JP2007026320A - Template matching method and image processor using the method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a template matching method capable of efficiently allowing a matching object image to correspond to a template image without being affected by a difference in size or minute change, and to provide an image processor using the method. <P>SOLUTION: A template processing part 4 obtains correspondence so as not to reverse an anteroposterior arrayal sequence between whole pixels, which correspond to a one-dimensional numeral value string at the side of the matching object image from the attribute value of each pixel on the contour line of the matching object image, and the whole pixels, which correspond to the one-dimensional numerical value string at the side of the template image previously stored in a template storage part 5. A difference amount calculating means 42 calculates the amount of difference between the numerical values of the corresponding pixels. Then the minimum total amount among the total amounts of the obtained difference amounts is outputted to a matching determining part 6. A matching determining part 6 compares the difference amount outputted from the template processing part 4 with a threshold and determines whether the matching object image is the same as the template image or not. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a template matching method and an image processing apparatus using the same.

  Conventionally, in the monitoring image processing, there is a background difference method as a commonly used method for detecting a moving object. In this background subtraction method, an image with no moving object shown in FIG. 12B is stored as a background image, and the background image is subtracted from a detection input image (FIG. 12A) that is sequentially input. Thus, the difference value is obtained. That is, when an object not in the background image appears in the input image, the difference value is not zero. The area where the difference value is not 0 is the moving object area, that is, the area of the matching target image. For example, a moving person (black portion in FIG. 12C) or a car image can be extracted in a plane.

  As a method for determining whether the extracted object is really a person or not, template matching is generally performed in image processing (for example, Patent Document 1).

  In general template matching, from the density values I (i, j) and T (i, j) for each pixel of a matching target image composed of a template image and a grayscale image extracted from an input image, the following equation 11 is used. The residual R shown and the cross-correlation coefficient C shown in Expression 12 (collectively referred to as “correlation value”) are obtained, and the degree of coincidence between the two images is determined from the correlation value.

JP-A-8-24946 (page 3-7, FIG. 1)

  By the way, in order to perform planar template matching for determining whether or not a person is present, it is necessary to match the position and size of the template with the position and size of the extracted object image. . However, it takes time to perform template matching for alignment, size change, and minute shape difference, and there are cases where erroneous detection and misjudgment occur.

  In other words, alignment is performed by searching the image with the highest correlation, and template images of a plurality of sizes are prepared for the size of the template image. Is searched while scanning the image. In this case, it is unclear how much the matching target image extracted from the input image has changed, and as a result, templates of various sizes are prepared. For example, even if templates of 0.9 times and 0.8 times are prepared, if the input image is 0.85 times, the correlation value decreases, and the possibility of erroneous detection misjudgment increases.

  On the other hand, if the judgment threshold value is lowered to absorb a slight variation of the shape, there is a possibility that a person other than a person is judged as a person. For example, as for the difference in fine shape, the shape of the extracted person differs depending on whether the hand is out, the degree of opening of the leg, the presence or absence of baggage, etc. The possibility increases.

  The above-described problem will be specifically described. When the gap between the legs of the person is small as in the matching target image shown in FIG. 13B compared to the person of the template image shown in FIG. When the matching process is performed using Equation 12), the position where the difference amount is the smallest is determined as the matched position.

  Therefore, I would like you to match at the position of FIG. 13 (c), but as shown in FIG. Although the position will be shifted, matching at this position is a mistake.

  In addition, compared with the person of the template image shown in FIG. 14A, the person of the matching target image shown in FIG. 14B is about twice the area ratio (about √2 times the size ratio) of the person of the template image. ) When the number is large, there are a plurality of positions having the smallest difference amount as shown in FIGS. 14C and 14D, but the magnification and the position are incorrect. Therefore, even if images having a person size from 1 times to 1.5 times are prepared as template images in increments of 0.1, it is determined that matching is performed at slightly different positions.

  Therefore, there is a problem that enormous processing time is required for matching at the correct position.

  The present invention has been made in view of the above points, and the object of the present invention is to efficiently associate the matching target image with the template image without being affected by a difference in size or a minute change. An object of the present invention is to provide a template matching method and an image processing apparatus using the template matching method.

