JP2014074977A - Image processing apparatus, image processing method, and image processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an image processing apparatus configured to detect and track an object, such as a person or an object, at a high speed, to improve detection accuracy.SOLUTION: An image processing apparatus includes: an object detection part which detects an object window including an object from an input frame image; a clustering integration part which integrates nearby object windows detected by the object detection part by clustering, to set an object area; and an object tracking part which registers an object candidate area in association with an object area of a past frame image, for an object candidate area, which is the object area, having a predetermined range of window width difference and located at the minimum distance from an object prediction area which is an area predicted based on the object area of the past frame image.

Description

  The present invention relates to an image processing apparatus and its program for detecting and tracking an object such as a person or an object, and detecting the object by performing high-speed detection and reliably tracking the object with a small amount of features. And tracking can be done in real time.

  A system that captures a visitor's video with a fixed-point camera, detects and tracks a person, has been studied to measure the number of visitors and investigate the age and gender.

  As a method for detecting a photographed object, particularly a person, an identification method called a HOG (Histograms of oriented Gradients) using a luminance gradient histogram as a feature amount is widely used.

N.Dalal and B. Triggs: “Histgrams of Oriented Gradients for Human Detection”, IEEE Computer Society Conference on ComputerVision and Pattern Recognition, vol.1, pp.886-893 (2005) X. Wang, T.X.Han and S. Yan: “An HOG-LBP Human Detector with Partial Occlusion Handling”, IEEE 12th International Conference on ComputerVision, pp.32-39 (2009)

  The person detection using the above HOG feature amount has a problem that it can perform person detection with high accuracy, but takes time for calculation. Also, in tracking people, it is difficult to apply to tracking people because it is necessary to process each complaint image in real time at high speed and discard the images in consideration of privacy. It was. In addition, it is necessary to track a person at high speed.

  The present invention relates to an image processing apparatus capable of detecting and tracking an object such as a person or an object at high speed and improving detection accuracy.

  The image processing apparatus of the present invention integrates an object detection unit that detects an object window in which an object exists from an input frame image, and a plurality of neighboring object windows detected by the object detection unit by clustering, thereby obtaining an object region. The clustering integration unit to be set, the object region as the object candidate region, the region predicted based on the object region of the past frame image as the object prediction region, and the object candidate region whose window width difference is within the predetermined range. And an object tracking unit that registers an object candidate area that is at the shortest distance from the object prediction area in association with the object area of the past frame image.

  In the image processing apparatus, the object tracking unit uses, as a related object candidate area, an object candidate area that has a window width difference within a predetermined range and is within a predetermined distance from the object prediction area among the plurality of object candidate areas. The related object candidate area that is the shortest distance from the object prediction area among the related object candidate areas may be extracted and registered in association with the object area of the past frame image.

  In the above image processing device, the object tracking unit, when the object candidate area is associated with two or more object prediction areas, displays the object area of the past frame image corresponding to the object prediction area at the shortest distance from the object candidate area. May be registered in association with each other.

  Further, in the image processing apparatus, the object tracking unit may be associated and registered by assigning the same identifier as the identifier assigned to the object region of the past frame image to the object candidate region.

  In the image processing apparatus, the object tracking unit may change the coordinates of the first object area of the frame image two frames before and the second object area of the frame image one frame before associated with the first object area. The object prediction region in the input frame image may be determined based on the amount.

  In the above image processing apparatus, when there is no object area associated with the frame image two frames before, the object tracking unit changes the input frame by changing a predetermined amount from the coordinates of the object area in the frame image one frame before. The object prediction area in the image may be determined.

  In the image processing apparatus, the object detection unit calculates a feature amount of the window by providing a window in the input frame image and performing an overlap scan in the window with a block having a predetermined area, and based on the calculated feature amount. An object window may be detected.

  In the above image processing apparatus, when the overlap detection is performed, the object detection unit calculates the feature amount of the window by reusing the feature amount once calculated as the overlap region feature amount for each scan layer. Also good.

  The present invention includes a step of detecting an object window in which an object exists from an input frame image, a step of setting an object region by integrating a plurality of detected object windows in a vicinity, and an object region as an object candidate region. An object predicted area based on an object area of a past frame image is set as an object predicted area, and an object candidate having a window width difference within a predetermined range and having the shortest distance from the object predicted area An image processing method including a step of registering an area in association with an object area of a past frame image is provided.

