JP5848665B2 - Moving object motion vector detection apparatus, moving object motion vector detection method, and program - Google Patents

Description

  The present invention detects a motion vector from at least two frame images input by a camera that captures a scene including an object moving in real space, and determines the number of moving objects and the motion vector on the moving object from the motion vectors. The present invention relates to a moving object motion vector detection apparatus, a moving object motion vector detection method, and a program.

  A motion vector on a moving object that detects a motion vector from at least two frame images input by a camera that captures a scene including an object moving in real space, and detects a motion vector on the moving object from the detected motion vector There is a detection device. In this moving object motion vector detection device, it is possible to detect, for example, the reverse direction of a pedestrian or a car using the detected motion vector.

  For example, in Non-Patent Document 1, a moving object motion vector detection device calculates a motion vector from one or more frame images using template matching or a spatio-temporal differential method, and calculates the calculated motion vector as a similarity between them. Describes a method of detecting a moving object by clustering using the above.

  However, since this method detects a moving object with a correct motion vector, there is a problem in that a moving object cannot be detected correctly if a false detection vector is included in the motion vector. Therefore, for example, in Non-Patent Document 2, the detection is performed by detecting / removing the false detection vector from the motion vector and separating the remaining false detection vector and the motion vector on the moving object as a Gaussian mixture distribution. It describes a method for accurately detecting a motion vector on a moving object even when a false detection vector is included in the motion vector.

Yoshiaki Shirai and Masahiko Taniuchi, “Pattern Information Processing”, Ohmsha, 1998, p. 139-142 Jun SHIMAMURA, Masashi MORIMOTO and Hideki KOIKE, “Robust vSLAM for dynamic scenes”, MVA2011 IAPR Conference on Machine Vision Applications, June 13-15, 2011, Nara, JAPAN, pp.344-347

  In the conventional method such as Non-Patent Document 2 described above, the number of moving objects is known, and the distribution of the feature amount calculated from the motion vector is regarded as a Gaussian mixture distribution in which Gaussian distributions corresponding to the number of moving objects are mixed. To separate. Therefore, when the number of moving objects is unknown, the motion vector on the moving object cannot be detected, or the motion vector on the moving object is detected because the number of moving objects in the real world is different from the number of moving objects set as known. There was a problem that mistakes occurred.

  The present invention has been made in view of the problems of the prior art as described above, and is accurate even when the number of moving objects included as an image is unknown from a plurality of frame images including moving object images. A moving object motion vector detection apparatus, a moving object motion vector detection method, and a program for detecting a motion vector and the number of moving objects on a moving object are provided.

In order to solve the above-described problem, the present invention calculates a motion vector of a feature point from at least two frame images, and outputs the calculated motion vector and a frame image of the latest time among the frame images. A motion vector detecting unit; receiving the motion vector output from the motion vector detecting unit and the frame image at the latest time; calculating a feature quantity of the motion vector in the frame image at the latest time; A feature quantity calculation unit that creates a feature quantity list that associates the identification information of the motion vector and the feature quantity calculated for the motion vector, and the feature quantity list that is created by the feature quantity calculation unit while changing the number of clusters. The motion vectors to be displayed are clustered based on the feature amount and the clustering result is good A motion vector clustering unit that calculates an information amount reference value representing badness, and determines a motion vector on the moving object and the number of clusters based on a result of clustering the motion vector when the calculated information amount reference value is the best; The motion vector clustering unit regards the distribution of the feature quantity of the motion vector indicated by the feature quantity list as a Gaussian mixture distribution, and changes the clustering number while changing the clustering number to the clustering number. Obtained as a result of the separation when the calculated information reference value is the best, and the information reference value representing the fit of the Gaussian mixture distribution to the feature value distribution is calculated. For each of the Gaussian distributions, the Gaussian distribution is a cluster of motion vectors on a moving object. Is determined based on a Gaussian distribution parameter, the number of Gaussian distributions determined to be a cluster of motion vectors on a moving object is determined as the number of clusters, and the Gaussian distribution separated into the specified Gaussian distribution is determined. A moving object motion vector detection apparatus characterized by determining a motion vector as a moving object motion vector.