  In order to achieve the above object, the template matching method according to claim 1 uses a one-dimensional numerical sequence consisting of numerical values based on the specification values of each pixel on the contour line of the template image and the matching target image. In performing the matching process, a process for associating all the pixels corresponding to the numerical sequence on the template image side with all the pixels corresponding to the numerical sequence on the matching target image side so that the order of arrangement before and after is not reversed is performed. The amount of difference between the obtained pixels is calculated and the sum of the amount of difference is calculated, and the association with the smallest obtained sum is determined to be the association with which the template image and the matching target image are most matched. It is characterized by doing.

  According to the template matching method of the present invention, the difference between the size of the template image and the matching target image is absorbed, and the matching target image and the template image are not affected even if there is a minute change. The best matching can be performed efficiently.

  In the invention of the template matching method of claim 2, in the invention of claim 1, if the sum of the differences determined to be the most matched correspondence is smaller than a preset threshold value, the matching target image and the template image are It is characterized by determining that they are the same.

  According to the template matching method of the present invention, the difference between the size of the template image and the matching target image is absorbed, and even if there is a minute change, the matching target image and the template target image are not affected. The matching judgment with can be performed accurately.

  In the invention of the template matching method of claim 3, in the invention of claim 1, each association when it is determined that the association with the smallest sum obtained is the association with which the template image and the matching target image are the most matched. The distribution of the difference amount between the pixels is examined, and it is determined whether the matching target image and the template image are the same based on the number of associations in which the difference amount is within a predetermined range.

  According to the template matching method of the present invention, by using the difference amount distribution state, it is possible to increase the accuracy of the matching determination in addition to the matching determination based on the sum of the difference amounts.

  According to a template matching method of a fourth aspect, in the invention according to any one of the first to third aspects, each numerical sequence is a tangential direction angle or normal line of the contour calculated based on a brightness change direction indicated by a pixel. It is a row in which direction angles are arranged.

  According to the template matching method of the present invention, the tangent or normal direction can be easily calculated by simply applying a differential filter to a contour with little unevenness.

  According to the invention of the template matching method of claim 5, in the invention of any one of claims 1 to 3, each numerical example is a row in which tangential direction angles or normal direction angles of the contour line are arranged, and on the contour line Using the pixel positions before and after a certain pixel of interest, the angle formed by the pixels before and after the pixel is the tangential direction angle of the contour line of the pixel of interest.

  According to the invention of the template matching method of the fifth aspect, an appropriate matching process can be performed on an image having severe unevenness.

  In a template matching method invention according to a sixth aspect, in the invention according to any one of the first to fifth aspects, the difference amount is a difference value of a tangential direction angle or a normal direction angle, and is constituted by two vectors. The smaller one of the angles is a feature.

  According to the template matching method of the present invention, a plurality of difference probabilities can be uniquely expressed depending on the numerical value, and therefore matching can be correctly determined.

  In a template matching method invention according to a seventh aspect, in the invention according to any one of the first to third aspects, the numerical value of the numerical sequence may be a luminance value of a pixel corresponding to a contour position or a color image if the image is a grayscale image. For example, the HLS value of the pixel corresponding to the contour line position is used.

  According to the template matching method of the present invention, it is effective when detecting a person using a human hair or skin color, or when tracking a specific person from sequentially input images.

  In the invention of the template matching method of claim 8, in the invention of any one of claims 1 to 3, a matrix is created based on the difference between the numerical sequence on the template image side and the numerical sequence on the matching target image side. A path that follows the grid where the cumulative added value of the grid has the smallest value while sequentially moving from the bottom grid to the top grid in the grid direction from the rightmost grid to the leftmost grid in each stage of the matrix. The route determination at the attention cell other than the rightmost cell at the lowermost stage, which is the starting point of the route, is performed by comparing the cumulative addition value of the cells existing on the right, lower right, and lower with respect to the attention cell. It is characterized by determining.

  According to the template matching method of the present invention, the determined route indicates the correspondence in which the template image and the matching target image are most matched, and the processing time when calculating the sum of the differences is reduced. Can be planned.

  In the template matching method invention according to claim 9, in the invention according to claim 8, when the cumulative addition value is larger than a threshold used for matching judgment at the time of determining the path of the target pixel, the cumulative addition value is compared at the time of path determination. It is excluded from the cumulative addition value.

