JP4743277B2 - Image data recognition method, image processing apparatus, and image data recognition program - Google Patents

Description

  The present invention relates to image data recognition. In particular, the present invention relates to an image data recognition method for detecting a subject in image data.

  In a system related to traffic, there is a technique for detecting a vehicle or a pedestrian from image data acquired by a fixed camera and image data acquired by a camera mounted on a moving body. A technique for detecting a vehicle or a pedestrian is used, for example, to acquire image data of an intersection, detect a vehicle in the intersection, and present the detected vehicle or pedestrian to a traveling vehicle. . By presenting the presence of the vehicle or the pedestrian to the traveling vehicle, the driver of the vehicle can know the situation in the intersection in advance.

  As a method for extracting a vehicle to be detected in image data, there are a background difference method, an edge detection method, and the like. The background difference method is a method of extracting a target object in image data by acquiring in advance background image data without the target object and comparing it with image data to be extracted. The edge detection method is a method of detecting an edge in image data, comparing template data based on edge data stored in advance with detected edge data, and specifying a target vehicle.

Since the background difference method detects an area including a shadow of a vehicle and the like, there is a problem that the position of only the vehicle in the image data cannot be accurately detected (Patent Document 1). The edge detection method also detects shadows. In addition, since the object is identified by comparing the shape data of the edge-detected portion and the shape data stored in the object shape database in advance, there is a problem that the detection process takes time (Patent Document) 2).
JP 2005-190142 A JP 2000-148777 A

(Problems to be solved by the invention)
It is an object of the present invention to provide an image data recognition method that detects a subject in image data without being affected by luminance and having a small amount of calculation.
(Means for solving the problem)
The present invention has the following configuration in order to provide a method for detecting a subject in image data that solves the above problems. First, the first pixel, the first region centered on the first pixel, the plurality of second pixels to be compared with the first pixel, and the plurality of second pixels, respectively. A plurality of second regions to be defined. Next, the similarity value calculated from the first region and each second region is associated with each second pixel. Next, a similarity value indicating a feature of the first region among the similarity values associated with the plurality of second pixels is associated with the first pixel. Next, an area where the shape determined according to the distribution state of the similarity value indicating the characteristics of the first area matches the shape previously determined as the characteristics of the object is detected as a portion of the object in the image data.

  In addition, a region in front of the three-dimensional space, which is the imaging target region of the image data, is large, and a region behind the three-dimensional space is reduced. With this configuration, it is possible to reduce the amount of calculation.

  Further, the shape defined as the feature of the subject is a condition that determines the shape of the vehicle.

  In addition, the coordinates at which the large similarity values are distributed on both sides of the detection range, the similarity values on both sides are symmetric, and the small similarity values are distributed below the detection range are displayed at the bottom of the front of the vehicle. It is characterized by the coordinates.

Further, the vehicle obtains the center position of the vehicle in the horizontal direction in the image data from the objectivity of the distribution of similar values of the vehicle, and specifies the position in the vertical axis direction from the conditions for determining the shape of the vehicle distribution and the shape of the vehicle. It is characterized by that.
(The invention's effect)
According to the present invention, for each pixel in the image data, a similarity value for each region centered on the pixel is detected, and a region in the image data whose similarity value distribution shape matches a predetermined condition is detected. . Therefore, it is possible to detect a subject in the image data without being influenced by luminance and with a small amount of calculation.

FIG. 1 is a configuration diagram of an image processing apparatus 1 according to the present embodiment. FIG. 2 shows image data taken by the camera 2 of the present embodiment. FIG. 3 is a configuration example of the database 20 in which the vertical coordinate 13 of the vehicle 16 existing in the image data 10 and the width of the vehicle 16 are associated with each type of the vehicle 16. FIG. 4 is a conceptual diagram for detecting the similarity value 38 from the acquired image data 10. FIG. 5 is a flowchart of a method for calculating the similarity value 38. FIG. 6 is an example of a pedestrian crossing 41. FIG. 7 is a flowchart of processing for excluding the pedestrian crossing 41 area from the image data 10 executed by the road mark deleting unit 5. FIG. 8 is a diagram showing an image (A) of the vehicle 16 portion of the original image data 10 and an image (B) of the vehicle 16 portion of the distribution map 50. FIG. 9 is a flowchart of processing for detecting an area where the vehicle 16 exists from the distribution chart 10 executed by the pattern feature quantity calculation unit 6. FIG. 10 is a diagram showing the relationship between the distribution map 50 and the object of the vehicle 16. FIG. 11 is a flowchart of a process for detecting a region where the vehicle 16 exists from the distribution chart 10 executed by the symmetric feature quantity calculation unit 7. FIG. 12 is an example of vehicle detection according to this embodiment. FIG. 13 is a hardware configuration example of the image processing apparatus 1 of the present embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Image processing apparatus 2 Camera 3 Image memory 4 Auto correlation difference calculation part 5 Road mark deletion part 6 Pattern feature-value calculation part 7 Objectivity feature-value calculation part 8 Vehicle position determination part 10 Image data 11 Pixel 12 Vanishing point 13 Vertical direction 14 Lateral direction 15 Roadway 16 Vehicle 17 Point 20 indicating the boundary between the roadway and the roadside belt in the foreground Database 22 Light vehicle 23 Normal vehicle 24 Large vehicle 31 Search range 32 Calculation area 33 Calculation target pixel 35 Comparison target pixel 34 Comparison area 38 Similar Value 39 Difference 41 Crosswalk 42 Striped pattern 43 Coordinate points 44 subject to calculation Coordinate 45 moved in the horizontal direction 14 by a predetermined amount L Coordinate 46 moved in the horizontal direction 14 by a predetermined amount (-L / 2) Predetermined amount ( The coordinate 47 moved in the lateral direction 14 by L / 2) Edge 48 along the line segment facing the vanishing point 12 Horizontal edge 50 Distribution map 51 Vehicle 16 The pixel group 52 located on the left and right sides of the vehicle group The pixel group 53 located on the outside of the left and right side surfaces of the vehicle 16 The pixel group 54 located on the front side of the vehicle 16 Pixel 55 The detection target pixel 56 Eleft
57 Eight
58 Eside
59 Ebottom
60 Value indicating size of vehicle feature 61 Calculation central coordinate 101 CPU
103 Memory 104 Output unit 105 Bus

  In this embodiment, an operation for acquiring a vehicle from image data of a fixed camera at an intersection will be described.

  FIG. 1 is a configuration diagram of an image processing apparatus 1 according to the present embodiment. The image processing apparatus 1 according to the present invention includes a camera 2, an image memory 3, an autocorrelation difference calculation unit 4, a road mark deletion unit 5, a pattern feature value calculation unit 6, an object feature value calculation unit 7, and a vehicle position determination unit 8. Composed.