  Further, the present invention includes a step of detecting an object window in which an object exists from an input frame image in a computer, a step of setting an object region by integrating a plurality of detected object windows in a vicinity by clustering, and the object The area is an object candidate area, the area predicted based on the object area of the past frame image is the object prediction area, the object candidate area has a window width difference within a predetermined range, and is the shortest from the object prediction area There is provided a program for executing a step of registering an object candidate area at a distance in association with an object area of a past frame image.

  According to the present invention, in an input frame image, an object prediction region is set based on a past frame image, and is an object candidate region detected by the object detection unit as an object, satisfying a predetermined condition and having the shortest distance Can be registered in association with the object area of the past frame image, so that the same person can be tracked between the frame images.

  Further, since tracking is performed without comparing histograms or color information in the tracking process, the tracking process can be performed in real time. Furthermore, since the object candidate area is detected based on a feature quantity such as luminance, tracking can be performed with high accuracy even with a small number of parameters.

  Further, in the object detection unit, a window is provided in the input frame image, overlap scan is performed with a block having a predetermined area in the window, and the feature amount calculated once is reused for each scan layer when performing overlap scan. Thus, object detection can be performed at high speed.

It is an example of the block diagram of the image processing apparatus in one embodiment concerning this invention. It is an example of the hardware block diagram of the image processing apparatus of this invention. 5 is an example of a flowchart when the object detection unit 12 detects an object from an input frame image in the image processing apparatus of the present invention. 3 is an example of a scan table stored in a scan table storage unit 11. It is a conceptual diagram which shows the relationship between a cell and a block. It is a conceptual diagram which shows the relationship between a window and a block. It is an example of a process in case the object tracking part 14 matches the same object area | region between frame images in the image processing apparatus of this invention. It is an image figure of correlation of the object area | region between frame images. It is a figure which shows the relationship between the object prediction area | region and object candidate area | region in step 703 of FIG. It is a table which shows the experimental result at the time of performing object detection in the image processing apparatus of this invention.

  An embodiment in which the present invention is applied to person tracking by detecting the upper body area of a person photographed in an image will be described below. In this embodiment, a camera is installed at a predetermined position, and a person passing through a shooting location is shot. A person is detected in the moving image by detecting a person from the captured image and performing tracking between the frame images. However, although it is the upper body area of the person here, the object to be detected is not limited to the upper body of the person, but may be a part of another person, eyes or nose, or an object such as a car or a car license plate. Or part of an object.

  FIG. 1 is an example of a block diagram of an image processing apparatus according to an embodiment of the present invention. The image processing apparatus 100 includes an image storage unit 10, a scan table storage unit 11, an object detection unit 12, a clustering convenience unit 13, and an object tracking unit 14. The image storage unit 10 stores an image input from a camera connected via a network or a wired cable. The camera captures a moving image, and image data that is time-series data of the moving image is sequentially input to the image processing apparatus 100 via, for example, a communication interface. Each input image data is a frame image. The input frame images are stored in the image storage unit 10 sequentially or temporarily or for a long time.

  The scan table storage unit 11 is a table that stores a window size, a scan stride value, the number of windows, and the like when scanning a frame image. When performing object detection processing from a frame image, a window is set in the image, and the feature amount in the window is calculated by scanning while overlapping the windows. Various window sizes can be set, and since the scan stride value and the number of windows necessary to scan the entire image data change depending on the window size, it is stored as a table in association with the window size. The detection unit 12 reads the scan table from the scan table storage unit 11 and performs scanning with a predetermined window size.

  The object detection unit 12 sequentially reads the frame images stored in the image storage unit 10 and performs an object window detection process in which an object exists from the input frame image. For example, the object detection unit 12 reads the scan table from the scan table storage unit 11 and searches the entire input frame image with a predetermined scan layer window size. As an example, by performing overlap scanning with a predetermined scan stride value in a window having a predetermined window size, a luminance gradient direction histogram in the window is generated, and the HOG feature amount is calculated.

  For example, when detecting the upper body of a person as an object, the object detection unit 12 uses a support vector machine (SVM) or the like to learn in advance the feature amount (HOG feature amount) of the luminance gradient direction histogram, Obtain an SVM classifier. Then, an SVM classifier evaluates whether or not the HOG feature amount extracted from the input frame image is the upper body of the person, thereby detecting an object window that is an area of the upper body of the person. That is, the SVM classifier detects a window having a feature value that is the upper body of a person as an object window.

