JP6516609B2 - Moving object detection apparatus and background model construction method - Google Patents

Description

  The present invention relates to a moving object detection apparatus for accurately detecting a moving object from a video taken by a fixed camera by a background subtraction method using a background image, and a background model constructing method thereof.

  A technique for accurately detecting moving objects by removing the effects of imaging noise, slight illumination fluctuation and periodically moving areas from fixed camera images is very important as a preprocessing of surveillance video analysis technology, It can be used for behavioral analysis and abnormal behavior detection.

  In order to detect a moving object from a fixed camera image, a background subtraction method has been proposed in which the difference between front and rear frames of the image is recognized as a moving object. Various moving object detection methods based on this background subtraction method have been devised. ing.

  Among these, Patent Document 1 and a non-patent document 1 and a method that can solve the problems such as imaging noise, slight illumination fluctuation, and the influence of a periodically moving area, and reduce the memory amount of background model retention while performing online processing. It is disclosed in Patent Document 1.

  In these prior arts, a background model is created based on a background image in which no moving object exists or a moving image in which there is almost no moving object in consideration of the entire frame, and a comparison is made between an input frame for detecting the moving object and the background model. Can detect the moving object.

  According to Patent Document 1, a background model is generated by dividing the background image into a two-dimensional grid (size W × H), and eight directions of upper, lower, left, and right and oblique to each pixel (target pixel) of the background image. Search for pixels. Then, the position of the pixel (reference pixel) whose absolute value of the difference in lightness with the pixel of interest is equal to or greater than the threshold Tp and the magnitude relation (positive or negative) of the lightness difference are obtained respectively, calculate. Here, the threshold Tp is determined from the imaging noise distribution.

  In the detection of the moving object from the input frame, the position in eight directions for each pixel generated by the background model and the code value (8 bits) of the lightness difference are similarly calculated for the input frame, and the code values of the lightness difference match This is performed by detecting a pixel whose rate is equal to or less than the threshold Tb as an animal. The threshold Tb is set to about "6" if the number of reference pixels is "8".

  Although the basic method of Non-Patent Document 1 is almost the same, in addition to performing the above processing for each background image using a background image at the time of background model construction, a constant brightness difference is obtained across all frames of the background image. By using only the positions of the generated pixels for comparison, it is possible to realize moving object detection that is more robust to illumination variations.

JP 2003-141546 A

Iwata, Sato, Ozaki, Sakagami, "Robust background subtraction based on statistical reach feature method," Transactions of the Institute of Electronics, Information and Communication Engineers, D Vol. J92-D No. 8 pp. 1251-1259, 2009.

  In the above-described conventional method, the reference pixel calculation method utilizes the fact that the absolute value of the lightness difference with the pixel of interest is equal to or greater than the threshold Tp, so the code value of the lightness difference between the reference pixel and the pixel of interest is biased. There is a technical problem that the accuracy of the moving object detection decreases if there is

  Here, that the sign value of the lightness difference is biased means that the lightness of all the reference pixels is larger than the target pixel (for example, the proportion of “0” in the 8-bit code value is high) or all the reference pixels Point when the lightness of is smaller than the target pixel (for example, the proportion of “1” in the 8-bit code value is high).

  For example, if the lightness of all the reference pixels is greater than the target pixel, the magnitude relationship of the lightness (positive and negative of the lightness difference) does not change even if the lightness of the target pixel changes in the input frame. Can not be extracted as the pixel of the moving object.

  Similarly, when the lightness of all reference pixels is smaller than the target pixel, the magnitude relationship of the lightness does not change even if the lightness of the target pixel in the input frame changes in the direction of increasing. It can not be extracted as

  Furthermore, in the above-described conventional method, when creating a background model from a background image or video, a pixel satisfying a predetermined condition of lightness difference with each noted image does not exist in the search direction (hereinafter referred to as an exceptional condition) In some cases, there has been a technical problem that the accuracy of moving object detection decreases.