  According to the template matching method of the present invention, the processing time can be further shortened.

  In the invention of the image processing apparatus according to claim 10, the contour line extracting means for extracting the contour line of the matching target image, and the primary value on the matching target image side from the specification value of each pixel on the contour line extracted by the contour line extracting means. Numerical value sequence generating means for generating the original numerical value sequence, all of the pixels corresponding to the one-dimensional numerical value sequence on the matching target image side generated by the numerical value sequence generating means, and the contour extracted from the template image stored in advance Corresponding all the pixels corresponding to the one-dimensional numerical sequence on the template image side obtained from the specification value of each pixel on the line so that the order of arrangement in the front and rear is not reversed, and the numerical value difference between the associated pixels Difference amount calculating means for calculating the amount and calculating the sum of the difference amounts, and the target image and the template image are the same based on the sum of the difference amounts determined by the difference amount calculating means Characterized by disconnection.

  According to the invention of the image processing device of claim 10, the difference in size between the template image and the matching target image is absorbed, and the image matching target image and the template image are not affected even if there is a minute change. Thus, it is possible to provide an image processing apparatus that can efficiently perform the best matching with the template image and can accurately determine the template image and the matching target image.

  The template matching method and the image processing apparatus using the same according to the present invention absorb the difference in size between the template image and the matching target image, and are not affected even if there is a minute change. There is an effect that the best matching between the template image and the template image can be performed efficiently.

  FIG. 1A shows the overall configuration of an image processing apparatus used for human detection according to an embodiment of the present invention. This image processing apparatus captures an image from a camera 1 that captures a monitoring area and an image from the imaging camera 1. After capturing, the image input unit 2 that performs A / D conversion and outputs, and the difference between the input image binarized by the image input unit 2 and the background image that has been captured and stored in advance in the monitoring area A region extraction unit 3 that extracts a region other than an image from an input image, a template storage unit 5 that stores data relating to a template image of a human body to be detected, and a region image and template extracted by the region extraction unit 3 A template processing unit 4 for performing a matching process by matching the images to be described later, and an image of an extraction region (hereinafter referred to as a matching target image) from the matching process in the template processing unit 4 A matching determination unit 6 for conducting identification determination of the template image, and a result output unit 7 for outputting a matching determination result.

  Here, the template storage unit 5 preliminarily calculates the tangential direction angle (or normal direction angle) using the predetermined coordinate information (hereinafter referred to as “N”) of each pixel on the outline of the person image that is the template image. A one-dimensional numerical value string A formed by arranging numerical values consisting of a numerical value of a pixel at a certain coordinate point and a numerical value of an end point pixel is stored as template image data.

  The area extraction unit 3 extracts the matching target image based on the difference between the background image and the input image, that is, the background difference, binarizes the image, and outputs it to the template processing unit 4. Alternatively, any other method may be used as long as it can extract the matching target region. For example, in a three-dimensional distance image obtained by stereo vision, if an object from a camera to a certain distance is cut out, a binarized matching target image can be easily obtained.

  As shown in FIG. 1B, the template processing unit 4 includes a contour line extracting unit 40 that extracts a contour line of a matching target image extracted by the region extracting unit 3 and binarized, and a contour line extracting unit 40. For each pixel on the extracted contour line, a numerical value sequence for obtaining a numerical value consisting of a tangential angle by a method described later, and generating a one-dimensional numerical value sequence B starting from a pixel at a predetermined position and ending with a pixel at the end point The generating unit 41, the numerical value sequence A from the template storage unit 5 and the association processing of each pixel of the numerical value sequence B extracted by the numerical value sequence generating unit 41, and the numerical difference value (difference amount) between the associated pixels And a difference amount calculating means 42 for performing a process of calculating and calculating the sum of the calculated difference amounts.

  Here, the operation of the contour line extracting means 40 will be briefly described with reference to FIG.