  Further, the present invention is the above-described moving object motion vector detection apparatus, wherein the feature amount is an angle of the motion vector, a luminance difference between both ends of the motion vector, or a luminance of one end point of the motion vector. It is characterized by being.

The present invention is also a moving object motion vector detection method executed by the moving object motion vector detection device, wherein the motion vector detection unit calculates a motion vector of a feature point from at least two frame images. A motion vector detection process for outputting the motion vector and a frame image at the latest time of the frame images, and a feature amount calculation unit that outputs the motion vector and the frame at the latest time output in the motion vector detection process. Upon receiving an image input, the feature amount of the motion vector in the frame image at the latest time is calculated, and a feature amount list in which the motion vector identification information and the feature amount calculated for the motion vector are associated is created. The feature quantity calculation process and the motion vector clustering unit change the clustering number while changing the feature number. Clustering the motion vectors indicated by the feature quantity list created in the quantity calculation process based on the feature quantities, calculating an information quantity reference value indicating whether the result of clustering is good or bad, and calculating the information quantity reference value possess a motion vector clustering process of determining the number of moving objects on the motion vectors and clusters based on the result of the clustering of the motion vector of the best case, a, in the motion vector clustering process, represented by the feature list The distribution of the feature quantity of the motion vector is regarded as a Gaussian mixture distribution, the Gaussian mixture distribution is separated into the Gaussian distribution of the clustering number while changing the clustering number, and the Gaussian mixture distribution with respect to the distribution of the feature quantity Information amount standard value indicating good / bad fit For each of the Gaussian distributions obtained as a result of separation when the calculated information amount reference value is the best, whether or not the Gaussian distribution is a cluster of motion vectors on a moving object based on a Gaussian distribution parameter The number of Gaussian distributions determined to be a cluster of motion vectors on the moving object is determined as the number of clusters, and the motion vector separated into the specified Gaussian distribution is determined as the motion vector on the moving object. This is a method for detecting a motion vector on a moving object.

  The present invention is a program that causes a computer to function as each unit of the above-described moving object motion vector detection apparatus.

  According to the present invention, it is possible to accurately detect a motion vector and the number of moving objects on a moving object from a plurality of frame images including an image of the moving object even when the number of moving objects included as an image is unknown.

1 is a block diagram illustrating a configuration of a moving object motion vector detection device 1 according to an embodiment of the present invention. FIG. It is a figure which shows the processing flow of the feature-value calculation part 12 by the embodiment. It is a figure which shows the processing flow of the motion vector clustering part 13 by the embodiment. It is a figure which shows an example of the frame image input into the moving-object top motion vector detection apparatus 1 by the embodiment. It is a figure which shows an example of the frame image input into the moving-object top motion vector detection apparatus 1 by the embodiment. It is a figure which shows an example of the motion vector calculation process result in the motion vector detection part 11 by the embodiment. It is a figure which shows an example of the feature-value list | wrist by the same embodiment. It is a figure which shows the histogram distribution of the feature-value of the motion vector which the feature-value calculation part 12 by the embodiment computed. It is a figure which shows the example of the normal distribution which the motion vector clustering part 13 by the same embodiment estimated from the histogram distribution of FIG. It is a figure which shows the example of the motion vector on a moving object which the motion vector clustering part 13 by the same embodiment detected.

  A moving object motion vector detection apparatus according to an embodiment of the present invention will be described below with reference to the drawings.

  FIG. 1 is a block diagram showing the configuration of a moving object motion vector detection apparatus according to an embodiment of the present invention, and shows only functional blocks related to the present embodiment. The moving object motion vector detection device 1 shown in FIG. 1 can be realized by a computer device, for example, and includes a motion vector detection unit 11, a feature amount calculation unit 12, a motion vector clustering unit 13, and a storage unit 14. Is done. Note that the “moving object” in the present embodiment includes an arbitrary moving subject such as a person or an object.