  Here, when it is not possible to search for a pixel satisfying the condition in the search direction, for example, as shown by the “x” mark in FIG. 11, the target pixel P of the search target is located near the end of the learning screen Vs. If the search range (distance) in the edge direction becomes short, or if the texture in the background area is small and the difference in lightness is difficult to occur, etc., in such a case, a pixel satisfying the condition can not be found. The probability is high.

  FIG. 12 is a diagram showing the background image for learning [FIG. (A)] and whether or not an exceptional condition has occurred under predetermined conditions in the background image [FIG. (B)]. It can be seen that the pixels for which the exceptional condition occurs are concentrated at the edge of the background image.

  For such technical problems, Patent Document 1 and Non-patent Document 1 (1) select pixels of maximum value (difference is less than Tp) in the scanning direction, (2) forcibly set pixels at the screen edge (See above, Patent Document 1), or there is no target pixel in the corresponding direction, and the code value comparison of the lightness difference in the same direction is not performed, that is, the code value determination result in the same direction is any input. The pixel is also dealt with by treating it as a background (above, Non-Patent Document 1).

  However, while the threshold Tp is set based on the imaging noise distribution, the determination based on the processing corresponding to the above exceptional condition is forced to be selected even if the brightness difference is a pixel less than the threshold Tp. Because of this, the condition for detecting the moving object is clearly determined gradually, and as a result, it has led to false detection or missing detection of the moving object.

  Furthermore, for the purpose of detecting a person's area indoors, (1) want to detect the moving object with the same accuracy as in the center of the screen as much as possible within the shooting range, (2) floor surface, table surface, etc. There is a need to detect moving objects with accuracy as in the case where there are regions with few textures but as much textures as possible, and improvements have been sought.

  In the object extraction method using GMM of Non-Patent Document 2, a GMM can be obtained for each pixel, but a GMM similar to the background image can be obtained for many pixels in the object, so distinction between the object and the background image Is difficult.

  A first object of the present invention is to solve the above technical problems and to provide a moving object detection device capable of accurately detecting a moving object from an input image and a background model construction method thereof.

  A second object of the present invention is a moving object detecting apparatus capable of accurately detecting a moving object from an input image even in moving object detection by the background subtraction method using a background image prone to exception processing, and a background model constructing method thereof To provide.

  In order to achieve the above object, the present invention is characterized in having the following configuration.

  (1) In the method of constructing a background model of the present invention, a procedure for sequentially selecting a pixel of interest from the background image and a pixel search in a plurality of radiation directions for each pixel of interest, and the feature amount difference satisfies a predetermined condition A search procedure including a search procedure for registering reference pixels in each radial direction, and a procedure for constructing a background model by associating the position of the reference pixel and the code value of the feature amount difference with the target pixel for each target pixel In this case, positive number reference value of the feature amount difference is positive and the difference width exceeds the first threshold, and predetermined number of negative side reference pixels where the difference value of the feature amount difference is negative and the difference width exceeds the second threshold I tried to search.

  (2) The moving object detection apparatus according to the present invention performs reference pixel search in a plurality of radiation directions for each pixel of interest selected from the background image, and registers reference pixels whose feature amount difference satisfies a predetermined condition for each radiation direction. Search means, means for correlating the position of the reference pixel and the code value of the feature amount difference with the target pixel for each target pixel to construct a background model, and comparing the input frame with the background model Means for identifying the pixel of the moving object based on the similarity of the code value of the feature value difference between the pixel of interest and each reference pixel, and the search means has a code value of the feature value difference positive and a difference width Positive side search means for searching for a predetermined number of positive side reference pixels exceeding one threshold, and negative side search for searching a predetermined number of negative side reference pixels for which the feature value difference has a negative code value and the difference width exceeds a second threshold Equipped with means.

  (3) From the set of reference pixels searched for each of the radial directions, some positive reference pixels and negative reference pixels closer to the target pixel are preferentially selected, and these selected reference pixels are selected. The background model is constructed by correlating the position of and the code value of the feature amount difference with the target pixel.

  (4) When a predetermined number of reference pixels can not be searched, the reference pixels for the missing number are searched using a different second search method.

  (5) In the second search method, the insufficient number of reference pixels are randomly searched on the background image.

  (6) In the second search method, the reference pixels corresponding to the missing part are searched in another radiation direction different from the radiation direction.