  The image data of the intersection imaged by the camera 2 is stored in the image memory 3. The autocorrelation difference calculation unit 4 creates a distribution map of similar value differences from the image data stored in the image memory 3. The autocorrelation difference calculation unit 4 obtains a difference between similar values for each pixel and creates a distribution map based on the difference between similar values. The road mark deletion unit 5 deletes a road mark area that appears in the image data stored in the image memory 3. As described above, the image processing apparatus 1 calculates a difference between similar values from the original image data, and obtains a distribution map from which road marks are removed. Next, the pattern feature quantity calculation unit 6 detects the vehicle pattern from the difference value distribution map created by the autocorrelation difference calculation unit 4. The pattern feature quantity calculation unit 6 obtains vertical coordinates in the image data indicating the front or rear part of the vehicle from the vehicle pattern. When the pattern feature quantity calculation unit 6 detects the vehicle pattern, the pattern feature amount calculation unit 6 also takes into account the area information deleted by the road mark deletion unit 5. The target characteristic quantity calculation unit 7 calculates the target characteristic of the vehicle from the distribution map of the difference between the similar values created by the autocorrelation difference calculation unit 4. The target characteristic amount calculation unit 7 acquires the center line where the vehicle in the image data exists from the target. The vehicle position determination unit 8 specifies the position of the vehicle in the horizontal direction in the image specified by the pattern feature quantity calculation unit 6 and the target feature quantity calculation unit 7. Further, the vehicle position determination unit 8 acquires a vehicle width determined from the position of the vehicle, and specifies a vehicle type by searching a database from the acquired vehicle width and a vertical position in the image data.

  Next, image data acquired by the image processing apparatus 1 with the camera 2 will be described. FIG. 2 shows image data 10 taken by the camera 2 of this embodiment. The image data 10 is stored in the image memory 3 in a digital format. The image data 10 is composed of a set of pixels 11. The pixel 11 is a minimum unit constituting the image data 10. In the following description, the vertical direction of the image data 10 indicates the direction of the vertical direction 13, and the horizontal direction indicates the direction of the horizontal direction 14. Further, the ordinate indicates coordinates that specify the position in the vertical direction 13 in the image data 10, and the abscissa indicates coordinates that specify the position in the horizontal direction 14 in the image data 10. The vanishing point 12 is a coordinate at the back when the image data 10 projected in two dimensions is viewed in the original three-dimensional space.

The vanishing point 12 of the image data 10 in the case where the camera 2 is installed at a place looking down from above is generally arranged at a position above the vertical direction 13 in the image data 10. Arranging the vanishing point 12 at an upper position in the image data 10 makes it possible to increase the area of the roadway 15 in the image data 10. Therefore, the image processing apparatus 1 can decompose the longitudinal direction 13 of the roadway 15 with a larger number of ordinates than when the vanishing point 12 is located at a lower position in the image data 10. If it becomes possible to disassemble more in the vertical direction 13, the image processing apparatus 1 can accurately acquire the position of the vehicle 16 on the roadway 15. On the other hand, when the camera 2 is mounted on the vehicle 16, the camera 2 is generally arranged so as to capture a horizontal direction with respect to the ground. When the camera 2 is attached to the vehicle 16 , the camera 2 is attached to a low position such as a front portion of the vehicle 16 . Therefore, when the camera 2 is at an angle directed downward, the camera 2 cannot capture the highest part of the vehicle 16 traveling forward.

The width of the roadway 15 is set to be the width of the roadway 15 to be acquired at a position below the vertical direction 13 in the image data 10. At this time, the point 17 indicating the boundary between the roadway 15 and the roadside belt in the foreground is the lowermost part of the image data 10. By disposing the roadway 15 so that the width of the roadway 15 is maximized at a lower position in the image data 10, the camera 2 can capture a large width of the vehicle 16 in front of the image data 10. By enlarging the width of the vehicle 16, the image processing apparatus 1 can be decomposed with more pixels than the width of the vehicle 16, and the image processing apparatus 1 can accurately acquire the width information of the vehicle 16. In addition, the angle of the camera 2 about the vanishing point 12 and the width of the roadway 15 can be changed according to road conditions and an object to be acquired.

  Next, data for detecting the vehicle 16 included in the image processing apparatus 1 will be described. The image processing apparatus 1 has coordinate information of the vanishing point 12 in the image data 10. Further, the image processing apparatus 1 has the coordinates of a point 17 indicating the boundary between the roadway 15 and the roadside belt in the foreground. The coordinates of the point 17 indicating the boundary between the roadway 15 and the roadside belt are two points on both sides of the road. Further, the image processing apparatus 1 obtains a straight line at the boundary between the roadway 15 and the roadside band from a straight line connecting the coordinates of the vanishing point 12 and the coordinates of the point 17 indicating the boundary between the roadway 15 and the roadside band. The image processing apparatus 1 also includes a database 20 that associates the coordinates of the vehicle 16 in the vertical direction 13 and the width of the vehicle 16 existing in the image data 10 for each type of the vehicle 16.

  FIG. 3 is a configuration example of the database 20 in which the coordinate value 21 in the vertical direction 13 of the vehicle 16 existing in the image data 10 and the width of the vehicle 16 are associated with each type of the vehicle 16. For example, the image data 10 in which the coordinate value 21 in the vertical direction 13 is composed of 480 pixels 11 has the width information of the vehicle 16 corresponding to the 480 coordinate values 21 in the vertical direction 13. The width of the vehicle 16 is substantially determined according to the light vehicle 22, the ordinary vehicle 23, and the large vehicle 24. For example, a typical vehicle width of the ordinary vehicle 23 is 1700 mm. The image processing apparatus 1 can obtain in advance the number of pixels in the width of the vehicle 16 in the horizontal direction 14 corresponding to the position of the coordinate value 21 in the vertical direction 13 of the vehicle 16 in the image data 10. The database 20 of the present embodiment has the width of the vehicle 16 in the lateral direction 14 such as the light vehicle 22, the ordinary vehicle 23, and the large vehicle 24 according to the position of the coordinate value 21 in the longitudinal direction 13 as the width information of the vehicle 16. The number of pixels is related and stored.

  Referring to FIG. 2, the length of the base of the quadrilateral indicating the vehicle 16 in the roadway 15 in the image data 10 is the number of pixels that is the width of the vehicle 16. The number of pixels of the width of the vehicle 16 changes in accordance with the coordinates in the vertical direction 13 where the quadrangle is detected.

  Next, a process of creating a difference map of similar values performed by the autocorrelation difference calculation unit 4 of the image processing apparatus 1 will be described. The autocorrelation difference calculation unit 4 calculates a difference 39 of similarity values 38 associated with each pixel 11 of the image data 10 stored in the image memory 3. The similarity value 38 is obtained by quantifying the degree of approximation of the entire area when the first area and the second area set in the image data 10 are compared.

  FIG. 4 is a conceptual diagram in which the autocorrelation calculation unit 4 detects the similarity value 38 from the acquired image data 10.

  The calculation target pixel 33 (first pixel) is a pixel 11 for obtaining the similarity value 38 in the image data 10 that is the current target of the calculation. The calculation target pixel 33 is for all the pixels 11 in the image data 10. It should be noted that the calculation of the similarity value 38 can be omitted for the pixels 11 located in the region where the vehicle 16 cannot exist in the image data 10. For example, the autocorrelation difference calculation unit 4 can omit the calculation for the pixel 11 that exceeds the roadside band and is in a range in which the vehicle 16 is clearly not present. An area in which the autocorrelation difference calculation unit 4 can omit the calculation is set in advance. The autocorrelation difference calculation unit 4 calculates the similarity value 38 for all the pixels 11 other than the region where the calculation can be omitted. Since the autocorrelation difference calculation unit 4 omits the calculation for the pixels in the range where the vehicle 16 is obviously not present, the amount of calculation performed by the image processing apparatus 1 is reduced. It is possible to shorten the time until obtaining.