  First, the contour line extracting means 40, when taking a binarized matching target image (FIG. 2A), scans from the left end as shown in FIG. The detected white pixel a is used as a start point, and eight neighboring pixels around the start point are searched counterclockwise, and the pixel is moved to the pixel b of the first white pixel (lower left of the start point pixel in FIG. 2C). Next, since the pixel has moved from the upper right pixel a, 8 neighboring pixels around the pixel b are searched in the order of upper, upper left, left, lower left, lower, lower right, and right as shown in FIG. Move to the first pixel c (FIG. 2E). After this movement, since the pixel b has moved from the upper right pixel b as in the case described above, eight neighboring pixels around the pixel c are searched and moved to the first white pixel d. Thereafter, returning to the starting point or repeating until no white pixel is found, the contour line consisting of white pixels a to j and its coordinate sequence are extracted as shown in FIG.

  Next, a method for obtaining the numerical sequence used for matching from the extracted contour line by the numerical sequence generation means 41 will be described.

  In this embodiment, the numerical sequence is generated by a method using the tangential direction angle (or normal direction angle) of the contour line based on the contour shape. As one example, the tangential direction angle of the contour line is defined as the contour line. A method is employed in which the angle formed by two pixels using the Nth pixel position before and after the target pixel on the line is determined as the tangential direction angle of the contour line of the target pixel.

  Specifically, the numerical value string generating means 41 to which the contour line coordinate string shown in FIG. 3B extracted by the contour line extracting means 40 is input is shown on the contour line (A) as shown in FIG. The tangential direction angle is calculated using the coordinate information of the Nth pixel before and after the predetermined target pixel, and the calculation result is set as the contour tangential direction angle of the pixel.

  In this case, when a pixel having coordinates (Y, X) of (14, 12) is a pixel of interest and N is 4, the coordinates (16, 10) of the pixel four pixels before the pixel of interest and the pixel four pixels after Is calculated as follows using the coordinates of (10, 10).

dx = X (+ N) -X (-N) = 10-10 = 0
dy = Y (+ N) −Y (−N) = 10−16 = −6
Ang = atan2 (−dx, dy) = atan2 (0,16) = 270 degrees = −90 degrees. FIG. 3B shows the relationship with the coordinates of the fourth pixel before and after the target pixel (14, 12). In FIG. 3B, −4 to −1 indicate the fourth to first pixels after the target pixel, and +1 to +4 indicate the first to fourth pixels before the target pixel.

  FIG. 4A shows a binary image obtained by extracting the human body region from the image obtained by photographing the human body, and the contour extraction unit 40 performs contour extraction from the matching target image (white portion). The result is shown in FIG. The numerical value string generating means 41 performs pixel tracking on the extracted contour image in the arrow direction (b) from the position of the circle, calculates the tangential direction angle of each pixel as described above, and calculates the calculated result. A one-dimensional numerical sequence B is generated by arranging the tracking from the start point to the end point. FIG. 4C shows the angular direction. Table 1 is a table showing the coordinates of each pixel on the contour line in FIG. 4B and a part of the calculated contour line tangent direction angle.

  This method has an effect of absorbing a sudden angle change due to unevenness, that is, smoothing. In other words, there is an effect that the influence of minute unevenness due to noise can be avoided.

  The larger the N value, the smaller the unevenness and the greater the smoothing effect. Therefore, the N value is a value determined according to the fine structure required for template matching. The smoothing effect can be varied depending on the object.

  As another example of the method for calculating the normal direction angle and the tangential direction angle of the contour line, there is a method of calculating based on the change direction of the value (brightness) of each pixel on the contour line. In this case, the direction in which the luminance value changes is calculated, the changing direction is the normal direction, and the perpendicular direction is the tangential direction. In the case of this method, the brightness change direction may be extracted using, for example, a Sobel differential filter. According to this method, it is an effective method for matching a simple shape with few uneven shapes, for example, a circle, an ellipse, or a heart shape.