  (7) In the second search method, a two-dimensional array of a plurality of background images is assumed, and the missing number of reference pixels are searched on the extension of the radial direction.

  (8) The first and second threshold values are determined so that the code value of the feature amount difference does not change due to the temporal change of the background image.

  According to the present invention, the following effects are achieved.

  (1) Since the plurality of reference pixels set for each pixel of interest in order to construct the background model include both reference pixels having larger or smaller feature amounts than the pixel of interest, the pixel of interest is a moving object. Even if the feature quantity changes only to either the increase side or the decrease side, the code value of the feature quantity difference changes with any reference pixel, so the pixel Can be recognized as the pixel of the moving object.

  (2) Background of the Invention According to the method of constructing a model of the present invention, it is possible to search for a predetermined number of reference pixels satisfying a predetermined condition with a low search cost and with extremely high probability. Therefore, while maintaining the conventional object detection performance based on the background difference with respect to most areas of the image, imaging noise, some illumination variation, and periodicity also for the vicinity of the screen edge and the background area with few textures. It is possible to construct a background model capable of maintaining sufficient detection performance of the moving object by removing the influence of the minute moving area.

  (3) If a predetermined number of reference pixels can not be searched even if a pixel search is performed in a plurality of radiation directions for each pixel of interest, a predetermined number of reference pixels are randomly searched on the background image . Therefore, even if a search in a specific radial direction can not find a reference pixel exceeding the threshold because there is some regularity in the pixel distribution of the background image, etc., the reference pixel exceeding the threshold can be found with high probability. Become.

  (4) If it is not possible to search a predetermined number of reference pixels satisfying a predetermined condition even if pixel search is performed in a plurality of radial directions for each target pixel, a two-dimensional array of a plurality of background images is assumed and each search direction is Only the search range is expanded while maintaining it. Therefore, the parameter for specifying the search position and the reference pixel can be managed not by the absolute or relative XY coordinate values with reference to the target pixel but only by the distance from the target pixel It is possible to reduce the search cost such as the search time and the search time.

  (5) According to the moving object detecting device of the present invention, when a person moving into the field of view from outside the field of view of the monitoring camera starts to appear on the screen edge, it can be recognized as the moving object. Therefore, immediately after a person appears in an image, it is possible to calculate as a moving object candidate region by background subtraction processing, so that improvement in detection speed and accuracy of a person region can be expected.

It is a functional block diagram of a moving object detection device concerning a 1st embodiment of the present invention. It is the figure which showed the search method of the reference pixel. It is a figure for demonstrating the effect of this invention. It is a functional block diagram of a moving object detection device concerning a 2nd embodiment of the present invention. It is the figure which showed the 1st search method of the reference pixel by the 2nd search part. It is the figure which showed the 2nd search method of the reference pixel by the 2nd search part. It is the figure which showed the 3rd search method of the reference pixel by the 2nd search part. It is the flowchart which showed the construction procedure of the background model by this invention. It is a functional block diagram (the 1) of a moving object detection device concerning a 3rd embodiment of the present invention. It is a functional block diagram (the 2) of the moving body detection apparatus which concerns on the 3rd Embodiment of this invention. It is a figure for demonstrating the exceptional condition which can not find a reference pixel. It is the figure which showed the occurrence condition of the exceptional condition.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the main part of a moving object detection apparatus 1 to which the background model building method of the present invention is applied, in which a background model used for moving object detection by the background subtraction method is constructed in advance A moving object is detected by applying a background model to the image Vin input in.

  The reference pixel search unit 10 includes the positive side search unit 101 and the negative side search unit 102, and searches for candidates of reference pixels according to a predetermined rule for each pixel of the background image Vs. The reference pixel is a pixel for which the magnitude relation of the feature amount with the target pixel satisfies a predetermined condition for each focused pixel (target pixel) of the background image Vs, and as the feature amount, lightness, R, G , B can be adopted.

  FIG. 2 is a diagram showing an example of a method of searching for reference pixel candidates by the reference pixel search unit 10, and for each pixel of interest P selected from the background image Vs, each radiation direction starting from the pixel of interest P In the embodiment, pixel search is sequentially performed in the upper, lower, left, and right directions and eight oblique directions.