  The comparison target pixel 35 (second pixel) is the pixel 11 that is the current target of the calculation for obtaining the similarity value 38 in the image data 10 for the calculation target pixel 33.

  The calculation area 32 (first area) is an area that is the current comparison target of the calculation for obtaining the similarity value 38 with the calculation target pixel 33 as the center. The comparison area 34 (second area) is an area to be compared with the calculation area 32 around the comparison target pixel 35.

  The search range 31 is a range in which the comparison target pixel 35 is set for the calculation target pixel 33. The search range 31 is set because it is not necessary to calculate the similarity value 38 when the distance between the coordinates in the image data 10 is large between the calculation target pixel 33 and the comparison target pixel 35. This is because the calculation area 32 and the comparison area 34 are considered not to target one vehicle 16 when the distance between the coordinates is greatly separated. On the other hand, the calculation of the similarity value 38 between completely different objects is a cause of erroneous detection of the vehicle 16 in the image data 10. For example, the search range 31 is a range in which the vehicle 16 can exist around the calculation target pixel 33.

  The comparison target pixels 35 are all the pixels 11 within the search range 31. When the search range 31 and the region where the vehicle 16 cannot exist on the image data 10 overlap, the autocorrelation difference calculation unit 4 calculates the similarity value 38 of the comparison target pixel 35 in the overlapping range. Can be omitted. If the calculation for calculating the similarity value 38 with the comparison target pixel 35 in the overlapping range is omitted, the amount of calculation performed by the image processing apparatus 1 is reduced, so that the image processing apparatus 1 has time to obtain the calculation result. It can be shortened.

  Next, a method of calculating the similarity value 38 performed by the autocorrelation difference calculation unit 4 will be described. FIG. 5 is a flowchart of a method for calculating the similarity value 38. The autocorrelation difference calculation unit 4 performs a calculation for calculating the similarity value 38 for each calculation target pixel 33 in the image data 10 according to the following procedure. The autocorrelation difference calculation unit 4 specifies the coordinates (x, y) of the calculation target pixel 33 (S01). The coordinates of the calculation target pixel 33 in FIG. 4 are (x, y). “X” is a coordinate value in the horizontal direction 14 of the calculation target pixel 33 in the image data 10. “Y” is the value of the coordinate in the vertical direction 13 of the calculation target pixel 33 in the image data 10. The area around the coordinates (x, y) of the calculation target pixel 33 is the calculation area 32. Further, each pixel 11 in the calculation area 32 is assumed to be (x + i, y + j). “I” is a variable that specifies the coordinate in the horizontal direction 14 of the pixel 11 in the area around the coordinate (x, y) of the calculation target pixel 33. “J” is a variable that specifies the coordinate in the vertical direction 13 of the pixel 11 in the region around the coordinate (x, y) of the calculation target pixel 33.

  The autocorrelation difference calculation unit 4 specifies the search range 31 (S02). The search range 31 is a value having a preset size, and is set around the coordinates (x, y) of the calculation target pixel 33.

  The autocorrelation difference calculation unit 4 specifies the coordinates (x + u, y + v) of the comparison target pixel 35 in the search range 31 (S03). “U” is a variable for specifying the coordinate in the horizontal direction 14 within the search range 31. “V” is a variable for specifying the coordinate in the vertical direction 13 within the search range 31. A region around the comparison target pixel 35 is a comparison region 34. Further, each pixel in the comparison area 34 is assumed to be (x + u + i, y + v + j). “I” is a variable for specifying the coordinate in the horizontal direction 14 of the pixel 11 in the region around the comparison target pixel (x + u, y + v). “J” is a variable for specifying the coordinate in the vertical direction 13 of the pixel 11 in the area around the comparison target pixel (x + u, y + v).

  Next, the autocorrelation difference calculation unit 4 calculates a similarity value 38 between the calculation area 32 and the comparison area 34. The calculation for calculating the similarity value 38 is performed by “i” and “j” of (x + i, y + j) indicating the coordinates of each pixel 11 in the area around the calculation target pixel 33 and in the area around the comparison target pixel 35. This is done by changing the values of “i” and “j” of (x + u + i, y + v + j) indicating the coordinates of each pixel 11 and comparing each pixel corresponding to each coordinate (S04).

  For example, the autocorrelation difference calculation unit 4 calculates the similarity value 38 between the calculation area 32 and the comparison area 34 by the following method. 1. 1. Method SAD (Sum Of Absolute Difference) for calculating from the sum of absolute values of differences between the luminance of the pixels in the calculation area 32 and the luminance of the pixels in the comparison area 34; 2. a method SSD (Sum Of Square Difference) for calculating from the sum of the square of the luminance of the pixel in the calculation area 32 and the luminance of the pixel in the comparison area 34, and 3. The calculation is performed by normalizing correlation (NCOR). By the above method, the autocorrelation difference calculation unit 4 calculates the luminance difference between the pixels between the calculation region 32 and the comparison region 34, and calculates the similarity 38 as the entire comparison region 34 from the sum of the luminance differences. .

  The autocorrelation difference calculation unit 4 performs the following calculation when calculating the similarity value 38 by SAD. The brightness P1 (x + i, y + j) of the pixel 11 at each coordinate (x + i, y + j) in the calculation area 32 is acquired. Next, the brightness P2 (x + u + i, y + v + j) of the pixel 11 at the coordinates (x + u + i, y + v + j) in the comparison area 34 is acquired. Next, the absolute value of the difference in luminance is calculated between the luminance P1 (x + i, y + j) and the luminance P2 (x + u + i, y + v + j). The sum of absolute values of the differences in luminance calculated between all the pixels 11 in the calculation area 32 and all the pixels 11 in the comparison area 34 is calculated. The autocorrelation difference calculation unit 4 sets the calculated value as a similarity value 38 corresponding to the comparison target pixel 35. An arithmetic expression for calculating the sum of absolute values of luminance differences is Expression (1). Note that “t” indicates that the currently acquired image data 10 is a target.

  The autocorrelation difference calculation unit 4 performs the following calculation when calculating the similarity value 38 by SSD.

  The autocorrelation difference calculation unit 4 calculates the square of the luminance difference between the luminance P1 (x + i, y + j) and the luminance P2 (x + u + i, y + v + j). The autocorrelation difference calculation unit 4 calculates the sum of squares of the differences in luminance calculated between all the pixels in the calculation region 32 and all the pixels in the comparison region 34. The autocorrelation difference calculation unit 4 sets the calculated value as a similarity value 38 corresponding to the comparison target pixel 35. An arithmetic expression for calculating the sum of the squares of the luminance differences is Expression (2). Note that “t” indicates that the currently acquired image data 10 is a target.