  When this Sobel differential filter is used, the Sobel filters Fx and Fy having a mask size of 3 × 3 are used as shown in FIG. And the differential value dx in the x direction and the differential value dy in the y direction at each pixel are obtained by the following equations. However, a to i represent pixel values (density values) in the pixel of interest and its eight neighboring pixels as shown in FIG.

dx = (a + 2d + g) − (c + 2f + i)
dy = (a + 2b + c) − (g + 2h + i)
Normal direction angle = atan (dy / dx)
In calculating the amount of difference between the associated pixels in the association between the numerical value sequence A and the numerical value sequence B, for example, the contour line tangential direction angle of the first pixel of the numerical value sequence A is θ1, the second pixel of the numerical value sequence B As shown in FIG. 5, the amount of difference between the pixels is the angle θ in the smaller direction of the angles θ1 and θ2 formed by the two vectors A1 and B2, as shown in FIG. It is a number. The calculated difference amount becomes the value of the intersection of A1 and B2 in the matrix of FIG. Also, depending on the numerical value, a differential angle with a plurality of expressions can be expressed uniquely, so that matching evaluation can be performed correctly.

For example,
If A1 = 100 degrees, B2 = 20 degrees, A1-B2 = 80 degrees A1 = 20 degrees, if B2 = 100 degrees, A1-B2 = -80 degrees → 280 degrees → 80 degrees A1 = 20 degrees, B2 = 250 degrees Then, A1−B2 = −230 degrees → 130 degrees When the numerical sequence by angle is generated by the numerical sequence generation means 41 as described above, the difference value calculation means 42 is stored in the template storage unit 5. A process of associating the numerical value string A and the numerical value string B for each pixel (element) is performed, a difference amount of numerical values of the associated pixels is obtained, and a process of obtaining a total sum of the difference values is performed. This sum indicates the degree of similarity between the region extracted image and the template image.

  The association between the numerical sequences A and B described above determines the correspondence with the partner while expanding and contracting the numerical sequence, and the correspondence that minimizes the sum of the difference amounts is set to the most matched state.

  For example, when the image of the region extracted from the input image with respect to the template image has the same size and a similar (same) shape, the correspondence between the pixels of the numerical sequences A and B is shown in FIG. As shown in FIG. Then, the difference between the numerical values of the two pixels is calculated and the sum of the differences is obtained. In this case, when the matching target image and the template image have the same shape, the difference amount is 0, so the sum is 0.

  On the other hand, when the matching target image extracted from the input image is a similar shape having an area ratio and a dimensional ratio of twice that of the template image (or vice versa), as shown in FIG. Each pixel of A corresponds to two pixels of the numerical sequence B. In this case, the numerical difference between the associated pixels is 0, and the total sum is 0. That is, the difference in size in the case of similar shapes can be absorbed.

  Further, when the matching target image and the template image have different shapes, the pixels corresponding to the same shape portion can be made to correspond one-on-one as shown in FIG. However, the amount of difference does not become zero in correspondence between pixels corresponding to different shape portions (indicated by bold lines). For this reason, the sum of the differences is not zero.

  Here, in the present embodiment, when the sum of differences (accumulated addition values) is obtained by accumulating the difference amounts between the associated pixels in the association of the numerical sequences A and B, the numerical sequences A and B are supported. The processing time is enormous in the method that assumes the correspondence in the form of brute force pixels, finds the sum of the differences in each correspondence, and uses the smallest sum among the obtained sums. .

  In view of this, a method for shortening the processing time and obtaining the minimum sum is adopted in this embodiment, and this method will be described below.

  First, when a matrix of the difference between the numerical value sequence A shown in FIG. 8A and the numerical value sequence B shown in FIG. 8B is created, it becomes as shown in FIG. 8C, and the right end of the bottom row of this matrix. In the two numerical value sequences A and B, all the pixels on the numerical value sequence B side are numerical values. Corresponds to any pixel on the column A side, all pixels on the numeric column A side correspond to any element on the numeric column B side, and all the correspondences are associated so that the order of the order is not reversed. It shows one correspondence among a plurality of things, and is equivalent to the cumulative addition of the difference values of the numerical values of the pixels having the correspondence.

  Here, as shown in FIG. 9B, when the template image α and the matching target image β are the same, the correspondence between the pixels of the numerical sequences A and B is only one-to-one as described above. Thus, the difference matrix is as shown in FIG. Then, the matrix image γ shown as the luminance value according to the difference amount between the corresponding pixels is as shown in FIG. 9B, and the difference amount between the corresponding pixels (as shown in FIG. 9A) ( In this case, a line connecting pixels indicated by luminance values corresponding to 0) is a diagonal line I of the matrix image γ.