  The positive side search unit 101 compares the feature amount Tp of the target pixel P with the feature amount Tij of each pixel in the radial direction in order of distance from the target pixel P in each radial direction, A pixel in which the difference (Tij−Tp) becomes “positive” and the difference width ΔT first becomes equal to or larger than the first threshold Tp_ref1 is registered as the positive reference pixel candidate p1.

  The negative side search unit 102 compares the feature amount Tp of the target pixel P with the feature amount Tij of each pixel in the radial direction in order of distance from the target pixel for each radial direction, A pixel in which Tij−Tp) is “negative” and the difference width ΔT is initially not less than the second threshold Tp_ref2 is registered as the negative reference pixel candidate p2.

  Therefore, in the present embodiment, a total of 16 reference pixel candidates are searched for one each of the positive side reference pixel candidate p1 and the negative side reference pixel candidate p2 for each of eight search directions. The first and second threshold values Tp_ref1 and Tp_ref2 may be the same value or different values.

  The reference pixel selection unit 20 selects a part of the reference pixel candidate as a reference pixel. In the present embodiment, the distances from the pixel of interest P are compared for eight positive reference pixel candidates p1, and the top four positive reference pixel candidates p1 with short distances are preferentially adopted as reference pixels.

  Similarly, the distances from the target pixel P are compared for eight negative reference pixel candidates p2, and the top four negative reference pixel candidates p2 with short distances are preferentially adopted as reference pixels. As a result, a total of eight reference pixels are selected in the present embodiment.

  At this time, either the positive reference pixel candidate p1 or the negative reference pixel candidate p2 may be selected from the same search direction, or a restriction may be provided such that only one is selected. .

  The background model construction unit 30 constructs a background model M based on the positions of a plurality of (here, eight) reference pixels selected by the reference pixel selection unit 20 and the code values (8 bits) of the feature amount difference. .

  Next, moving object detection using the background model M will be described. The input pixel code value calculation unit 40 of the moving object detection device 1 calculates, for each pixel (target pixel) of the input image Vin, a feature amount with each pixel at a position corresponding to the reference pixel selected by the reference pixel selection unit 20. The difference is determined and the code value is similarly calculated.

  The moving object identification unit 50 compares the code value of each reference pixel with the corresponding pixel of the background model M for each pixel of the input image Vin, and determines the pixel whose number of coincidence of code values is equal to or less than a predetermined threshold value. Identify the pixels of the body and identify the others as the pixels of the background image.

  Such a moving object detection device 1 can be configured by mounting an application (program) for realizing the above-described functions on a general-purpose computer or server. Alternatively, a part of the application may be configured as a dedicated machine or a single-purpose machine in which hardware or ROM is implemented.

  FIG. 3 shows the detection result of the moving object according to the related art [the same drawing (b)] for the frame of the input image Vin [the same drawing (a)] and the moving object detection result of the present embodiment [the same drawing (c)]. It is a figure shown and it turns out that the loss of the moving object in the case where the lightness value of a photographic subject is small is eliminated in this embodiment.

  As described above, according to the present embodiment, a plurality of reference pixels set for each pixel of interest to construct a background model include both reference pixels with large and small feature amounts of the pixel of interest than the pixel of interest. Therefore, even if the pixel of interest is a moving object, and the feature quantity changes only to either the increase side or the decrease side, the code value of the feature quantity difference changes with any of the reference pixels. Therefore, the pixel can be recognized as a pixel of the moving object.

  FIG. 4 is a functional block diagram showing the configuration of the second embodiment of the present invention, and the same reference numerals as described above indicate the same or equivalent parts, so the description thereof will be omitted.

  In the present embodiment, when the (first) primary side search unit 101 and the (first) negative side search unit 102 can not search for reference pixel candidates within the search range, the reference pixels are searched using different search methods. A feature is that a second positive side search unit 103 and a second negative side search unit 104 are provided.