  The autocorrelation difference calculation unit 4 performs the calculation of Expression (3) when calculating the similarity value 38 by NCOR. “T” indicates that the image is currently acquired. Note that “t” indicates that the currently acquired image data 10 is a target.

  The autocorrelation difference calculation unit 4 repeats the process until the luminance difference is calculated for all the pixels between the comparison area 34 and the calculation area 32 (S05: no). When the autocorrelation difference calculation unit 4 calculates the sum of the luminance differences (S05: yes), it is associated with the comparison target pixel 35 as a similarity value 38 (similar value calculated from the first region and the second region). Store (S06).

  The autocorrelation difference calculation unit 4 repeats the process until the similarity value 38 between all the comparison target pixels 35 and the calculation target pixel 33 in the search range 31 is calculated (S07: no). When the autocorrelation difference calculation unit 4 calculates the similarity value 38 for all the comparison target pixels 35 in the search range 31 (S07: yes), each comparison target pixel 35 in the search range 31 is included in the calculation target pixel 33. The difference 39 of the similarity value 38 indicating the feature (the similarity value indicating the feature of the first area) is associated with the difference 39 of the similarity value 38 that is included (S08). The difference 39 of the similarity value 38 indicating the feature is a large value when the image of the calculation area 32 includes a corner, and is a small value when the image of the calculation area 32 has little change such as a straight line or an asphalt surface. Become. The difference 39 of the similarity values 38 associated in S08 is used for a vehicle detection process described later.

  Here, a method of detecting the difference 39 of the similarity values 38 that the autocorrelation difference calculation unit 4 associates with the calculation target pixel 33 in S08 will be described. The autocorrelation difference calculation unit 4 calculates the difference 39 of the similarity values 38 associated with the calculation target pixel 33 in the following procedure.

  The comparison target pixel 35 calculated by the autocorrelation difference calculation unit 4 within the search range 31 includes the calculation target pixel 33 itself. The similarity value 38 calculated between the calculation target pixels 33 itself is a value indicating the most similarity among the similar values 38 corresponding to the comparison target pixels 35 in the search range 31. Therefore, it is necessary to exclude the similarity value 38 calculated between the calculation target pixel 33 itself.

  When calculated by SAD and SSD, the calculation target pixel 33 itself is the same, and the similarity value 38 indicating the most similarity is “0”. When the calculation target pixel 33 itself is included in the comparison target pixel 35 when calculated by SAD or SSD, the autocorrelation difference calculation unit 4 has the second highest similarity value 38 in the search range 31. Is detected. For example, the autocorrelation difference calculation unit 4 performs a process of removing the similarity value 38 between the calculation target pixel 33 itself, and among the similar values 38 stored corresponding to each comparison target pixel 35 in the search range 31. The similarity value 38 that is the minimum value is acquired. The autocorrelation difference calculation unit 4 sets the acquired value as the difference 39 of the similarity value 38 that is associated with the calculation target pixel 33. The calculation obtained from SAD is shown in Equation (4). The calculation obtained from the SSD is shown in Expression (5). Note that “t” indicates that the currently acquired image data 10 is a target.

  In the case of NCOR, the calculation target pixel 33 itself is the same, and the similarity value 38 is “1”. In the case of NCOR and when the calculation target pixel 33 itself is included in the comparison target pixel 35, the autocorrelation difference calculation unit 4 detects the comparison target pixel 35 having the second-like similarity value 38 in the search range 31. For example, the similar value 38 that is the maximum value among the similar values 38 stored corresponding to each comparison target pixel 35 in the search range 31 is obtained by performing the process of removing the similar value 38 between the calculation target pixel 33 itself. To get. Thereafter, the autocorrelation difference calculation unit 4 sets a value 39 obtained by subtracting the second-like similarity value 38 from “1” as the difference 39 of the similarity value 38 that is associated with the calculation target pixel 33. The calculation obtained from NCOR is shown in Equation (6). Note that “t” indicates that the currently acquired image data 10 is a target.

  Next, the feature of the difference 39 of the second-like similarity value 38 will be described. The difference 39 of the similarity values 38 calculated by the autocorrelation difference calculation unit 4 has a feature that increases when the calculation area 32 has a corner. The vehicle 16 has many corners when viewed from the front. For example, a frame portion on both sides of a windshield attached to the vehicle 16, a fender mirror portion of the vehicle 16, a front grill portion on the front of the vehicle 16, a license plate portion on the front of the vehicle 16, and the like. These parts of the vehicle 16 are shown as corners in the image data 10, and the calculated similarity value 38 is a larger value than the similarity value 38 of a region where other corners are shown. The reason why the similarity value 38 is large is that the calculation area 32 and the comparison area 34 have few corners having the same shape, and that the number of pixels having different luminance between the calculation area 32 and the comparison area 34 is large. .

  On the other hand, when the calculation area 32 has no corner, there is a feature that the difference 39 of the similarity value 38 calculated by the autocorrelation difference calculation unit 4 becomes small. The case where there is no corner is, for example, a case where a straight edge portion such as a roof portion of the vehicle 16 is used as a region, or a case where the inside of a region such as asphalt is configured with similar colors.

  The autocorrelation difference calculation unit 4 performs a calculation process of acquiring the difference 39 of the similarity value 38 with respect to the calculation target pixel 33 according to the above procedure as the calculation target pixel 33 for all the pixels 11 in the image data 10 (S09). When the difference 39 of the similarity value 38 is not associated with all the calculation target pixels 33 in the image data 10 (S09: no), the autocorrelation difference calculation unit 4 sets the calculation target pixel 33 to the next in the image data 10. Similarly, the difference 39 between the similar values 38 is calculated. On the other hand, when the difference 39 of the similarity value 38 is associated with all the calculation target pixels 33 in the image data 10 (S09: yes), the autocorrelation difference calculation unit 4 sets each calculation target pixel 33 in the image data 10. A distribution map 50 based on the difference 39 of the similarity value 38 associated with is created (S10).

  Here, the road mark in the image data 10 will be described. The distribution map 50 is constituted by the difference 39 of the similarity value 38. The pixels 11 having a large difference 39 of the similarity values 38 in the distribution map 50 are distributed in a large area in the image data 10 where the corner is reflected. However, the difference 39 of the similarity value 38 has a large value even in an area where the corner of the road mark is shown. Therefore, such a road mark causes erroneous detection. Therefore, the excess portion of the background is deleted from the image data 10 stored in the image memory 3. The extra portion of the background is, for example, a road mark. A road mark is a line indicating a boundary between a pedestrian crossing, a center line of a road, and a sidewalk. In this embodiment, processing for deleting information related to a pedestrian crossing in the image data 10 is performed. In this embodiment, it is assumed that a pedestrian crossing is drawn in parallel with the horizontal direction 14 in the image data 10. The road mark deletion process is performed every time the image data 10 is acquired. This is because in the process of deleting the road mark area in advance, the road mark area is deleted even when the vehicle 16 is on the road mark, and the vehicle 16 on the road mark may not be detected.