  In addition, the object detection unit 12 sequentially scans the input frame image with windows having various window sizes, and detects an object region that is a region where the upper body is photographed.

  The clustering integration unit 13 integrates object windows, which are windows detected by the object detection unit 12, by clustering. That is, when the input frame image is scanned in various windows, a plurality of neighboring object windows exist for one upper body. In other words, a plurality of windows having different positions and windows having different sizes are overlapped to be detected as an object window. Therefore, the plurality of object windows are integrated and set as one object region using, for example, Meanshift clustering.

ここにいうbandwidth_positionとbandwidth_scaleは、位置とスケールにおけるそれぞれのクラスタリング探索領域として設定するパラメータである。 For example, when using Meanshift clustering, the clustering integration unit 13 performs three-dimensional clustering integration of the identified object window position (x _i , y _i ) and the scale scale _i using the following Gaussian kernel function: The position and scale shift values are calculated from the accumulated values, and the positions and scales are corrected and updated from the respective shift values to be converged to form one object region.

Here, bandwidth_position and bandwidth_scale are parameters set as respective clustering search regions in the position and scale.

  The object tracking unit 14 sets the object area in the input frame image as an object candidate area, and sets the area in which the object is predicted to exist in the input frame image based on the object area in the past frame image as the object prediction area, and satisfies a predetermined condition. The object candidate area is registered in association with the object area of the past frame image corresponding to the object prediction area, thereby tracking the object and outputting the result. The past frame image is, for example, an image one frame before the input frame image.

  For example, the object tracking unit 14 determines an object candidate area that has a window width difference within a predetermined range and that has the shortest distance from the object prediction area as compared to other object candidate areas, as a past frame. Register in association with the object area of the image. For example, when registering in association with each other, the object tracking unit 14 assigns the same identifier as the identifier (person ID) assigned to the object area of the past frame image to the object candidate area. Details of the tracking method will be described later.

  FIG. 2 is a block diagram showing an example of a hardware configuration of the image processing apparatus 100 according to the embodiment of the present invention. In FIG. 2, the computer constituting the image processing apparatus 100 can be realized by a general-purpose hardware configuration that has existed before. That is, the computer forming the image processing apparatus 100 includes a CPU 101, a ROM 102, a RAM 103, an external storage device 104, a communication interface 105, a mouse 107 and a keyboard 108 connected to the input / output interface 106, and a display as shown in FIG. A display 109 provided as a device is connected to a bus.

  In one aspect, the image processing apparatus 100 is realized by cooperation of hardware resources and software. Specifically, various functions of the image processing apparatus 100 are realized by the CPU 101 executing programs recorded on a recording medium such as the ROM 102 and the external storage device 104. Further, the image processing apparatus 100 may be physically realized by a single device or may be realized by a plurality of devices.

  FIG. 3 is an example of a flowchart when the object detection unit 12 detects an object from an input frame image in the image processing apparatus of the present invention. The object detection unit 12 receives a frame image input from the image storage unit 10 (step 301). The object detection unit 12 starts detection processing of an object window in which an object in the frame image that has received the input exists.

  The object detection unit 12 reads the scan table from the scan table storage unit 11 and determines the window size for scanning the input frame image (step 302). The window size is determined based on, for example, a predetermined scan layer. The input frame image may be scanned with the window size of one or two of the scan layers stored in the scan table. Alternatively, the scan processing of all the scan layers stored in the scan table storage unit 11 may be performed.

  Next, the object detection unit 12 sets a calculation window in the input frame image based on the determined window size (step 303). The number of windows required to scan the entire input frame image varies depending on the window size, but the position of the window for calculating the feature amount in the input frame image is set based on a predetermined window size. The position of the calculation window is determined by the window size and the number of scan windows in the window size.