  The motion vector detection unit 11 receives input of at least two or more frame images obtained by capturing a scene including a moving object moving in real space from a camera or the like. The motion vector detection unit 11 extracts, from the input frame image, feature points that are corners or intersections on the image from the frame image at the latest time, and at least two or more frame images are used as motion vectors for the extracted feature points. From the above, it is calculated using template matching or spatiotemporal differentiation. The motion vector detection unit 11 outputs the calculated motion vector to the feature amount calculation unit 12 together with the frame image at the latest time.

  The feature amount calculation unit 12 receives the motion vector output from the motion vector detection unit 11 and the latest frame image and calculates the feature amount from the latest time frame image over all the motion vectors. The feature quantity calculation unit 12 creates a feature quantity list in which the motion vector ID uniquely assigned to each motion vector is associated with the feature quantity calculated for the motion vector, and the created feature quantity list is stored in the motion vector clustering unit 13. Output.

  The motion vector clustering unit 13 receives a feature amount list from the feature amount calculating unit 12 and performs a plurality of processes for clustering the motion vectors indicated by the input feature amount list based on the feature amount while changing the number of clustering. The information amount, which is a value for evaluating whether the model is fit or not, is calculated. Specifically, the motion vector clustering unit 13 regards the distribution of feature values as a Gaussian mixture distribution, and separates (clusters) the Gaussian mixture distribution into a Gaussian distribution of the clustering number while changing the clustering number. For example, an EM (Expectation Maximization) algorithm is applied to the clustering process, and Gaussian distribution parameters associated with each cluster are output by the clustering process. When the motion vector clustering unit 13 determines the optimum clustering number using the calculated information amount, the motion vector clustering unit 13 performs threshold processing on the Gaussian distribution parameters of each cluster output in the clustering process based on the determined clustering number, thereby moving the moving object. The number and the motion vector on the moving object are determined and output.

  The storage unit 14 appropriately stores data used for each process of the moving object motion vector detection device 1.

  Next, detailed processing executed by the moving object motion vector detection apparatus 1 shown in FIG. 1 will be described. First, the processing operation of the motion vector detection unit 11 will be described with reference to FIGS. 4 and 5.

  FIGS. 4 and 5 are examples of frame images input to the moving object motion vector detection apparatus 1. FIG. 5 shows a frame image F (t) obtained by imaging a real space at the current time with a camera. FIG. 4 shows a frame image F (t−1) obtained by imaging the same real space one hour before the current time. Show.

  When a frame image is input to the moving object motion vector detection device 1, the motion vector detection unit 11 is first activated. The motion vector detection unit 11 executes the following motion vector calculation process.

  First, the motion vector detection unit 11 extracts feature points that are corners or intersections on the image from the frame image F (t) that is the frame image at the latest time, and the motion vector for the extracted feature points is extracted from the frame image F (t ) And the frame image F (t−1). Here, the motion vector calculation processing is performed by, for example, template matching processing for the frame image F (t−1). That is, the motion vector detection unit 11 obtains image data in the vicinity of feature points that have been extracted from the frame image F (t), for example, image data in the vicinity of 8 feature points and image data of the frame image F (t−1). By comparing, the image coordinates of the point corresponding to the feature point to be processed on the frame image F (t) in the frame image F (t−1) are determined. The motion vector detection unit 11 determines the determined image coordinates and the image coordinates of the feature point to be processed on the frame image F (t) as a motion vector. This process is performed over all feature points extracted from the frame image F (t). Finally, the motion vector detection unit 11 outputs the determined motion vector together with the frame image F (t) to the motion vector detection unit 11.

  FIG. 6 is a diagram illustrating an example of a motion vector determined as a result by the motion vector detection unit 11. In the present embodiment, the subsequent processing will be described on the assumption that the motion vector detection unit 11 has calculated the motion vector shown in FIG. Note that the number of motion vectors determined by the motion vector detection unit 11 is P.

  After the motion vector calculation process by the motion vector detection unit 11 described above, the feature amount calculation unit 12 is subsequently activated.