  FIG. 6 is an example of a pedestrian crossing 41. 6A, 6B, and 6C is an area in which a striped pattern 42 is drawn in white on the roadway 15. The road mark deletion unit 5 detects the area of the pedestrian crossing 41 from the image data 10 according to the following procedure. The area of the pedestrian crossing 41 detected by the road mark deletion unit 5 is excluded from the calculation when the vehicle 16 is detected. FIG. 7 is a flowchart of processing for excluding the pedestrian crossing 41 area from the image data 10 executed by the road mark deleting unit 5.

  First, the road mark deleting unit 5 deletes the edge 47 along the line segment facing the vanishing point 12 in the image data 10 (S21). Next, the road mark deleting unit 5 deletes the edge 48 in the horizontal direction 14 periodically present in the image data 10 (S22). This is because the stripe pattern 42 of the pedestrian crossing 41 exists in the horizontal direction 14 in the image data 10. Next, the road mark deletion unit 5 detects a repeating pattern having a period length in the horizontal direction 14 that matches the pedestrian crossing 41 (S23). The width 49 of the stripe pattern 42 of the pedestrian crossing 41 is defined. Therefore, the road mark deleting unit 5 can store in advance the width 49 of the striped pattern 42 in the horizontal direction 14 according to the coordinates in the vertical direction 13 in the image data 10.

  Ap (x, y) is a predetermined amount L from the luminance P3 (x, y) of the coordinate 43 (x, y) that is the target of the current calculation and the coordinate 43 (x, y) that is the target of the current calculation. It is an average value of luminance between the coordinates 44 (x + L, y) moved in the horizontal direction 14 and the luminance P4 (x + L, y). The predetermined amount L is the width in the horizontal direction 14 of the striped pattern 42 of the pedestrian crossing 41 according to the coordinate “y” in the vertical direction 13 in the image data 10. Ap (x, y) is an average value of the luminances of P3 and P4, and thus increases when the white portion of the pedestrian crossing. The road mark deletion unit 5 calculates Ap from the equation (7).

  Am (x, y) is a coordinate 45 (x−L / 2, y) moved in the lateral direction 14 by a predetermined amount (−L / 2) with respect to the coordinate 43 (x, y) that is the object of the current calculation. ) With a luminance P5 (x−L / 2, y) and a coordinate 46 (x + L /) that is moved in the lateral direction 14 by a predetermined amount (L / 2) with respect to the coordinate 43 (x, y) that is the object of the current calculation. 2, y) is the average value of the luminance between the luminance P6 (x + L / 2, y). The average value of luminance increases when the white part of the pedestrian crossing. The road mark deletion unit 5 calculates Am from the equation (8).

  The relationship between Ap and Am is a relationship in which the luminance is inverted because P5 and P6 are coordinates shifted by “L / 2” by 14 in the horizontal direction with respect to P3 and P4. The relationship in which the luminance is inverted is a relationship in which P5 and P6 are located in the asphalt portion of the striped pattern 42 of the pedestrian crossing 41 when P3 and P4 are positioned in the white painted portion of the striped pattern 42 of the pedestrian crossing 41. . That is, Am (x, y) has a minimum value when Ap (x, y) has a maximum value, and Am (x, y) has a maximum value when Ap (x, y) has a minimum value.

  Cp (x, y) is a difference absolute value between the luminance P3 (x, y) and the luminance P4 (x + L, y). Therefore, it becomes a small value when P3 and P4 are in the same state. For example, this is the case where P3 and P4 are located in white portions of the stripe pattern 42 of the pedestrian crossing 41, or P3 and P4 are located in the asphalt portion of the stripe pattern 42 of the pedestrian crossing 41, respectively. The road mark deletion unit 5 calculates Cp from the equation (9).

  Cm (x, y) is an absolute difference between the luminance P5 (x−L / 2, y) and the luminance P6 (x + L / 2, y). Therefore, the value is small when P5 and P6 are in the same state. For example, this is the case where P5 and P6 are located in white portions of the stripe pattern 42 of the pedestrian crossing 41, or where P5 and P6 are located in the asphalt portion of the stripe pattern 42 of the pedestrian crossing 41, respectively. The road mark deletion unit 5 calculates Cm from the equation (10).

  Based on the calculated Ap, Am, Cp, and Cm, the road mark deletion unit 5 calculates the feature point Ecw of the pedestrian crossing 41 from the equation (11).

  Ecw takes a large value when Ap and Am are in a reversed relationship. Further, Ecw becomes a large value when the absolute difference values Cp and Cm are small. Therefore, a high feature point is shown in the area of the pedestrian crossing 41.

  The road mark deletion unit 5 associates the calculated Ecw value with the corresponding pixel 11 (S24). The road mark deletion unit 5 performs the above calculation for each pixel 11 in the image data 10 (S25). The road mark deletion unit 5 excludes the area determined to be the pedestrian crossing 41 from the process of specifying the vehicle 16 (S26). Specifically, the pixel detected by the road mark deletion unit 5 as the pedestrian crossing 41 is associated with the pixel in the distribution map 50 created by the autocorrelation difference calculation unit 4, and is excluded from the target of subsequent vehicle 16 detection processing. To do. The autocorrelation difference calculation unit 4 can also create the distribution map 50 from the image data 10 after the crosswalk 41 is removed by the road mark deletion unit 5. By excluding the area of the pedestrian crossing 41, the image processing apparatus 1 can reduce erroneous detection of the vehicle 16 in the process of specifying the vehicle 16. The road mark deletion unit 5 acquires the area of the pedestrian crossing 41 to be deleted in FIG. 6C with respect to the original image data 10 in FIG.

  Next, the image processing apparatus 1 sets a region in which the shape determined according to the distribution state of the similarity value indicating the feature of the calculation region 32 matches the shape previously determined as the feature of the vehicle (subject) as the vehicle portion in the image data. Detect as. The image processing apparatus 1 detects the vehicle 16 from the distribution map 50 created by the autocorrelation difference calculation unit 4 and the road mark deletion unit 5.

  First, the relationship between the vehicle 16 and the difference 39 (similar value indicating characteristics) between the similar values 38 will be described. The image processing apparatus 1 detects the vehicle in the distribution map 50 using the shape of the left and right objects of the vehicle 16. FIG. 8 shows an image (A) of the vehicle 16 portion of the original image data 10 and an image (B) of the vehicle 16 portion of the distribution map 50. The camera 2 installed in the present embodiment photographs the vehicle 16 from the front or the back. The front or back of the vehicle 16 has a right / left target shape.

  The difference 39 between the similarity values 38 of the pixel groups 51 located on the left and right of the vehicle 16 is increased. This is because there are many corners on the left and right side surfaces of the vehicle 16. The difference 39 between the similarity values 38 of the pixel groups located in the front portion of the vehicle 16 becomes large. This is because the front portion of the vehicle 16 has many corner portions. The front part is also symmetrical from the center of the vehicle 16. The difference 39 between the similarity values 38 of the pixel group 52 located outside the left and right side surfaces of the vehicle 16 is reduced. This is because the difference 39 of the similarity value 38 is small because the outside of the vehicle 16 is asphalt. The difference 39 of the similarity value 38 corresponding to the pixel group 53 located in front of the front portion of the vehicle 16, that is, in front of the vehicle 16 becomes small. This is because the similarity value 38 is small because the outside of the vehicle 16 is asphalt. Small values are distributed among the similar values of the pixel group located in the front portion of the vehicle 16. It should be noted that the front of the front portion of the vehicle 16 facing forward is the coordinate in the vertical direction 13 below the vehicle 16 in the two-dimensional image data 10.

