imaging unit 2 to obtain a non-hiding area of the moving object model, and this is estimated as a non-hiding area of the moving object in real space. Then, the feature amount of the target object is extracted in the estimated non-hidden region, and the uncorrected likelihood L0 corresponding to the target hypothesis is calculated based on the extraction degree (S10).

  The likelihood calculating unit 52 performs a correction process on the likelihood L0 and calculates a corrected likelihood L (S11). The correction process S11 will be described in detail later.

  When the hypothesis for which the likelihood L has not been calculated remains (“NO” in S12), the control unit 5 repeats steps S9 to S11. When the likelihood L is calculated for all α hypotheses (in the case of “YES” in S12), the object position determination unit 53 calculates the position (X, Y) of each hypothesis of the target object and each of the hypotheses. The object position of the object of interest is calculated using the obtained likelihood L (X, Y) (S13). The object position calculated for the current time is added in association with the object position of the object of interest stored in the storage unit 4 one hour before. In the case of a newly appearing moving object, a new identifier is assigned and registered. Further, when the likelihood in all hypotheses is less than the lower limit of likelihood, it is determined that there is no object position.

  The penalty setting unit 520 updates the individual presence degree map of the target object using the method (A1) using the object position calculated in step S13 (S14). Thereby, in the object position determination process (steps S7 to S14) of a certain object of interest, the object position determination result of another moving object processed before that is reflected. Since the presence level based on the object position can be expected to have higher accuracy than the presence level based on the hypothesis, the accuracy of the subsequent object position determination is improved by updating the above-described presence level map, and thus loop processing (steps S7 to S15). Is repeated once for all moving objects, it can be expected that a suitably converged state of the positional relationship between the moving objects can be obtained. Further, by determining the object position from the front object, the concealment state of the rear object by the front object is accurately evaluated, and the accuracy of the likelihood L0 calculated in step S10 is improved. This also improves the object position determination accuracy.

  When an unprocessed moving object remains ("NO" in S15), the control unit 5 repeats the process of determining the object position for the moving object (steps S7 to S14). On the other hand, when the determination of the object positions for all the moving objects is completed (in the case of “YES” in S15), the appearance and disappearance of the objects are determined (S16). Specifically, the object position determination unit 53 combines the object regions estimated for each object position, detects a change pixel outside the combination region among the change pixels extracted by the change pixel extraction unit 50, Among the detected change pixels, adjacent change pixels are labeled. If the label is of a size that can be regarded as a moving object, a new appearance is recorded in the storage unit 4 together with the label position (appearance position). If there is a moving object without an object position, the fact that the moving object has disappeared is recorded in the storage unit 4. When the above processing is completed, the processing returns to step S1 in order to perform processing on the monitoring image at the next time.

  FIG. 9 is a schematic process flow diagram of the likelihood correction process S11. The penalty setting unit 520 reads the presence P at the coordinates represented by the hypothesis of interest from the presence map created in step S8, and sets the presence P as a penalty value. The likelihood calculating unit 52 inputs the attention hypothesis and penalty value P to the relaxation value setting unit 521 (S20).

  If penalty value P is 0 (“NO” in S21), relaxation value setting unit 521 proceeds to step S12 of FIG. 8 without setting relaxation value E. In this case, the likelihood L is equal to the likelihood L0 of the image feature calculated in step S9.

On the other hand, if penalty value P set in step S20 is 0 or more (“YES” in S21), processes S22 to S26 for setting relaxation value E are performed. In this process, the relaxation value setting unit 521 first reads the object positions P _t-1 and P _t-2 of the target object calculated in the past from the storage unit 4, and calculates the movement vector V _t at the current time from these. (S22). Then, the moving distance | V _t |, which is the magnitude, is compared with the stop determination threshold value Tv (S23). If | V _t |> Tv, it is determined that the target object is in a moving state (“YES” in S23), and if | V _t | ≦ Tv, it is determined that the target object is in a stopped state (in S23). ("NO").

If it is determined in step S23 that the movement state, the relaxation value setting unit 521 obtains the movement vector v from the object position P _t-1 one time before read out in step S22 to the target hypothesis, and calculates the magnitude d1. Further, the magnitude d2 of the movement vector Vt direction component of the movement vector v is calculated (S24). Then, the relaxation value setting unit 521 calculates the relaxation value E according to the equation (3).