  FIG. 2 is a diagram illustrating a processing flow of the feature amount calculation unit 12. In the present embodiment, the following description will be made assuming that the angle of the motion vector is used as the feature amount. However, another feature amount such as a luminance difference between both ends of the motion vector or a luminance itself at one end point may be used. This process is the same as the method described in Non-Patent Document 2.

  First, a motion vector is input from the motion vector detection unit 11 to the feature amount calculation unit 12 (step S11). The feature amount calculation unit 12 determines whether or not the processing after step S13 has been processed for all input motion vectors (step S12).

When it is determined that there is an unprocessed motion vector (step S12: NO), the feature amount calculation unit 12 selects one unprocessed motion vector and calculates the angle of the motion vector on the image (step S13). . For example, when the start point coordinate of the motion vector is (Xs, Ys) and the end point coordinate is (Xe, Ye), the angle θ (radian) is given by tan ⁻¹ ((Ys−Ye) / (Xs−Xe)). Calculated. The angle θ is not limited to the isolation method, and may be expressed by other units such as a frequency method, or may be expressed by two subparameters such as θ ₁ = sin θ and θ ₂ = cos θ.

  The feature amount calculation unit 12 associates the motion vector ID uniquely assigned to the selected motion vector with an angle parameter indicating the angle θ calculated for the selected motion vector, and the feature amount list recorded in the storage unit 14 Is added and written (step S14). The feature amount calculation unit 12 repeats the processing from step S12.

  When it is determined that the processing has been performed for all the motion vectors (step S12: YES), the feature amount calculation unit 12 outputs the feature amount list stored in the storage unit 14 to the motion vector clustering unit 13 (step S15). The process ends.

  FIG. 7 is a diagram illustrating an example of the feature amount list. In the present embodiment, the following processing will be described on the assumption that the feature quantity list in which the angle parameter of the motion vector is recorded in association with the motion vector ID as shown in FIG. In the motion vector recorded with the motion vector ID in this feature quantity list, in addition to the motion vector on the moving object, occlusion in which the object is concealed by another object or an error caused by another cause is detected. Motion vectors are also included.

  Following the processing of the feature quantity calculation unit 12 shown in FIG. 2, the motion vector clustering unit 13 is activated.

FIG. 3 is a diagram illustrating a processing flow of the motion vector clustering unit 13.
First, the feature amount list output from the feature amount calculating unit 12 is input to the motion vector clustering unit 13 (step S21). The motion vector clustering unit 13 determines whether or not N (N is an integer of 2 or more) indicating the number of repetition processes is equal to or less than Nmax indicating the maximum number of repetition processes (step S22). When the motion vector clustering unit 13 determines that N is equal to or less than Nmax (step S22: YES), the motion vector clustering unit 13 clusters the motion vector IDs indicated by the feature amount list into N based on the angle parameter (step S23). ).

FIG. 8 shows a histogram distribution of the angles of the motion vectors shown in the feature quantity list.
In the present embodiment, the histogram distribution of the angle of the motion vector is modeled and expressed by a Gaussian mixture distribution that is a combination of a plurality of Gaussian distributions (same as the “normal distribution”), and the motion vector clustering unit 13 uses the parameters of each peak. Is determined by performing processing such as an EM algorithm. Thereby, the motion vector clustering unit 13 clusters the motion vectors shown in the feature quantity list into a motion vector on a moving object, an erroneous motion vector due to occlusion, and an erroneous motion vector due to another cause. To do.

  Since details of the EM algorithm are generally known, the description thereof is omitted in this specification. Note that the EM algorithm is an algorithm that estimates the maximum likelihood of the parameters of a probability model (normal distribution curve in this embodiment) numerically by an iterative method, and is used particularly when the probability model depends on hidden parameters that cannot be observed. It is done. The EM algorithm is effective because it easily converges to a good solution compared to other estimation methods, is often simple to implement, and the processing speed is basically high.