Based on the state in the distribution map 50 of the pixel group 51 located on the left and right of the vehicle 16, the pixel group 52 located outside the left and right side surfaces of the vehicle 16, and the pixel group 53 located in front of the vehicle 16, the vehicle Define the pattern that is characteristic of. The pattern that is a feature of the vehicle 16 in this embodiment is defined by the following conditions.
・ Condition 1 A pixel group 51 having a large difference 39 of the similarity values 38 is distributed on the left and right of the vehicle 16.
・ Condition 2 A pixel group 52 having a small difference 39 of the similarity values 38 is distributed in a region further outside both sides of the vehicle 16.
-Condition 3. A pixel group 53 having a small difference 39 of the similarity values 38 is distributed below the vehicle 16.

  The pattern feature quantity calculation unit 6 of the image processing apparatus 1 detects a region satisfying the above conditions from the distribution map 50.

  FIG. 9 is a flowchart of processing for detecting an area where the vehicle 16 exists from the distribution chart 10 executed by the pattern feature quantity calculation unit 6. The pattern feature quantity calculation unit 6 performs the following calculation for each pixel 54 in the distribution diagram 50 of FIG. First, the pattern feature quantity calculation unit 6 sets a pixel 54 as a current calculation target in the distribution map 50 as a detection target pixel 55 (S31). The pattern feature quantity calculator 6 calculates the feature quantity using the detection target pixel 55 as the reference position (S32). Specifically, the pattern feature quantity calculation unit 6 acquires the distribution state of the difference 39 of the similarity value 38 from the results of the following formulas (12) to (16). Note that “t” indicates that the currently acquired image data 10 is a target.

  The Elft 56 obtained by the equation (12) is the sum of the differences 39 of the similarity values 38 of the pixels 54 in the region defined by “k1” and “l1” with the detection target pixel 55 as a reference coordinate. Eight 57 obtained by the equation (13) is the sum of the differences 39 of the similarity values 38 of the pixels 54 in the region defined by “k2” and “l2” with the detection target pixel 55 as a reference position. Eleft 56 and Eight 57 correspond to the range from the center of the vehicle 16 to the left and right side portions. The side portion and the front portion of the vehicle 16 have many corner portions and many values having a large difference 39 between the similar values 38 are distributed. Note that “t” indicates that the currently acquired image data 10 is a target.

  Eside 58 obtained by Expression (14) is the sum of the differences 39 of the similar values 38 of the pixels 54 in the region defined by “k3” and “l3” on the left side of the distribution map 50 with the detection target pixel 55 as a reference position. Eside 58 corresponds to the outside of the side surface of the vehicle 16. A region further outside the side surface of the vehicle 16 is asphalt, and many values having a small difference 39 of the similarity values 38 are distributed.

  Ebottom 59 obtained by equation (15) is the sum of the differences 39 of the similarity values 38 of the pixels 54 in the region defined by “k4” and “l4” on the lower side of the distribution map 50 with the detection target pixel 55 as a reference position. . Ebottom 59 is on the lower side of the vehicle 16. When the reference position of the detection target pixel 55 is the lower end of the front portion of the vehicle 16, the area of the Ebottom 59 is a road. Therefore, the value of Ebottom 59 becomes small.

  The pattern feature quantity calculation unit 6 substitutes the values calculated by the above equations (12) to (15) into the following equation (16), and determines the feature size of the vehicle 16 corresponding to the detection target pixel 55. The indicated value 60E (x, y) is obtained. Note that “t” indicates that the currently acquired image data 10 is a target.

E (x, y) increases when the values of Eleft 56 and Eight 57 are large.
In addition, when the detection target pixel 55 is in the center of the vehicle 16, the Elft 56 and the Eight 57 are positioned symmetrically. Therefore, the value of Eleft 56 and the value of Eight 57 are close to each other. E (x, y) becomes small when there is a balance between the left and right of Eleft 56 and Eight 57. E (x, y) becomes smaller when the value of Eside 58 is small. E (x, y) decreases when the value of Ebottom 59 is small. Note that the values of Cside and Cbottom are values set by the designer according to the state of the road taken by the camera 2 and the like. The pattern feature quantity calculation unit 6 associates the calculated E (x, y) with the detection target pixel 55 (S33).

  The pattern feature quantity calculation unit 6 calculates the above processing for all the pixels 54 in the distribution map 50 (S34). The pattern feature quantity calculation unit 6 can acquire the coordinate in the vertical direction 13 and the coordinate in the horizontal direction 14 of an area where the vehicle 16 is considered to exist in the distribution map 50 based on the calculation result.

  In the calculation of the detection of the vehicle 16 by the pattern feature quantity calculation unit 6, the autocorrelation difference calculation unit 4 obtains the difference 39 of the similarity value 38 based on the calculation target pixel 33 for the original image data 10, and the difference of the similarity value 38. Since the distribution map 50 is created by 39, the influence of shadows and the like in the image data 10 is reduced.

  On the other hand, the target characteristic amount calculation unit 7 of the image processing apparatus 1 performs a calculation for obtaining the left and right center coordinates of the vehicle 16 from the distribution map 50. FIG. 10 is a diagram showing the relationship between the distribution map 50 and the object of the vehicle 16. The shape of the vehicle 16 is a right and left object. Therefore, the difference 39 of the similarity value 38 is distributed to the target from the center of the vehicle 16 to the left and right. Therefore, the objectivity feature quantity calculation unit 7 specifies the position of the vehicle 16 using the left-right symmetry of the vehicle 16. Evaluation of objectivity is performed according to the following procedure.

  FIG. 11 is a flowchart of a process for detecting a region where the vehicle 16 exists from the distribution chart 10 executed by the symmetric feature quantity calculation unit 7.

  The target characteristic amount calculation unit 7 sets the calculation central coordinates 61 (x, y) in FIG. 10 as the coordinates to be the target of the current calculation (S41). Next, the target characteristic amount calculation unit 7 calculates the characteristic amount of the calculation central coordinate 61 (x, y) (S42). More specifically, the symmetric feature quantity calculation unit 7 has coordinates 63 ”+ i” and a negative coordinate that are separated in the positive horizontal direction 14 around the coordinate “x” in the horizontal direction 14 of the calculation central coordinates 61 (x, y). The luminances of the coordinates separated by the coordinates 62 ″ -i ″ separated in the horizontal direction 14 are acquired. “I” is appropriately determined within a range in which the target property is to be calculated. For example, the range of “i” is set to half the width of the vehicle 16 corresponding to the coordinate in the vertical direction 13 in the image data 10. The characteristic feature amount calculation unit 7 sets the coordinate 63 away from the calculation central coordinate 61 (x, y) in the positive lateral direction 14 to “x + i, y”, and sets the luminance of the coordinate 63 “x + i, y” to P7 (x + i, y). y). The characteristic feature amount calculation unit 7 sets the coordinates 62 away from the calculation center coordinates 61 (x, y) in the negative lateral direction 14 as “xi, y”, and sets the luminance of the coordinates 62 “xi, y”. Let P8 (xi, y).