  The object detection unit 12 generates a luminance gradient histogram in the block within the window set as the calculation target (step 304). Here, for example, the HOG feature value is used to detect an object that is the upper body. For this purpose, the luminance gradient in the horizontal and vertical directions is obtained from the four neighboring pixels in each pixel in the input frame image, and the luminance gradient intensity and direction are calculated. Then, the luminance gradient strength for each pixel is assigned to the number of bin images 9 every 20 degrees from 0 degrees to 180 degrees according to the luminance gradient direction. Filter each bin image using the kernel. You can use a 7x7 kernel here. Then, an integral image for each bin image is generated. As an example, a luminance gradient strength accumulation in each cell for a block is used as a histogram, luminance gradient direction histograms for four cells in the block are concatenated, and each block is normalized to obtain a 36-dimensional luminance gradient direction histogram in the block. Generate. Note that scanning means analyzing the feature amount of each pixel of an image.

  Further, the object detection unit 12 performs an overlap scan in the window with blocks (step 305). An overlap scan is performed based on the scan layer and cell width stored in the scan table. That is, the brightness gradient histogram in the block is sequentially created while shifting the position in the window by a predetermined cell width. That is, a total of 49 scans are performed in the window, 7 times in the horizontal direction and 7 times in the vertical direction. At this time, the object detection unit 12 uses the feature amount calculated when the previous block is scanned as the feature amount of the overlap region. This is for speeding up the object detection process by reusing the calculation result.

  Then, the object detection unit 12 calculates the feature amount in the calculation window (step 306). As an example, the object detection unit 12 obtains a 1764-dimensional HOG feature value in the window by performing 49 scans in the calculation window.

  The object detection unit 12 identifies the feature amount in the calculation window by SVM (step 307). Here, a support vector machine (SVM) is used as the classifier, but the classifier is not limited to this. HOG feature values are learned in advance from an image that includes the upper body and an image that does not include the upper body, and an SVM classifier is obtained. Using the SVM classifier, the object detection unit 12 detects a window having a feature value that is the upper body of a person as an object window. If it is determined that the calculation window is an object window, the process proceeds to the clustering integration unit 13.

  The object detection unit 12 determines whether or not the feature amount has been calculated for all the windows in the predetermined scan layer (step 308). If the feature amount calculation has not been completed for all windows, the next calculation window is set (step 303). Similarly, the feature amount is calculated. When the feature amount calculation has been completed for all windows, the object detection unit 12 ends the object detection process in the scan layer.

  In the above description, the object detection process in one scan layer has been described. However, in the case where all the scan layers stored in the scan table are performed, or when the process of two or more scan layers is performed, the process described with reference to FIG. The flow is performed in the same manner for a predetermined number of times. Changing the scan layer changes the window size. Therefore, when the size of the person shown in the image is various, it is desirable to perform the object detection process with a plurality of scan layers.

  FIG. 4 is an example of a scan table stored in the scan table storage unit 11. This scan table is an example when the resolution of the input image is 640 × 480 pixels. A scan layer number is assigned for each window size, and a window size, a cell size, a scan stride value, and the number of windows are stored in association with the scan layer. The window size is a window that is set when an object is detected in the input image. Here, the window size is represented by the number of pixels. The cell size is the cell size set in the window as described above. The scan stride value is a pixel distance value by which the window is shifted in order to scan the entire image. The number of windows indicates the number of windows necessary to scan the entire image. The object detection unit 12 scans the input frame image while referring to the scan table stored in the scan table storage unit 11.

  FIG. 5 is a conceptual diagram showing the relationship between cells and blocks. In step 304 of FIG. 3, the object detection unit 12 provides a block in the window and generates a luminance gradient histogram. At this time, cells are provided in the block. In each cell, the luminance gradient in the horizontal and vertical directions is obtained, and the cell 1, cell 2, cell 3, and cell 4 are connected in this order to generate a 36-dimensional in-block luminance gradient direction histogram.

  FIG. 6 is a conceptual diagram showing the relationship between windows and blocks. In one example, each area obtained by dividing the window into four equal parts in the horizontal and vertical directions is used as a block, and four areas obtained by further dividing the window into two equal parts in the horizontal and vertical directions are used as cells. Since overlap scanning is performed by shifting the cell width, the dotted block area is the next block for obtaining the luminance gradient. In this way, when a block is subjected to overlap scan, in the next block scan, the previous block and two cells overlap (overlap). Therefore, the object detection unit 12 reuses the feature amount of the cell as it is when the brightness gradient of the previous block is calculated in the calculation of the feature amount of the two cells. Here, scanning is first performed in the vertical direction, and then scanned in the horizontal direction. However, the present invention is not limited to this, and scanning may be performed first in the horizontal direction and then in the vertical direction.