  If it is determined in step S23 that the stop state has occurred, the relaxation value setting unit 521 generates random numbers r1 and r2 by the random number generation means, and calculates the relaxation value E according to the equation (4) (S26).

  The likelihood calculating unit 52 uses the likelihood L0 of the image feature calculated in step S10, the penalty value P set in step S20, and the relaxation value E set in step S25 or S26. The likelihood L obtained by correcting the likelihood L0 of the image feature is obtained (S27). The likelihood calculating unit 52 returns to step S12, and repeats the above-described likelihood correction processing S11 for the attention hypothesis for all hypotheses.

In the above-described embodiment, the corrected likelihood is defined by equation (1). However, the likelihood correction equation is used to reduce the concession of other objects while reducing the presence of other objects. Other forms of mitigating / compensating for the reduction can be used. For example, the likelihood L (X, Y) after correction is subtracted from the upper limit value Pmax of the presence level (here, normalized to “1”) as shown in the following equation (4). The corrected value is multiplied by the likelihood before correction as a correction coefficient that reflects the exclusive effect according to the degree of existence in the likelihood, and concessioned by multiplying the correction coefficient by the compensation coefficient (1-E) according to the relaxation value. The effect may be reflected.
L (X, Y) = L0 (X, Y) × {1-P (X, Y) × (1-E)} (4)

  In the above-described embodiment, Pmax = 1 and the degree of presence is defined in the range of [0, Pmax]. On the other hand, the penalty value itself is allowed to be set beyond Pmax, and P (X, Y) or P (X, Y) × (1− When E) exceeds Pmax, L (X, Y) may be defined as 0.

  DESCRIPTION OF SYMBOLS 1 Moving object tracking apparatus, 2 Imaging part, 3 Setting input part, 4 Storage part, 5 Control part, 6 Output part, 40 3D model, 41 Camera parameter, 42 Hypothesis, 43 Object position, 44 Presence map, 45 Object Features, 50 Changed pixel extraction unit, 51 Hypothesis setting unit, 52 Likelihood calculation unit, 53 Object position determination unit, 54 Abnormality determination unit, 100 Reference plane, 101 Camera position, 103 Moving object model, 520 Penalty setting unit, 521 Relaxation Value setting part.

Claims (5)

A moving object tracking device for tracking a plurality of objects moving in the space using time-series images of a predetermined space,
A storage unit for storing position information of each object in the past from the attention time;
A position prediction unit that predicts and distributes destination candidates for each object at the time of interest from the position information;
An evaluation value calculation unit that obtains an evaluation value of the likelihood of the movement destination for the movement destination candidate from the image feature of the position corresponding to each of the movement destination candidates in the image at the attention time;
An object position determination unit that determines a movement destination position of each object based on the height of the evaluation value;
With
The evaluation value calculation unit uses one of the objects as a target object, and calculates the evaluation value for the destination candidate of the target object predicted within the distribution range of the destination candidates of other objects by a predetermined value. Performing a lowering correction, and reducing the predetermined value for some of the destination candidates in the distribution range to reflect the concessions of the other objects in the evaluation value,
A moving object tracking device. The moving object tracking device according to claim 1,
The evaluation value calculation unit, when it is determined from the position information that the target object is moving, sets the destination candidate predicted ahead of the target object as the partial candidate. Moving object tracking device. In the moving object tracking device according to claim 1 or 2,
The evaluation value calculation unit, when it is determined from the position information that the target object is moving, reduces the predetermined value as the destination candidate closer to the traveling direction of the target object. Object tracking device. In the moving object tracking device according to any one of claims 1 to 3,
The evaluation value calculation unit, when it is determined from the position information that the object of interest is moving, makes the predetermined value smaller for the movement destination candidate that is closer to the closest position of the object of interest. Object tracking device. In the moving object tracking device according to any one of claims 1 to 4,
The evaluation value calculation unit randomly selects the partial candidates from the movement destination candidates within the distribution range when the target object is determined to be stopped from the position information. Object tracking device.

Images