  The target characteristic feature amount calculation unit 7 calculates the total value of the current luminance P7 and luminance P8 by S (i) in Expression (17).

  S (i) takes a large value when the luminances P7 and P8 are large. Note that “t” indicates that the currently acquired image data 10 is a target. The target characteristic amount calculation unit 7 calculates a difference value between the current luminance P7 and the luminance P8 by D (i) in Expression (18).

  D (i) is a small value when the luminance P7 and the luminance P8 are close to each other. Therefore, D (i) indicates the balance between the luminance P7 and the luminance P8. Note that “t” indicates that the currently acquired image data 10 is a target.

  The objectivity feature amount calculation unit 7 calculates a line passing through the center of the vehicle by Esym (x, y) in Expression (19).

Esym (x, y) is the sum of absolute values of D (i) from the sum of absolute values of S (i) when “i” is changed within a predetermined range from the calculation central coordinates 61 (x, y). Subtracted value. Therefore, Esym (x, y) takes a large value when the left and right luminances in the region to be calculated are equal and the left and right luminances are large. The target characteristic amount calculation unit 7 associates Esym (x, y) calculated in S32 with the calculation central coordinate 61 (S43).

  By performing the above calculation for all the pixels 54 in the distribution map 50 by the symmetric feature value calculation unit 7 (S44), it is possible to detect a line passing through the center of the vehicle 16 in the distribution map 50.

  Further, the target characteristic amount calculation unit 7 can also detect a line passing through the center of the vehicle 16 by the equation (20) for obtaining the center of gravity Rsym (x, y) of the vehicle 16 by the following equation. Note that “t” indicates that the currently acquired image data 10 is a target.

  Next, processing performed by the vehicle position determination unit 8 will be described.

  The vehicle position determination unit 8 obtains the coordinate in the vertical direction 13 and the coordinate in the horizontal direction 14 where the vehicle 16 exists in the distribution map 50 from the pattern recognition processing of the vehicle 16 by the pattern feature value calculation unit 6. The coordinates in the vertical direction 13 are the lower end of the front of the vehicle 16 when the vehicle 16 is facing forward in the image data 10. Further, the vehicle position determination unit 8 obtains the coordinate in the horizontal direction 14 that is the center of the vehicle 16 in the distribution map 50 from the process of obtaining the line passing through the left and right centers of the vehicle 16 by the target characteristic amount calculation unit 7. The vehicle position determination unit 8 superimposes the calculation result obtained by the pattern feature quantity calculation unit 6 and the calculation result obtained by the target characteristic quantity calculation unit 7 to obtain a reference coordinate in which the vehicle 16 exists in the image data 10. The vehicle position determination unit 8 detects the luminance peak of the image data 10 or the distribution map 50 from the acquired reference coordinates, and acquires the number of pixels indicating the width of the front portion of the vehicle 16.

  A matching vehicle type in the database 20 is detected from the coordinates of the vertical direction 13 as a reference of the vehicle 16 and the number of pixels of the width of the front portion of the vehicle 16 obtained by peak detection. As described above, the image processing apparatus 1 can detect the position of the vehicle 16 in the image data 10 and specify the vehicle type of the vehicle 16.

  FIG. 12 is an example of vehicle detection according to this embodiment.

The original image data (A) is an image taken by the camera 2.
The vehicle portion (B) is an area where the vehicle portion of the original image data (A) is extracted.
Detection (C) according to the present embodiment is a detection range when detection is performed by the processing of the present embodiment.
Detection by background difference (D) is a detection range in the case of detection by the background difference method which is a conventional technique. Edge detection (E) is a detection range in the case of detection by an edge detection method that is a conventional technique. In the background difference method and the edge detection method, the range of only the vehicle cannot be detected because it is affected by the shadow portion.

  FIG. 13 is a hardware configuration example of the image processing apparatus 1 of the present embodiment. The image processing apparatus 1 includes a CPU 101, a camera 2, a memory 103, an output unit 104, and a bus 105. The CPU 101, the camera 2, the memory 103, and the output unit 104 are connected by a bus 105 and exchange data between the units. The CPU 101 is a central processing unit that reads and executes the image processing program 100 stored in the memory 103 in the work area. The image processing program 100 causes the CPU 101 of the image processing apparatus 1 to function as the autocorrelation difference calculation unit 4, the road mark deletion unit 5, the pattern feature amount calculation unit 6, the target feature amount calculation unit 7, and the vehicle position determination unit 8.

  The memory 103 stores various information of the image processing apparatus 1. For example, a RAM, a hard disk device, or the like. The memory 103 also serves as a work area when the CPU 101 performs calculations. In addition, information stored in the database 20 is also stored.

  The output unit 104 is an interface for outputting information of the position of the vehicle 16 detected in the image data 10 and the specified vehicle type. For example, the output destination is a wireless terminal possessed by another vehicle. Other vehicles receive information indicating that the vehicle 16 is present in the image data 10. In the case of wireless communication, an interface communicates with a wireless terminal using a wireless protocol (for example, a procedure defined in IEEE 802.11).

  In this embodiment, the difference 39 of the similarity values 38 between the pixels 11 is used as a basic feature amount for detecting the vehicle 16. In the feature amount due to the difference 39 of the similarity value 38, the values of the curve and the corner portion are large. Further, the fact that the front or back of the vehicle is symmetrical is utilized. Therefore, the feature amount based on the difference 39 of the similarity value 38 is effective when detecting the front or back of the vehicle 16. On the other hand, the feature amount due to the difference 39 of the similarity value 38 has a small response in a linear portion or a portion with a small curvature. Therefore, the feature amount based on the difference 39 of the similarity value 38 is less responsive to a straight portion such as a shadow or a portion that tends to be a simple shape. As a result, the feature amount based on the difference 39 of the similarity value 38 can reduce erroneous detection due to shadows and erroneous detection due to road surface reflection.

  Further, when the present embodiment is applied to a system that acquires the image data 10 at regular time intervals, the moving speed of the vehicle 16 is calculated by acquiring the difference in the position where the vehicle 16 exists among the plurality of image data 10. It is also possible to do.

  In the present embodiment, calculation processing is performed using the pixel 11 as a reference. However, a region in which a plurality of pixels 11 in the image data 10 are grouped can be used as a unit of calculation. That is, when the size is changed so as to reduce the number of pixels of the digital image data 10, the size is reduced so that a plurality of pixels 11 are gathered to form one pixel. On the other hand, when the size is changed so as to increase the number of pixels of the digital image data 10, the size of the pixel 11 interpolated from the relationship with the surrounding pixels 11 is increased to a new pixel 11. Therefore, performing detection in which a plurality of pixels 11 are regarded as a single region is substantially the same as performing detection processing for each pixel 11 after changing the size of the image data 10.