  On the other hand, when the template image α and the matching processing image β are different as shown in FIG. 10, for example, in human detection, the template image α does not have both hands, but the region extraction image β has both hands. When the numerical sequences A and B are associated, a plurality of pixels are associated with one pixel, and there is an association in which the difference amount is not 0, and the sum of the difference amounts is not 0. Shows a matrix image γ, and a line I connecting the matrix images corresponding to the difference amounts of the associated pixels cannot be associated with a difference amount of 0, and the luminance value is not 0. The part passing through (vertical line part) II will appear.

  Therefore, a path that minimizes the sum of the difference amounts (cumulative addition value) is obtained using a matrix image by the following method.

  Here, in order to simplify the description, a case where the number of pixels (elements) in the numerical sequences A and B is four will be described. First, the difference amount calculation means 42 creates 4 × 4 matrix image data having pixel values based on the difference amounts between the respective pixels from the numerical sequences A and B, and applies each pixel from the upper left to the lower right of the matrix image data. Pixel numbers are assigned sequentially. FIG. 11A shows the contents of this matrix image data.

  The shortest path from the lower right pixel (16) of the matrix image data to each pixel in the lowermost stage is determined. FIG. 11B shows the contents at the time of determination, and in the figure, the path cost indicates the cumulative sum of pixel values on the path. The shortest path to the lower right pixel (16) in each pixel in the lowermost stage is only in the right direction.

  Next, as shown in FIG. 11C, each of the pixel values of the shortest path in the pixels (12), (16), and (15) located on the right, lower right, and lower from a target pixel (for example, 11), respectively. Cumulative <5 (= 4 + 1) on the right, 4 on the lower right, 13 (= 4 + 9) on the lower>, determines the shortest path from the lower right pixel (16) to the target pixel (11), and calculates the sum Ask. That is, the shortest path is at the lower right, and the cumulative sum of pixel values in that case is 4 + 6 = 9 (see FIG. 11D).

  In the next pixel of interest (10), the cumulative addition values 9, 13, and 15 of the pixel values of the shortest path in the pixels (11), (15), and (14) positioned on the right, lower right, and lower are compared. Then, the shortest path from the lower right pixel (16) to the target pixel (10) is determined, and the cumulative sum thereof is obtained. In this case, the shortest path is on the right, and the cumulative sum of pixel values in that case is 9 + 10 = 19 (see FIG. 11E). When the shortest path is sequentially determined as described above and the shortest path to the upper left pixel (1) is determined as shown in FIG. 11F, the pixels from the upper left pixel (1) to the lower right pixel at that time. That is, the shortest path to (16) and the minimum cumulative sum of the pixel values are determined. Here, since the pixel value corresponds to the amount of difference between the corresponding pixels in the numerical sequences A and B, the accumulation of the pixel value corresponds to the sum of the amounts of difference.

  Since it is assumed that the path does not follow backwards (upper, left, upper left), if the pixel of interest is at the lowest level, the lower right and lower pixels that compare the minimum cumulative sum do not exist and are compared. Is only the right pixel. Further, the pixel at the right end of the lowermost row does not exist in other pixels to be compared. In the case of the rightmost pixel of interest, although not the lowest stage, since there is no pixel on the right, the minimum cumulative sum is compared with the lower right and lower pixels.

  Thus, according to this method, if the minimum cumulative sum to the target pixel (cell) and its path are determined, it is not necessary to repeat the calculation process for the minimum cumulative sum and path of the next target pixel, and the processing time Will be shortened.

  When the minimum cumulative sum of a certain pixel of interest (a grid) exceeds a threshold used for the determination by the matching determination unit 6, the cumulative sum of the pixel of interest is referred to in other pixels of interest. If excluded from the comparison target, the calculation time can be shortened.

 If there is no right or lower right or lower cell to be compared if it is at the bottom or right end, it is compared otherwise.

  Note that the movement of the cells for obtaining the minimum cumulative sum is not only the movement from the rightmost cell of the bottom row to the left end of the top row, but also the movement from the leftmost cell of the top row to the right end of the bottom row, etc. Also good.