FIG. 9 shows an example of a result of estimating a normal distribution curve by applying the EM algorithm to the angular distribution of FIG. In the figure, an example in the case of N = 4 is shown, and normal distribution curves D _{1 to} D ₄ corresponding to the _four clusters C _{1 to} C ₄ are estimated. In this embodiment, the EM algorithm is used to estimate the normal distribution curve, but other algorithms can naturally be used. The motion vector ID is clustered by executing the EM algorithm with the feature quantity list as input, and the cluster number of the cluster to which the motion vector ID belongs is assigned in association with each motion vector ID. In addition, Gaussian distribution parameters associated with each cluster are output for the number of clusters. The Gaussian distribution parameter includes at least an average value indicating the center position of the Gaussian distribution and a variance value indicating the spread of the Gaussian distribution, but may also include a weight value indicating the weight of the Gaussian distribution with respect to the Gaussian mixture distribution.

  Next, the motion vector clustering unit 13 calculates the information amount reference value IC using the maximum likelihood that is one of the outputs of the EM algorithm (step S24). Here, the information criterion is an amount representing the goodness of fitting of the model in statistics, and in the case of the present embodiment, indicates whether the fitting of the Gaussian mixture distribution to the histogram distribution of the angle of the motion vector indicated by the feature amount list is good or bad. . A value calculated based on the information amount reference is referred to as an information amount reference value IC. The motion vector clustering unit 13 calculates the information amount reference value IC by, for example, the following equation (1-1) or the following equation (1-2).

IC = -2ln (L) + 2k ... Formula (1-1)
IC = -2ln (L) + k * ln (n) ... Formula (1-2)

In Expressions (1-1) and (1-2), L is the maximum likelihood obtained in the EM algorithm, and k is the number of free parameters. In the case of a Gaussian mixture distribution, the number k of free parameters is k = N × {M + M × (M + 1) / 2 + 1}. Here, M is the number of dimensions of data. Similar to the description in Non-Patent Document 2, when the angle of the motion vector is expressed by _two subparameters θ ₁ and θ ₂ as the feature quantity, M = 2 and the number of free parameters k = N {2 + 2 × 3/2 + 1} = 6N. In Expression (1-2), n is the number of data, and in the case of this embodiment, is the number P of motion vectors.

  Equation (1-1) is the Akaike information reference amount (AIC), and Equation (1-2) is the Bayes information reference amount (BIC). In addition, although there exist bootstrap reference | standard quantity (EIC), MDL (minimum description length) reference | standard quantity etc. as another information amount reference | standard, any may be used in this embodiment.

  Next, the motion vector clustering unit 13 determines whether or not the information amount reference value IC is better than the best information amount reference value ICopt (step S25). When the information amount reference value IC of the expression (1-1) or the expression (1-2) is used, the motion vector clustering unit 13 indicates that the information amount reference value IC is the best information amount in order to indicate that a smaller value is better. It is determined whether it is smaller than the reference value ICopt. It is assumed that a sufficiently large value such as 99999999 is set as the initial value of the best information amount reference value ICopt.

When it is determined that the information amount reference value IC is better than the best information amount reference value ICopt (step S25: YES), the motion vector clustering unit 13 replaces the current best information amount reference value ICopt with the information amount reference value IC and stores it. It records in the part 14 (step S26). In addition, the motion vector clustering unit 13 also stores Gaussian distribution parameter values output in association with the clusters C ₁ to C _N and the cluster numbers associated with each motion vector ID recorded in the feature quantity list. It records in the part 14 (step S27). If the Gaussian distribution parameter value or the cluster number associated with the motion vector ID has already been recorded in the storage unit 14, the motion vector clustering unit 13 overwrites and records it.

  After the process of step S27, or when it is determined in step S25 that the information amount reference value IC is not better than the best information amount reference value ICopt (step S25: NO), the motion vector clustering unit 13 increases N by one. Increment (step S28). After the increment, the motion vector clustering unit 13 returns to the process of step S22 for determining whether N is equal to or less than Nmax indicating the maximum number of repetition processes.