  FIG. 7 shows an example of processing when the object tracking unit 14 associates the same object area between frame images in the image processing apparatus of the present invention. The object tracking unit 14 sets the region set as the object region by the clustering integration unit 13 in the input frame image as the object candidate region (step 701).

  Next, the object tracking unit 14 sets an object prediction region in the input frame image based on the object region in the past frame image (step 702). For example, the object tracking unit 14 reads past frame images such as a frame image two frames before and a frame image one frame before from the input frame image stored in the image storage unit 10. Then, the amount of change in coordinates between the object area in the frame image two frames before and the object area in the frame image one frame before, which is associated with the object area in the frame image two frames before Calculate Then, the object tracking unit 14 calculates and sets the object prediction area in the input frame image based on the coordinates of the object area in the frame image one frame before and the calculated change amount of the coordinates. As another example, a predetermined change amount may be set, and the object prediction area in the input frame image may be calculated based on the object area in the frame image one frame before. In this case, by setting the coordinate change amount in advance based on the moving speed of the object target such as the walking speed of a human, the object prediction area can be set even when the object is not shown two frames ago. It becomes possible to do. As another example, an object prediction based on an object area of a frame image several frames before and a predetermined coordinate change amount using a frame image several frames before instead of a frame image one frame previous as a past frame image. An area may be set. That is, the past frame image is not limited to the previous frame image.

ここで、Ｔは、所定の閾値である。 The object tracking unit 14 extracts an object candidate area in which a difference in distance and window width from the object prediction area is within a predetermined range (step 703). Specifically, the coordinates of the upper left point of the rectangle of the window is an object prediction area _{P (a (x 1),} a (y 1)) and then the window width of the object prediction region and W _a, window object candidate region When the coordinate of the upper left corner of the quadrangle is P (b (x ₁ ), b (y ₁ )) and the window width of the object candidate area is W _b , the object tracking unit 14 satisfies the following four conditions. Extract candidate areas.

Here, T is a predetermined threshold value.

  (2) and (4) in the above formulas are the object areas in which the difference in coordinate values is within a predetermined threshold range among the object candidate areas, that is, within a predetermined distance from the object prediction area to the object candidate area. An object candidate area is to be extracted. Further, (3) and (5) are intended to extract object candidate areas whose difference from the window width of the object prediction area is within a predetermined range from the object candidate areas. Here, the object tracking unit 14 extracts a window candidate area that satisfies these four conditions, and sets the extracted object area as a related object candidate area.

In addition, although T which is a threshold value can be set to arbitrary values, you may determine T based on the following formula as an example. β is an arbitrary variable.

  Next, the object tracking unit 14 extracts the object candidate area that is the shortest distance from the extracted object candidate areas, that is, the related object candidate areas (step 704). An object candidate area having a coordinate value closest to the object prediction area, that is, the shortest distance is extracted from the related object candidate areas that satisfy the predetermined condition.

  Further, the object tracking unit 14 checks whether or not the extracted object candidate area is associated with another object prediction area (step 705). This is because the extracted object candidate area may already be associated with another object prediction area. This check may be related to the “object prediction area” or may be replaced with “an object area of a past frame image (for example, a frame image one frame before) corresponding to the object prediction area”. When the extracted object candidate area is associated with other objects (Yes in step 705), the object that is not the shortest is associated with only the object prediction area that is the shortest distance from the extracted object candidate area among the plurality of object prediction areas The association with the prediction region is discarded (step 706).

  Next, for the discarded object prediction area, an object candidate area having the shortest distance is extracted from the remaining object candidate areas among the related object candidate areas extracted for the object prediction area (step 707). The discarded object prediction area is recalculated because the corresponding object candidate area disappears. The re-extraction is performed by extracting the related object candidate region with the shortest distance from the related object candidate regions extracted in advance.

  If the extracted object area is not associated with another object prediction area (No in step 705), a person ID associated with the object prediction area is assigned (step 708). That is, a person ID that is an identifier assigned to the object area in the past frame image (for example, the frame image one frame before) corresponding to the object prediction area is assigned. By assigning the same identifier, object tracking can be performed between frames. The assigned person ID may be displayed on the display unit in the vicinity of the object area.