  Thus, the difference amount calculating means 42 corresponds to the numerical value sequence A corresponding to the pixels on the contour line of the template image, all the pixels corresponding to the numerical value sequence A on the template image side, and the numerical value sequence B on the matching target image side. Assume that all the pixels are associated with each other so that the arrangement order of the front and rear is not reversed, the difference between numerical values of the associated pixels is obtained, and the sum of the minimum differences is the highest in similarity The result is output to the matching judgment unit 6.

  The matching determination unit 6 is preset with a threshold value for determining that the template image and the matching target image are the same, and compares this threshold value with the sum total output by the difference amount calculation means 42 to determine the sum total. Is smaller than the threshold value, it is determined that the template image and the matching target image are the same. Conversely, when the sum is larger than the threshold value, it is determined that they are not the same, and the determination result is sent to the result output unit 7. Output. The result output unit 7 outputs a result signal in a predetermined format to the outside according to the determination result.

  If there is a pixel of interest whose cumulative addition value exceeds the threshold when obtaining the shortest path described above, the arithmetic processing is shortened by excluding the cumulative addition value of the pixel of interest from the comparison target when obtaining the shortest path. You may do it.

  In the present embodiment, the tangential direction angle (or normal direction angle) of the contour line of each pixel is used as the numerical value constituting the numerical sequences A and B. However, the gray level image is used to represent the pixels on the contour line. The luminance value may be used as a numerical value, or in the case of a color image, the HLS value on the contour line may be used. In the former case, the human hair is detected by using the black color of the human hair, the skin color of the face or the hand. The latter is effective when a person image detected at a certain moment is used as a template image, and the person is identified and tracked from matching target images that are sequentially input based on the color of the person's clothes and the range of hair. It is valid. In the case of this tracking, after capturing an image to be a template image, the above-described region extraction, contour line extraction, and numerical sequence generation are performed before the start of the matching process, and the generated numerical sequence is stored in the template storage unit. 5 is stored.

  Furthermore, in the present embodiment, the matching determination unit 6 compares the threshold value with the sum of the difference amounts output from the difference amount calculation means 42 to determine whether or not the matching target image is the same as the template image. However, normalization of the sum of the differences may be performed as in the following equation (1) based on the number of associations (the number of lines connecting the pixels in FIG. 7) Y. In this case, it is possible to correctly determine the degree of similarity between images with extremely different sizes. Furthermore, even if the number of contour pixels corresponding to the template image N1 and the number of contour pixels corresponding to the matching target image N2 are normalized as shown in the following equation (2), There are significant advantages.

Normalized sum of differences = total difference / number of associations (1)
Normalized sum of differences = total sum of differences / (N1 × N2) (2)
Further, the distribution state of the difference amounts in all the difference amount calculation means 42 is checked, and based on the number of associations where the difference amounts are within a predetermined range, whether the matching target image and the template image are the same or not. You may use the method of judging. For example, it is determined whether or not the ratio of the correspondence relationship in which the difference amount is within a predetermined range (the difference amount is 0 or more and 5 or less) with respect to the number of associations Y is equal to a preset value. .

(A) is a block diagram of an image processing apparatus using the method of the present invention, and (b) is a block diagram of a matching processing unit of the image processing apparatus of the above. It is explanatory drawing of the outline tracking method of the outline extraction part in the matching process part of an image processing apparatus same as the above. It is explanatory drawing of the calculation method of the tangent direction angle of the numerical value sequence generation part in the matching process part of an image processing apparatus same as the above. It is explanatory drawing of the processing operation with respect to the specific example of the matching process part of an image processing apparatus same as the above. It is explanatory drawing of another angle calculation method of the numerical value sequence generation part of the matching process part of an image processing apparatus same as the above. It is explanatory drawing of the angle calculated | required as a numerical value of a numerical sequence in the numerical sequence generation part of the matching process part of an image processing apparatus same as the above. It is explanatory drawing of the difference amount calculating part of the matching process part of an image processing apparatus same as the above. It is explanatory drawing of the calculation method of the sum total of the difference amount calculating part of the matching process part of an image processing apparatus same as the above. It is explanatory drawing of an example of the difference matrix produced | generated by the difference amount calculating part of the matching process part of an image processing apparatus same as the above, and a matrix image. It is explanatory drawing of the other example of the difference matrix produced | generated by the difference amount calculating part of the matching process part of an image processing apparatus same as the above, and a matrix image. It is explanatory drawing of the method of calculating | requiring the shortest path | route where the cumulative addition value of the image processing apparatus same as the above becomes the minimum. It is explanatory drawing of the method of obtaining a matching object image from the input image by the difference with a background image. It is explanatory drawing of the conventional template matching. It is a subject explanatory drawing of the conventional template matching.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Imaging camera 2 Image input part 3 Area extraction part 4 Template processing part 40 Outline line extraction means 41 Numerical value sequence generation means 42 Difference amount calculation means 5 Template storage part 6 Matching judgment part 7 Result output part