  Furthermore, in the present embodiment, the calculation area 32, the comparison area 34, and the search range 31 set by the autocorrelation difference calculation unit 4 are all assumed to be uniform in the image data 10. However, it is also possible to change the range of the search range 31, the calculation region 32, and the comparison region 34 in accordance with the region that is in front or the region that is behind when viewed as a three-dimensional space in the image data 10. . For example, for the region that is the back when viewed as a three-dimensional space, the ranges of the search range 31, the calculation region 32, and the comparison region 34 are reduced. On the other hand, the range of the search area 31, the calculation area 32, and the comparison area 34 is enlarged for the area that comes to the front when viewed as a three-dimensional space. By changing the range, the amount of calculation executed by the autocorrelation difference calculation unit 4 can be reduced. Further, by changing the range, it is possible to detect the accuracy according to each coordinate position. Specifically, data for changing the search range 31, the calculation region 32, and the comparison region 34 according to the coordinate value in the vertical direction 13 in the image data 10 of the calculation target pixel 33 is provided in advance. Is possible.

  In the present embodiment, the method of calculating the similarity value based on the luminance of the image data 10 has been described. In this embodiment, not only the luminance but also the similarity value can be calculated by a combination of brightness, hue, and saturation.

  INDUSTRIAL APPLICABILITY The present invention can be used in a system that notifies the state of a vehicle in image data by enabling detection of the position of the vehicle in image data with a small amount of computation.

Claims (9)

An image data recognition method for detecting a subject in image data,
Determining a detection target area for detecting a subject in the image data ;
Determining a plurality of first regions each first pixel of each of the plurality including a respective center and a vital first predetermined number of pixels of the plurality of first pixels included in the determined detection target area And
Determining a search range to a respective one pixel of each of the plurality including respective center Toshikatsu second predetermined number of pixels of the plurality of first pixels,
It determined the respective search ranges INCLUDED multiple second plurality including the number of pixels corresponding to a predetermined number each of centered and vital said first pixel and the second region each plurality of Determined for each second pixel,
By comparing the characteristics of the determined images of the plurality of second regions and the images of the first region, the similarity values between the first region and the plurality of second regions are respectively determined. For each of the plurality of second regions of
Associating the calculated similarity values for the plurality of second regions with the plurality of second pixels,
Second similar indicating that most similar to said first region within each similarity value associated with the second pixel of the plurality other than the first pixels included in each search range the Select a value
Associating the selected second similarity value with the first pixel corresponding to the respective search range;
The distribution is obtained by mapping the second similarity value associated with each of the plurality of first pixels in the detection target region to the position of the first pixel,
A method of recognizing an image data, wherein an area where a shape determined according to the distribution state matches a shape previously determined as a feature of the subject is detected as a portion of the subject in the image data. The image data is imaged from obliquely above the road so that the road, the subject vehicle on the road, and the vanishing point of the road are shown.
The number of pixels that should be included in the first area, the second area, and the search range used to detect an area corresponding to the front side of the roadway in the image data is displayed on the back side of the roadway in the image data. The image data recognition method according to claim 1 , wherein the number of pixels to be included in the first region, the second region, and the search range used for detection of the corresponding region is larger. .   2. The image data recognition method according to claim 1, wherein the shape defined as the feature of the subject is a condition for determining the shape of the vehicle. In the obtained distribution of the second similarity values, a rectangular third region is detected in which similarity values indicating that the similarity values are less similar than the similarity values distributed in the surroundings are detected,
Similar values indicating dissimilarity on the left and right sides in the detected third region are distributed, and the similar values indicating dissimilarity on the left and right sides in the third region are There is symmetry sandwich the center, and, if the similarity value indicating that similar to the ambient on the lower side of said third region is distributed, the coordinates corresponding to the lower side of said third region and the 4. The image data recognition method according to claim 3, wherein the coordinates are front coordinates of the vehicle. The vehicle obtains the center position of the vehicle in the lateral direction in the image data from the target of the distribution of the second similarity value of the vehicle, and from the conditions for determining the shape of the vehicle distribution and the shape of the vehicle 4. The image data recognition method according to claim 3, wherein a position in the vertical axis direction is specified. Image data recognizing method according to claim 2, wherein the determining the number of pixels the search range to be in the range of the vehicle may be present around the first pixel. After erasing the edge of the line segment along the vanishing point direction from the edge of the image data and the edge of the line segment periodically present in the direction intersecting the roadway of the image data, the distribution of similarity values is calculated The image data recognizing method according to claim 2 . An image data recognition device for detecting a subject in image data,
Means for determining a detection target region for detecting a subject in the image data ;
Determining a plurality of first regions each first pixel of each of the plurality including a respective center and a vital first predetermined number of pixels of the plurality of first pixels included in the determined detection target area Means to
It means for determining a search range for each first pixel of each of the plurality including respective center Toshikatsu second predetermined number of pixels of the plurality of first pixels,
It determined the respective search ranges INCLUDED multiple second plurality including the number of pixels corresponding to a predetermined number each of centered and vital said first pixel and the second region each plurality of Means for determining each second pixel ;
The similarity values between each and said first region the plurality of second regions by you compare the characteristics of the respective image and the image of the first region of the determined plurality of second regions Means for calculating each of the plurality of second regions ;
Means the calculated plurality of second the similarity value for each area that associates to a second pixel of each of the plurality,
Select a similarity value indicating that the most similar to said first region within each similarity value associated with the second pixel of the plurality other than the first pixels included in each search range the Means to
Means that the selected said similarity value to correspond to the first pixels corresponding to the search range of each said,
Means asking you to distribution by mapping the similar values associated with each of the plurality of first pixel within the detection target region to a position of the first pixel,
An image data recognition apparatus comprising: means for detecting, as a portion of a subject in image data, a region in which a shape determined according to the distribution state matches a shape previously determined as a feature of the subject. An image data recognition program for detecting an object in the image data, the computer,
Determining a detection target area for detecting a subject in the image data ;
Determining a plurality of first regions each first pixel of each of the plurality including a respective center and a vital first predetermined number of pixels of the plurality of first pixels included in the determined detection target area Means to
Determining a search range to a respective one pixel of each of the plurality including respective center Toshikatsu second predetermined number of pixels of the plurality of first pixels,
It determined the respective search ranges INCLUDED multiple second plurality including the number of pixels corresponding to a predetermined number each of centered and vital said first pixel and the second region each plurality of Determined for each second pixel ,
By comparing the characteristics of the determined images of the plurality of second regions and the images of the first region, the similarity values between the first region and the plurality of second regions are respectively determined. second was calculated for each region of the plurality of,
Associating the calculated similarity values for the plurality of second regions with the plurality of second pixels,
Select a similarity value indicating that the most similar to said first region within each similarity value associated with the second pixel of the plurality other than the first pixels included in each search range the Means to
Associating the selected similarity value with the first pixel corresponding to the respective search range;
The distribution is obtained by mapping the similarity value associated with each of the plurality of first pixels in the detection target region to the position of the first pixel,
Image data recognition characterized by functioning to execute a process of detecting, as a portion of the subject in the image data, an area in which the shape determined according to the distribution state matches the shape previously determined as the feature of the subject program.

Images