FIG. 8 is an image diagram of association of object areas between frame images. The frame images I ₁ , I ₂ , and I ₃ are frame images acquired in time series. For example, when I ₃ is an input frame image, I ₁ is a frame image two frames before and I ₂ is one frame previous It is a frame image. When each detected object area exists and the object is a moving object such as a human, even if the object is the same object as shown in FIG. 8, the detection position in each frame image is different. Therefore, in order to associate these objects, the object tracking unit 14 realizes object tracking by associating objects between frame images according to the flow shown in FIG.

  FIG. 9 is a diagram showing the relationship between the object prediction area and the object candidate area in step 703 of FIG. A is an object window which is an object prediction area, and B is an example of a window of an object candidate area. As described above, whether or not the difference between the position of the coordinate P of the upper left corner of the object prediction area and the window width W is within the threshold is determined based on the four expressions, and object candidate areas that satisfy the conditions are extracted.

FIG. 10 is a table showing experimental results when object detection is performed in the image processing apparatus of the present invention. It shows the time taken to calculate the feature values for each scan layer and window size. In the present invention, in the object detection, overlap scanning is performed to calculate the window feature amount. At this time, the object detection processing is speeded up by reusing the past calculation result for the feature amount of the overlap region for each scan layer. As a result, even when a feature amount that can be detected with high accuracy such as a HOG feature amount is employed, object detection can be performed in a very short time as shown in the figure. Therefore, even when frame images are input one after another, object detection can be performed in real time.

Claims (10)

An object detection unit for detecting an object window in which an object exists from an input frame image;
A clustering integration unit for setting an object region by integrating a plurality of neighboring object windows detected by the object detection unit by clustering; and
The object area is an object candidate area, an area predicted based on an object area of a past frame image is an object prediction area, and the object prediction area has a window width difference within a predetermined range, and the object prediction An object tracking unit that registers an object candidate area that is at the shortest distance from an area in association with an object area of the past frame image.   The object tracking unit extracts an object candidate area having a window width difference within a predetermined range and within a predetermined distance from the object prediction area as a related object candidate area from among the plurality of object candidate areas. The image processing apparatus according to claim 1, wherein among candidate areas, a related object candidate area located at a shortest distance from the object prediction area is registered in association with the object area of the past frame image.   When the object candidate region is associated with two or more object prediction regions, the object tracking unit registers only the object region of the past frame image corresponding to the object prediction region at the shortest distance from the object candidate region. The image processing apparatus according to claim 1, wherein the image processing apparatus is an image processing apparatus.   The image according to claim 1, wherein the object tracking unit associates and registers an identifier identical to an identifier assigned to an object area of the past frame image by assigning to the object candidate area. Processing equipment.   The object tracking unit is based on the first object area of the frame image two frames before and the amount of change in coordinates in the second object area of the frame image one frame before associated with the first object area. The image processing apparatus according to claim 1, wherein an object prediction area in the input frame image is determined.   The object tracking unit may change the object prediction area in the input frame image by changing a predetermined amount from the coordinates of the object area in the previous frame image when there is no object area associated with the previous frame image. The image processing apparatus according to claim 5, wherein:   The object detection unit calculates a feature amount of the window by providing a window in the input frame image and performing an overlap scan within the window with a block having a predetermined area, and detects the object window based on the calculated feature amount. The image processing apparatus according to claim 1, wherein:   The object detection unit calculates a feature value of a window by reusing a feature value once calculated as a feature value of an overlap region for each scan layer when performing an overlap scan. Item 8. The image processing apparatus according to Item 7. Detecting an object window in which an object exists from an input frame image;
A step of setting an object region by integrating a plurality of detected object windows in a cluster by clustering;
The object area is an object candidate area, an area predicted based on an object area of a past frame image is an object prediction area, and the object prediction area has a window width difference within a predetermined range, and the object prediction An image processing method comprising a step of registering an object candidate area at a shortest distance from an area in association with an object area of the past frame image. On the computer,
Detecting an object window in which an object exists from an input frame image;
A step of setting an object region by integrating a plurality of detected object windows in a cluster by clustering;
The object area is an object candidate area, an area predicted based on an object area of a past frame image is an object prediction area, and the object prediction area has a window width difference within a predetermined range, and the object prediction Registering the object candidate area at the shortest distance from the area in association with the object area of the past frame image;
A program that executes

Images