Claims (10)

When performing the object matching process using a one-dimensional numerical sequence consisting of numerical values based on the specification values of each pixel on the contour line of the template image and the matching target image, the pixels corresponding to the numerical sequence on the template image side A process for associating all of the pixels corresponding to the numerical sequence on the matching target image side so that the order of arrangement before and after is not reversed, calculating a numerical difference amount between the associated pixels, and calculating the difference amount A template matching method characterized in that a process for obtaining a sum is performed, and a correspondence with the smallest sum obtained is determined as a correspondence in which a template image and a matching target image are most matched. The template matching according to claim 1, wherein if the sum of the differences determined to be the most matched correspondence is smaller than a preset threshold value, the matching target image and the template image are determined to be the same. Method. By examining the distribution of the difference amount between each associated pixel when it is determined that the association with the smallest sum obtained is the association in which the template image and the matching target image are the most matched, the difference amount is within a predetermined range. The template matching method according to claim 1, wherein it is determined whether the matching target image and the template image are the same based on the number of associations in the template. 4. The numerical value sequence according to claim 1, wherein the numerical value sequence is a sequence in which tangential direction angles or normal direction angles of the contour line calculated based on a brightness change direction indicated by a pixel are arranged. The template matching method described. Each numerical example described above is a column in which the tangential direction angle or normal direction angle of the contour line is arranged, and using the pixel positions before and after a certain target pixel on the contour line, 4. The template matching method according to claim 1, wherein a tangential direction angle of a contour line of the target pixel is set. 6. The template according to claim 1, wherein the difference amount is a difference value between a tangential direction angle and a normal direction angle, and is a smaller one of angles formed by two vectors. Matching method. 4. The luminance value of a pixel corresponding to the contour line position in the case of a grayscale image, and the HLS value of a pixel corresponding to the contour position in the case of a color image are used as the numerical values in the numerical sequence. The template matching method according to any one of the above. A matrix is created based on the amount of difference between the numerical sequence on the template image side and the numerical sequence on the matching target image side, and the cells that are noticed in order from the rightmost cell to the leftmost cell in each stage of this matrix are the first. It has a process of determining the route that follows the cell where the accumulated value of the value of the cell is the smallest while moving sequentially from the lower level to the uppermost level. The template matching method according to any one of claims 1 to 3, wherein the route is determined by comparing cumulative addition values of the cells existing on the right, lower right, and lower with respect to the target cell. 9. The template according to claim 8, wherein, when the path of the target pixel is determined, when the cumulative addition value is larger than a threshold used for matching determination, the cumulative addition value is excluded from the cumulative addition value to be compared at the time of path determination. Matching method. Contour line extracting means for extracting the contour line of the matching target image, and numerical value string generating means for generating a one-dimensional numerical value sequence on the matching target image side from the specification values of each pixel on the contour line extracted by the contour line extracting means And all the pixels corresponding to the one-dimensional numerical sequence on the matching target image side generated by the numerical sequence generation means, and the specification values of each pixel on the contour line extracted from the template image stored in advance All the pixels corresponding to the one-dimensional numerical sequence on the template image side are associated with each other so that the order of arrangement in the front and rear is not reversed, and the amount of numerical difference between the associated pixels is calculated and the sum of the differences is calculated. An image processing apparatus comprising: a difference amount calculation unit to be calculated; and determining that the target image and the template image are the same based on a sum of the difference amounts calculated by the difference amount calculation unit