  The first positive side search unit 101 and the first negative side search unit 102 select reference pixels whose feature amount difference satisfies a predetermined condition even if searching is performed to the end of the background image Vs in the radial direction for each pixel of interest. When the search can not be performed, the second positive search unit 103 and the second negative search unit 104 are requested to search for the missing reference pixels to the second search units 103 and 104, respectively.

  The second positive side search unit 103 searches for the requested number of reference pixels from the directions other than the respective radiation directions, and the feature amount difference (Tij−Tp) with respect to the target pixel becomes “positive”, and the feature amount difference A pixel whose width ΔT first becomes equal to or larger than the first threshold Tp_ref1 is additionally registered as the positive reference pixel candidate p1.

  Similarly, the second negative side search unit 104 searches for the requested number of reference pixels from the directions other than the radiation directions, and the feature amount difference (Tij−Tp) with respect to the target pixel becomes “negative”, and A pixel in which the width ΔT of the feature amount difference first becomes equal to or larger than the second threshold Tp_ref2 is additionally registered as the negative reference pixel candidate p2.

  FIG. 5 is a view schematically showing a first search method by the second search units 103 and 104. The same reference numerals as in the above denote the same or equivalent parts.

  In the illustrated example, the positive reference pixel candidate p1 and the negative reference pixel candidate p2 can all be searched for a total of three directions of upper, left, and upper left around the pixel of interest P. On the other hand, with respect to the other search directions, at least one of the positive reference pixel candidate p1 and the negative reference pixel candidate p2 is searched because the distance to the image edge is short and the search range is narrow mainly. Not done.

  Therefore, each of the second search units 103 and 104 randomly performs an image search on the entire area of the background image Vs. Then, pixels with feature amount differences satisfying the respective conditions are added as reference pixel candidates p1 and p2. Such random search of reference pixel candidates is repeated for the number of reference pixels for which the first search units 101 and 102 could not search.

  Note that the method for randomly searching for the reference pixel candidates p1 and p2 is not limited to the above, and the radiation direction centered on the pixel of interest P is not limited to the second search method shown in FIG. It is also possible to add at random the requested number of insufficiency and to carry out the image search in each of the radial directions.

  According to this search method, since the reference pixels are searched at random, there is no regularity in the pixel distribution of the background image, etc. When a search limited to a specific radiation direction can not find a reference pixel that exceeds the threshold However, it is possible to find reference pixels that exceed the threshold with high probability.

  FIG. 7 is a diagram schematically showing a third search method by the second search units 103 and 104. The same reference numerals as in the above denote the same or equivalent parts.

  In this embodiment, a search range in which a plurality of identical background images Vs are two-dimensionally tiled is assumed. Then, on the extension of the radial direction in which the first search units 101 and 102 could not search for the reference pixels for each pixel of interest, the search for the reference pixels was performed by ignoring the boundary between the background images The corresponding pixel in each background image Vs of the pixels is added to the reference pixel.

  According to this search method, only the search range can be expanded while maintaining each search direction. Therefore, the parameters for specifying the search position and the reference pixel can be absolute or relative XY coordinates with reference to the target pixel P It can manage only by the distance from the focused pixel P, not the value. Therefore, it is possible to reduce search cost such as memory consumption and search time at the time of search.

  FIG. 8 is a flow chart showing the construction procedure of the background model M in the second embodiment, which is repeated for each image Vin input in frame units.

  In step S1, a background image Vs is acquired. In step S2, one pixel is selected from the background image Vs, and is set as the current pixel P of interest. In step S3, the first search units 101 and 102 search for a predetermined number (16 in total from eight search directions in this embodiment) of the target pixel P for the target pixel P.

  In step S4, it is determined whether or not a predetermined number of reference images have been searched for the current target pixel P. Here, if a predetermined number of reference pixels can not be searched due to the occurrence of the exceptional condition, the process proceeds to step S5, and the second search units 103 and 104 search for the reference images of the insufficient number.

  Note that GMMs of lightness values may be created for target pixels for which the second search units 103 and 104 can not search for a predetermined number of reference pixels, and a moving object may be detected according to the model.