  When the motion vector clustering unit 13 determines that N has exceeded Nmax indicating the maximum number of iterations (step S22: NO), the motion vector clustering unit 13 reads the Gaussian distribution parameter value of each cluster recorded in the storage unit 14 at that time. . The motion vector clustering unit 13 specifies a cluster number of a cluster that exceeds a predetermined threshold and determines the number of clusters that exceed the predetermined threshold by threshold processing on the read Gaussian distribution parameter. This threshold processing can be realized, for example, by comparing the variance value of the Gaussian distribution parameter with a predetermined value (threshold). Alternatively, the threshold processing can also be realized by comparing the weight value of the mixed distribution with a predetermined value (threshold). A cluster whose Gaussian distribution parameter exceeds a predetermined threshold is a cluster of motion vectors on a moving object, and a cluster which does not exceed a predetermined threshold is caused by, for example, a cluster of erroneous motion vectors due to occlusion or other causes. For example, a cluster of erroneous motion vectors.

  Finally, the motion vector clustering unit 13 outputs the number of clusters determined that the Gaussian distribution parameter has exceeded a predetermined threshold as the number of moving objects. Furthermore, the motion vector clustering unit 13 reads out the motion vector ID associated with the cluster number of the cluster for which the Gaussian distribution parameter has been determined to exceed a predetermined threshold from the feature amount list, and uses the read motion vector ID as the motion on the moving object. It outputs as a vector and finishes a process (step S29). As the output, when the feature amount calculation unit 12 generates the feature amount list, the start point and end point of the motion vector are recorded in the storage unit 14 together with the angle of the motion vector in association with the motion vector ID. The vector clustering unit 13 may output the start point and the end point of the motion vector associated with the motion vector ID specified as the motion vector on the moving object.

  FIG. 10 is a diagram illustrating a motion vector on a moving object detected by the motion vector clustering unit 13 as a result of the above processing. In the figure, two moving object motion vectors, moving object A and moving object B, are shown. Three moving object motion vectors of the moving object A are obtained from one cluster, and three moving object motion vectors of the moving object B are obtained from one cluster different from the moving object A.

  According to the above-described embodiment, the moving object motion vector detection device 1 regards the distribution of the feature amount calculated from the feature amount of the motion vector as a Gaussian mixture distribution, and separates the Gaussian mixture distribution into a Gaussian distribution (clustering). This process is repeated a plurality of times while changing the number of separations (clustering number), and the optimum number of separations is determined using the amount of information for evaluating whether the model fit is good or bad. The moving object motion vector detection apparatus 1 accurately detects the motion vector on the moving object in order to determine the number of moving objects and the motion vector on the moving object using the Gaussian mixture distribution separation result at the optimal separation number. can do.

  The above-described moving object motion vector detection apparatus 1 has a computer system therein. The operation processes of the motion vector detection unit 11, the feature amount calculation unit 12, and the motion vector clustering unit 13 of the moving object motion vector detection device 1 of the present embodiment are recorded on a computer-readable recording medium in the form of a program. The above processing is performed by the computer system reading and executing this program. The “computer system” here includes a CPU, various memories, an OS, and hardware such as peripheral devices. The “computer system” includes a WWW system having a homepage providing environment (or display environment). The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Further, the “computer-readable recording medium” refers to a volatile memory (RAM) in a computer system that becomes a server or a client when a program is transmitted via a network such as the Internet or a communication line such as a telephone line. In addition, those holding programs for a certain period of time are also included.

  The program may be transmitted from a computer system storing the program in a storage device or the like to another computer system via a transmission medium or by a transmission wave in the transmission medium. Here, the “transmission medium” for transmitting the program refers to a medium having a function of transmitting information, such as a network (communication network) such as the Internet or a communication line (communication line) such as a telephone line. The program may be for realizing a part of the functions described above. Furthermore, what can implement | achieve the function mentioned above in combination with the program already recorded on the computer system, what is called a difference file (difference program) may be sufficient.

Although the present invention has been specifically described above based on the embodiments, the description of the above embodiments is for explaining the present invention, and limits the invention described in the claims. Should not be construed as reducing the range.
Moreover, each means structure of this invention is not restricted to the said embodiment, Of course, a various deformation | transformation is possible within the technical scope as described in a claim.

DESCRIPTION OF SYMBOLS 1 Moving object motion vector detection apparatus 11 Motion vector detection part 12 Feature-value calculation part 13 Motion vector clustering part 14 Memory | storage part

Claims (4)

A motion vector detection unit that calculates a motion vector of a feature point from at least two frame images, and outputs the calculated motion vector and a frame image at the latest time of the frame images;
Receiving the motion vector output from the motion vector detection unit and the frame image at the latest time, calculating a feature quantity of the motion vector in the frame image at the latest time, and identifying the motion vector identification information and the motion A feature quantity calculation unit that creates a feature quantity list that associates the feature quantities calculated for the vector;
While changing the number of clustering, the motion vector indicated by the feature amount list created by the feature amount calculation unit is clustered based on the feature amount, and an information amount reference value representing the quality of the clustering result is calculated, A motion vector clustering unit that determines the motion vector on the moving object and the number of clusters based on the result of clustering the motion vectors when the calculated information amount reference value is the best;
Equipped with a,
The motion vector clustering unit regards the distribution of the feature quantity of the motion vector shown in the feature quantity list as a Gaussian mixture distribution, and changes the Gaussian mixture distribution to the Gaussian distribution of the clustering number while changing the clustering number. The Gaussian distribution obtained as a result of the separation when the calculated information amount reference value is the best, and the information amount reference value indicating whether the Gaussian mixture distribution is fit to the feature amount distribution is calculated. For each, it is determined whether the Gaussian distribution is a cluster of motion vectors on a moving object, based on Gaussian distribution parameters, and the number of Gaussian distributions determined to be a cluster of motion vectors on a moving object is determined. The number of clusters is determined, and the motion vector separated into the specified Gaussian distribution is transferred. Is determined as an object on the motion vector,
An apparatus for detecting a motion vector on a moving object. 2. The moving object motion vector detection device according to claim 1, wherein the feature amount is an angle of the motion vector, a luminance difference between both ends of the motion vector, or a luminance of one end point of the motion vector. . A moving object motion vector detection method executed by a moving object motion vector detection apparatus,
A motion vector detection process in which a motion vector detection unit calculates a motion vector of a feature point from at least two frame images, and outputs the calculated motion vector and a frame image at the latest time of the frame images;
A feature amount calculation unit receives the motion vector output in the motion vector detection process and the frame image at the latest time, calculates a feature amount of the motion vector in the frame image at the latest time, and the motion vector A feature quantity calculation process for creating a feature quantity list that associates the identification information of the motion vector and the feature quantity calculated for the motion vector;
A motion vector clustering unit clusters the motion vectors indicated by the feature quantity list created in the feature quantity calculation process based on the feature quantities while changing the number of clustering, and indicates the amount of information indicating whether the clustering result is good or bad A motion vector clustering step of calculating a reference value and determining a motion vector on the moving object and the number of clusters based on a result of clustering of the motion vector when the calculated information amount reference value is the best;
I have a,
In the motion vector clustering process, the distribution of the feature quantity of the motion vector indicated in the feature quantity list is regarded as a Gaussian mixture distribution, and the Gaussian mixture distribution is converted into a Gaussian of the clustering number while changing the clustering number. The Gaussian mixture obtained as a result of the separation when the information amount reference value that represents whether the Gaussian mixture distribution is fit to the distribution of the feature amount is calculated. For each distribution, whether the Gaussian distribution is a cluster of motion vectors on a moving object is determined based on Gaussian distribution parameters, and the number of Gaussian distributions determined to be a cluster of motion vectors on a moving object Is determined as the number of clusters, and the motion vector separated into the specified Gaussian distribution is determined. The determined as a moving object on the motion vector Torr,
A method for detecting a motion vector on a moving object. Computer
A program that functions as each unit of the moving object motion vector detection device according to claim 1 .

Images