  In step S6, a part of the searched many reference pixel candidates is selected as a reference pixel from the viewpoint of the accuracy of the moving object detection and the calculation cost. In the present embodiment, from the set of the positive side reference pixel candidate p1 and the set of the negative side reference pixel candidate p2, the top four at a short distance from the target pixel P are selected and set as reference pixels.

  In step S7, it is determined whether the predetermined number of reference pixel search has been completed for all the pixels of the background image. If the process has not been completed, the process returns to step S2, and the pixel of interest is switched to repeat the above processes. In step S8, a background model M is constructed as a set of positions and code values of each reference pixel.

  In the above embodiment, although the positive side reference pixel p1 and the negative side reference pixel p2 are selected by the same number for each target pixel P, the present invention is not limited to this, and positive If the side reference pixels p1 and the negative side reference pixels p2 are selected by at least a number larger than the determination threshold of the moving object, the numbers may not be the same.

  That is, if it is determined that the pixel of the object is the same as the pixel of the same object when the number of reference pixels matching the code value among eight reference pixels is less than six, at least three positive reference pixels p1 and at least three negative reference pixels p2 It is desirable to be included.

  Next, moving object detection using the above background model M will be described. The input pixel code value calculation unit 40 of the moving object detection device 1 focuses on each pixel of the input image Vin, and each pixel corresponding to the reference pixel searched for each target pixel of the background image Vs by the reference pixel search unit 10 The feature amount is determined from the above and the code value is similarly calculated.

  The moving object identification unit 50 compares the code value of each reference pixel with the corresponding pixel of the background model M for each pixel of the input image Vin, and determines the pixel whose number of coincidence of code values is equal to or less than a predetermined threshold value. Identify the pixels of the body and identify the others as the pixels of the background image.

  FIGS. 9 and 10 are functional block diagrams showing the configuration of the third embodiment of the present invention. The same reference numerals as described above denote the same or equivalent parts, and therefore the description thereof will be omitted.

  In each of the above embodiments, it is assumed that the brightness of the background image does not change with time. However, under the condition that the feature value of each pixel of the background image changes with time regardless of the presence or absence of the moving object, the change of the feature value exceeds the threshold depending on the setting of the threshold, even in pixel regions other than the moving object. It can not be denied that it may be mistaken as a pixel of an animal.

  Therefore, in the present embodiment, when the background image Vs is provided as a moving image, the temporal variation of the feature amount of each pixel is analyzed, and the first threshold Tp_ref1 and the second threshold Tp_ref2 are optimized for each pixel of interest. It is characterized in that the determination unit 60 is provided.

  In the threshold determination unit 60, the background video analysis unit 601 analyzes temporal variation of the feature amount for each pixel of the background video. The threshold values setting units 602 and 603 on the positive side and the negative side are searched for in consideration of the feature amount variation of each pixel so that the feature amount variation of each pixel in the background image does not affect the moving object determination. The first and second threshold values Tp_ref1 and Tp_ref2 are set for each pixel.

  In the present embodiment, the feature amount difference does not satisfy the condition for adopting the pixel as a reference pixel depending on only the time-series fluctuation of the pixel of interest and each fluctuation component of the time-series fluctuation of the pixel at each search position. The first and second threshold values Tp_ref1 and Tp_ref2 are set.

  For example, if the feature amount of a certain target pixel P fluctuates in the range of Tp ± ΔTp and the feature amount of the pixel pij at each search position fluctuates in the range of Tij ± ΔTij, the pixel pij is referred to as a reference pixel (candidate) It is desirable that each of the threshold values Tp_ref1 and Tp_ref2 adopted as is larger than ΔTp + ΔTij.

  According to the present embodiment, even when the feature value of each pixel of the background video changes with time, it can be accurately recognized whether each pixel of the input video is a moving image.

  DESCRIPTION OF SYMBOLS 10 ... Reference pixel search part, 20 ... Reference pixel selection part, 30 ... Background model construction part, 40 ... Input pixel code value calculation part, 50 ... Moving body identification part, 60 ... Threshold value determination part, 101 ... 1st positive side search Part 102: first negative search part 103: second positive search part 104: second negative search part 601: background image analysis part 602: positive side threshold setting part 603: negative side threshold setting Department

Claims (14)

In the method of constructing a background model used for detecting a moving object by the background subtraction method,
A step of sequentially selecting a pixel of interest from the background image;
A search procedure for performing pixel search in a plurality of radiation directions for each pixel of interest, and registering, for each radiation direction, reference pixels whose feature amount difference satisfies a predetermined condition;
And, for each pixel of interest, associating the position of the reference pixel with the code value of the feature amount difference with the pixel of interest to construct a background model.
In the search procedure, positive side reference pixels in which the code value of the feature amount difference is positive and the difference width exceeds the first threshold, and negative side reference pixels in which the code value of the feature amount difference is negative and the difference width exceeds the second threshold A background model construction method characterized by searching by a predetermined number. Including a step of preferentially selecting some positive reference pixels and negative reference pixels closer to the target pixel from the set of reference pixels searched for each of the radial directions,
The background model according to claim 1, wherein in the step of constructing the background model, the background model is constructed by associating the position of the selected reference pixel with the code value of the feature amount difference with the target pixel. How to build a model.   The second search procedure according to claim 1 or 2, further comprising a second search procedure for searching for a predetermined number of reference pixels by a different search method when a predetermined number of reference pixels can not be searched in the search procedure. Background model construction method.   4. The background model construction method according to claim 3, wherein the second search procedure randomly searches the background image for the insufficient number of reference pixels.   4. The background model construction method according to claim 3, wherein the second search procedure searches for a few minutes of reference pixels in another radiation direction different from the radiation direction.   4. The background according to claim 3, wherein the second search procedure assumes a two-dimensional array of a plurality of background images, and searches for the missing number of reference pixels on an extension of the radial direction. How to build a model. In a moving object detection apparatus that detects an moving object by comparing an input frame with a background model,
Search means for performing reference pixel search in a plurality of radiation directions for each pixel of interest selected from the background image, and registering reference pixels for which the feature amount difference satisfies a predetermined condition for each radiation direction;
A means for correlating the position of the reference pixel and the code value of the feature amount difference with the target pixel for each target pixel to construct a background model;
Means for comparing the input frame with the background model and identifying the pixel of the moving object based on the similarity of the code value of the feature amount difference between the corresponding pixel of interest and each reference pixel;
The search means is
Positive side search means for searching for a predetermined number of positive side reference pixels in which the code value of the feature amount difference is positive and the difference width exceeds the first threshold;
A moving object detection apparatus comprising: a negative search means for searching for a predetermined number of negative reference pixels in which the code value of the feature amount difference is negative and the difference width exceeds the second threshold. The apparatus further comprises means for preferentially selecting some positive reference pixels and negative reference pixels closer to the target pixel from the set of reference pixels searched for each of the radiation directions,
The animal according to claim 7, wherein the means for constructing the background model constructs the background model by correlating the position of the selected reference pixel and the code value of the feature amount difference with the target pixel. Body detection device.   9. The apparatus according to claim 7, further comprising second search means for searching for a predetermined number of reference pixels by a different search method when the search means can not search a predetermined number of reference pixels. The moving object detection device as described.   10. The moving object detection apparatus according to claim 9, wherein the second searching means searches for the reference pixels of the missing number at random on the background image.   10. The moving object detection apparatus according to claim 9, wherein the second searching means searches for the reference pixels corresponding to the shortfall in another radiation direction different from the radiation direction.   10. The animal according to claim 9, wherein the second searching means assumes a two-dimensional array of a plurality of background images, and searches for the missing number of reference pixels on an extension of the radial direction. Body detection device. 13. The apparatus according to any one of claims 7 to 12, further comprising threshold value determining means for determining the first and second threshold values so that the code value of the feature amount difference does not change due to the temporal change of the background image. The moving object detection device as described in. The background image is a moving image, and
A means for analyzing the time-series variation of the feature amount for each pixel of the background image;
Threshold setting means for setting the first and second thresholds such that the feature amount difference does not satisfy the predetermined condition only by the time-series fluctuation of the pixel of interest and each fluctuation component of the time-series fluctuation of the pixel at each search position The apparatus according to any one of claims 7 to 13, further comprising: