JP2009276910A - Image processor, method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To extract a person candidate area having high accuracy that it is an area corresponding to a person from an image without causing complication of a configuration or processing. <P>SOLUTION: This image processor searches for/extracts the person candidate area estimated that it corresponds to the person with the whole area of the processing target image as a target (52), extracts edge pixels inside the individual extracted person candidate area (58-68), searches for/extracts an artificial linear line segment or curved line segment inside the area (70-86), sequentially performs processing for excluding the artificial line segment from an output target (90) when the artificial linear line segment or curved line segment clearly exists inside the area (affirmation in 88), and outputs the person candidate area left without being excluded from the output target as the area having the high accuracy that it is the person area (94). <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an image processing apparatus, method, and program, and more particularly to an image processing apparatus that extracts a human candidate region estimated to correspond to a person from an image, an image processing method applicable to the image processing apparatus, and a computer. The present invention relates to an image processing program for functioning as an image processing apparatus.

  Conventionally, image processing is performed in which an image of a vehicle is captured by a camera mounted on the vehicle, and an image portion (person candidate region) estimated to correspond to a person is searched for and extracted from the image obtained by the imaging. Thus, a technique for detecting a person such as a pedestrian existing around the vehicle has been proposed. In this technique, various algorithms such as a neural network method and a subspace method described in Non-Patent Document 1 exist as algorithms applicable to search / extraction of human candidate regions. When searching / extracting a human candidate area by applying an algorithm, if the image part has an image part similar to the human candidate area, for example, on a general road, trees, utility poles, signs, white lines, An image portion corresponding to a pedestrian crossing may be erroneously extracted as a person candidate region.

  Preventing erroneous extraction of a person candidate area can be realized by changing the threshold used to determine whether the person candidate area is a person candidate area in an algorithm for searching and extracting a person candidate area to a value that can reduce erroneous extraction. In this case, although the erroneous extraction of the person candidate area is reduced, the omission of extraction in which the image portion corresponding to the person is not extracted as the person candidate area increases. Also, if the threshold value is set with an emphasis on preventing omission of extraction of human candidate regions, the omission of extraction decreases while the extraction of human candidate regions increases. Thus, there is a trade-off between preventing erroneous extraction of a person candidate area and avoiding an extraction omission where an image portion corresponding to a person is not extracted as a person candidate area. It was a difficult task to achieve both.

  In relation to the above, Patent Document 1 detects, as an infrared image, an area in a range corresponding to the amount of infrared radiation emitted from a person in an infrared image obtained by photographing a detection target range, By detecting a high-luminance part in a visible image with a clear relationship between positions on the image, and excluding the detected image area of the infrared image corresponding to the high-luminance part and its periphery from the human detection target area. In addition, a technique is disclosed that reduces the rate of erroneously detecting an object having a similar area to an image, such as a headlight of an automobile, and having the same area as the human candidate area.

  Further, in Patent Document 2, when detecting a pedestrian with an infrared camera, a lane marking of a road is imaged with a visible light camera, and an area near the lane marking position is set as a pedestrian search area. A technology that shortens the processing time and enables early warning by setting a sandwiched area as a pedestrian search area and setting only the search area as a pedestrian search / detection target is disclosed.

Patent Document 3 proposes a technique for extracting a human face area from a still image. As a pre-process for specifying a face area, a straight line is drawn on the outline of the divided area on the binarized image. It is described that an artificial object is judged based on an existing ratio.
Kenichiro Ishii, Eisaku Maeda, Noriyoshi Ueda, Hiroshi Murase, “Intuitive Pattern Recognition”, Ohmsha, August 1998 JP 2001-108758 A JP 2002-362302 A JP-A-8-122944

  However, the technique described in Patent Document 1 detects a person candidate area using two images, an infrared image and a visible image, and it is necessary to provide an infrared camera and a visible light camera, respectively. There is a problem of increasing complexity. In particular, in the technique described in Patent Document 1, an area in a range corresponding to the amount of infrared radiation emitted from a person in an infrared image is detected as a person candidate area. Cameras are essential, but visible images can be used to detect pedestrians and other vehicles in the vicinity of vehicles, as well as detecting preceding vehicles, road signs, lane marks, etc. It is not suitable for detection of people other than pedestrians and the like, and its use is limited and low in versatility. It is costly to install an infrared camera on a vehicle only for the detection of pedestrians, etc. Not desirable. In the technique of Patent Document 1, the correspondence between the position on the infrared image and the position on the visible image needs to be clear, but it is necessary to make each camera an optical system with small aberration. However, there is a problem in that the correspondence between two different types of images must be derived by image processing, which further complicates the apparatus configuration or increases the processing load.

  Further, the technique described in Patent Document 2 also detects a human candidate region using two images, an infrared image and a visible image, as in the technique described in Patent Document 1, so that the apparatus configuration is complicated. There is a problem of becoming. In addition, the technique described in Patent Document 2 has a drawback that it cannot be applied to an image in which no lane marking exists (it cannot be applied to a road on which no lane marking exists), and lane marking is present in the image. Even if it is, the process of extracting the lane marking from the image is complicated, and a processing load is added to the processing device. If there is an up / down on the road that is running, this up / down is not accurately detected. There is also a problem that the search area for pedestrians cannot be set appropriately.

  Moreover, in the technique described in Patent Document 3, region segmentation is performed after binarizing a still image, but a complex image such as an image (moving image or the like) obtained by capturing an image of the periphery of the vehicle. Even if this technique is applied, it is difficult to properly divide the image area by binarization using a single threshold. Moreover, even if the linearity of the contour of the region that is not properly divided is examined, an expected result cannot be obtained.

  The present invention has been made in consideration of the above-described facts, and an image processing apparatus and image processing that can extract a human candidate area with a high degree of accuracy, which is an area corresponding to a person, from an image without complicating the configuration and processing. The object is to obtain a method and an image processing program.

  There may be various subjects in the image, but if there are artifacts in the image (for example, if the image is taken from the vehicle, the above-mentioned artifacts are the utility poles, signs, white lines, crossings, etc. There may be sidewalks, roadside trees, etc.), and the image part corresponding to the artifacts in the image has a straight line, a line segment close to a straight line, a curve with a certain curvature, a curve with a certain curvature, etc. It is very likely that a straight line segment exists. It is very likely that there is no image portion corresponding to a person in an area where an image portion corresponding to an artificial object exists.

  Based on the above, the image processing apparatus according to the first aspect of the present invention includes an artificial line segment detecting unit that detects an artificial line segment existing in a visible image to be processed, and the visible image to be processed. A human candidate region extraction unit that extracts a human candidate region that is estimated to correspond to a person that exists in a region excluding a region in which the artificial line segment detected by the artificial line segment detection unit exists. It is configured to include.

  In the first aspect of the present invention, since the artificial line segment existing in the visible image to be processed is detected, based on the detection result of the artificial line segment, the artifact in the visible image to be processed It is possible to determine a region having a high probability that an image portion corresponding to the above exists. In the first aspect of the present invention, the region where the artificial line segment detected by the artificial line segment detection means exists in the visible image to be processed, that is, the image portion corresponding to the person does not exist. Since the candidate candidate area that is estimated to be equivalent to a person is extracted from the area that excludes a highly likely area, the artifact area in which the image portion corresponding to the artifact exists is used as the candidate person area. It is possible to prevent erroneous extraction, and it is possible to extract a person candidate area with high accuracy that is an area corresponding to a person.

  In the first aspect of the present invention, since the human candidate region is extracted using the visible image, an infrared image is not necessary for extracting the human candidate region, and the configuration is complicated to obtain the infrared image. Or processing for associating the position on the infrared image with the position on the visible image. Therefore, according to the first aspect of the present invention, it is possible to extract a human candidate area with high accuracy, which is an area corresponding to a person, from an image without complicating the configuration and processing.

  Note that the invention described in claim 1 does not judge an artificial object based on the outline of an area obtained through binarization and area division as in the technique described in Patent Document 3, but in a complicated image. Detects artificial line segments (straight lines, curves, etc.) that can be detected using image edges (differential values) that can be detected stably, and extracts human candidate regions based on the detection results of artificial line segments. ing. The edge of the image also appears as the outline of the area when binarization and area division are performed as described above, but also appears in other parts. In the technique described in Patent Document 3, the accuracy of artifact determination depends on the accuracy of binarization and area segmentation. However, when an artificial line segment is detected using an edge of an image, an artificial line is detected. The minute detection accuracy is not affected by the accuracy of other processes such as binarization. Therefore, the invention according to claim 1 can accurately detect an artificial line segment, and can extract a human candidate region with high accuracy, which is a region corresponding to a person, with high accuracy.

  In the invention described in claim 1, for example, as described in claim 2, the artificial line segment detection means searches for the person candidate area for the entire area of the visible image to be processed by the person candidate area extraction means. Is detected, an artificial line segment is detected for the candidate human region extracted in the search, and the human candidate region extracting means searches for the human candidate region for the entire region of the visible image to be processed. After the artificial line segment is detected by the artificial line segment detection means, an artificial line segment is detected in the area by the artificial line segment detection means among the candidate human regions extracted in the search. The person candidate area can be excluded from the person candidates. According to the second aspect of the present invention, since the human candidate area is searched for the entire area of the visible image to be processed, when the human candidate area does not exist in the visible image to be processed (result of the search) In the case where no person candidate area is extracted), it is possible to reduce the processing load when the person candidate area does not exist in the visible image to be processed by stopping the subsequent processing.

  Further, in the invention described in claim 1, for example, as described in claim 3, the artificial line segment detection means detects the artificial line segment in the entire region of the visible image to be processed, and the person candidate The region extraction means is configured to search for a human candidate region for a region excluding the region where the artificial line segment detected by the artificial line segment detection unit exists in the visible image to be processed. May be. In the invention described in claim 3, since the human candidate area is searched for the area excluding the area where the artificial line segment exists in the visible image to be processed, the algorithm is generally complicated and the processing time is long. The search range for searching for a candidate human area that is likely to be timed becomes narrower than that of the invention according to claim 2, and the time required for searching for the candidate person area can be shortened.

  Further, in the invention described in claim 1, the artificial line segment detecting means, as described in claim 4, for example, as an artificial line segment, the edges in the visible image to be processed are substantially linearly continuous. To detect a straight line segment or a curved line segment in which edges in a visible image to be processed are formed continuously and the variation of curvature in the continuous edge transitions within a predetermined range with respect to a reference value. Can be configured. By detecting the straight line segment or the curved line segment as the artificial line segment, the artificial line segment existing in the artifact region in the image can be reliably detected.

  In the invention according to claim 4, the artificial line segment detecting means, for example, includes all straight line segments and curved line segments existing in the visible image to be processed (regardless of the extending direction of the line segment, etc.). Although it may be configured to extract, for example, as described in claim 5, as an artificial line segment, among straight line segments, in a visible image to be processed corresponding to a vertical direction in real space Alternatively, only a straight line segment extending substantially along the vertical direction may be detected. In this case, all the curved components are excluded from the detection target, and the linear components extending along the direction different from the vertical direction in the visible image are also excluded from the detection target, and the artificial components existing in the visible image to be processed are present. Although all the line segments cannot be detected, the artificial line segment can be detected by the artificial line segment detection means in a short time. In addition, in an image captured from the vehicle and the like, most of the artifacts that may appear in the image are artifacts that generate a straight line segment extending substantially along the vertical direction in the image. When an image is a processing target, even if the artificial line segment detection means is configured to detect only a straight line segment extending substantially along the vertical direction in the visible image to be processed as described above, It can be avoided that the extraction accuracy of the region corresponding to is reduced.

  Further, in the invention according to claim 4, the extraction of the straight line segment or the curved line segment by the artificial line segment detection means is specifically, for example, as described in claim 6, the visible image to be processed. Is extracted, edge pixels corresponding to the edges in the image to be processed are extracted, and then a single edge pixel is selected as the processing target from the extracted edge pixels, and the processing is performed in the vicinity of the processing target edge pixel. It is determined whether there is another edge pixel within the angle range corresponding to the linear component or the curved component with respect to the target edge pixel, and if the corresponding edge pixel exists, the corresponding edge pixel is determined. By repeating the process of selecting as a new processing target, it can be realized by detecting a straight line segment or a curved line segment composed of a series of edge pixels as an artificial line segment.

  In addition, in the invention according to claim 4, the extraction of the straight line segment or the curved line segment by the artificial line segment detection means is not limited to the above. For example, as described in claim 7, the processing target After the edge pixels corresponding to the edges in the image to be processed are extracted by differentiating the visible image, the Hough transform is applied to the extracted edge pixels, so that the edge pixels are connected as an artificial line segment. This can also be realized by detecting a straight line segment or a curved line segment configured as described above.

  Moreover, in the invention according to any one of claims 1 to 7, as the visible image to be processed, for example, as described in claim 8, the periphery of the vehicle is imaged by an imaging means mounted on the vehicle. The moving image obtained by this is suitable.

  The image processing method according to the invention described in claim 9 searches for a human candidate area estimated to correspond to a person for all areas of a visible image to be processed, and targets a human candidate area extracted by the search. By detecting the artificial line segment and excluding the human candidate area in which the artificial line segment is detected in the human candidate area extracted in the search from the human candidate, the visible image to be processed Since the person candidate area existing in the area excluding the area where the artificial line segment exists is extracted, the configuration and processing are not complicated as in the inventions according to claims 1 and 2. It is possible to extract a human candidate area with high accuracy, which is an area corresponding to a person, from the image, and to reduce the processing load when the human candidate area does not exist in the visible image to be processed.

  The image processing method according to the invention of claim 10 detects an artificial line segment for the entire region of the visible image to be processed, and the artificial line segment exists in the visible image to be processed. By searching for a candidate human area that is estimated to be equivalent to a person with an area that excludes the existing area as a target, it exists in the area that excludes the area where the artificial line segment exists in the visible image to be processed Since the person candidate area is extracted, similarly to the first and third aspects of the invention, it is possible to extract a person candidate area having a high degree of accuracy, which is an area corresponding to a person, without complicating the configuration and processing. At the same time, the time required for searching for the person candidate area can be shortened.

  According to an eleventh aspect of the present invention, there is provided an image processing program comprising: an artificial line segment detecting unit that detects an artificial line segment existing in a visible image to be processed; Among them, function as a human candidate area extracting means for extracting a human candidate area estimated to be equivalent to a person existing in an area excluding the area where the artificial line segment detected by the artificial line segment detecting means exists Let

  The image processing program according to the invention described in claim 11 is a program for causing a computer to function as the above-described artificial line segment detection means and person candidate area extraction means. By executing the image processing program, the computer functions as the image processing apparatus according to claim 1, and, similarly to the invention according to claim 1, the computer can function without complication of configuration and processing. It is possible to extract a human candidate area with high accuracy, which is a corresponding area, from an image.

  As described above, the present invention detects an artificial line segment that exists in a visible image to be processed, and extracts a human candidate area that exists in an area that excludes the artificial line segment from the visible image to be processed. Since extraction is performed, there is an excellent effect that it is possible to extract a human candidate region having a high degree of accuracy, which is a region corresponding to a person, from an image without complicating the configuration and processing.

  Hereinafter, an example of an embodiment of the present invention will be described in detail with reference to the drawings.

[First Embodiment]
FIG. 1 shows an object detection apparatus 10 according to the present embodiment. The object detection device 10 is mounted on a vehicle 12 (see FIG. 2) and detects an object existing around the vehicle 12 (for example, in front of the vehicle 12). The visible light camera 14 and the image processing unit 24 are connected to a bus 26. It is connected and configured. The object detection device 10 is a part of a vehicle control system that is mounted on the vehicle 12 and has various functions, and various sensors and various control units are actually connected to the bus 26. However, these illustrations are omitted.

  The visible light camera 14 includes a CCD image sensor, a MOS type solid-state imaging device, and the like, and an imaging device 16 having an sensitivity in the visible light region, an analog processing unit 18, an A / D (analog / digital) converter 20, and a digital processing unit 22. And these are connected in order. As shown in FIG. 2, the visible light camera 14 is located in the vehicle 12 near the upper end of the windshield glass of the vehicle 12 and approximately in the center in the left-right direction of the vehicle 12 (for example, in the vicinity of the room mirror). Are arranged in an orientation in which the front of the image pickup range. The image sensor 16 receives light from the imaging range in front of the vehicle 12 on the light receiving surface, and photoelectrically converts the light incident on the light receiving surface to represent an analog image representing a situation in the imaging range in front of the vehicle 12. The signal is output at a constant frame rate.

  The analog processing unit 18 amplifies the image signal output from the image sensor 16 and corrects white balance or the like on the amplified image signal. The image signal output from the analog processing unit 18 is converted into digital image data by the A / D converter 20 and input to the digital processing unit 22, and after processing such as color correction is performed by the digital processing unit 22. And output to the image processing unit 24 via the bus 26. The image data output from the visible light camera 14 to the image processing unit 24 corresponds to a visible image to be processed according to the present invention (more specifically, a moving image according to claim 8).

  The image processing unit 24 includes a microcomputer and the like, and includes a CPU 24A, a memory 24B including a ROM and a RAM, a non-volatile storage unit 24C including an HDD (Hard Disk Drive), a flash memory, and the like. Image data input via the bus 26 is sequentially stored in the storage unit 24C. In order to save the storage capacity of the storage unit 24C, the image data input from the visible light camera 14 may be temporarily compressed and stored in the storage unit 24C. The storage unit 24C is preinstalled with a human candidate region extraction program for performing a human candidate region extraction process described later by the CPU 24A. This person candidate area extraction program corresponds to the image processing program according to claim 11, and the image processing unit 24 functions as an image processing apparatus according to the present invention by the person candidate area extraction program being executed by the CPU 24 </ b> A. To do.

  Next, as an operation of the first embodiment, a person candidate region extraction process according to the first embodiment will be described with reference to FIG. This person candidate region extraction process is repeatedly executed by the image processing unit 24 while the ignition switch of the vehicle 12 is turned on. The execution period of the person candidate area extraction process may be the same as the output period of the image signal from the image sensor 16 (the reciprocal of the frame rate), or may be longer than the output period of the image signal. In addition, the person candidate area extraction processing according to the first embodiment is processing to which the image processing method according to claim 9 is applied.

  In the candidate human region extraction process according to the first embodiment, first, in step 50, image data for one frame input from the visible light camera 14 and temporarily stored in the storage unit 24C is read from the storage unit 24C, and the processing target image is read out. The image data is stored in the memory 24B. In the next step 52, the entire region of the processing target image in which the image data is stored in the memory 24B is targeted, and the region (person candidate) estimated to correspond to the person existing in the processing target image by pattern recognition. Search / extract (region). As pattern recognition in step 52, pattern recognition by statistical machine learning represented by, for example, the neural network method and subspace method described in Non-Patent Document 1 is preferable, but other methods are applied. Also good.

  Pattern recognition by statistical machine learning involves collecting a large number of recognition target images, extracting some common features from the collected images by machine learning, and performing pattern recognition using the extracted features. . A set of a plurality of features extracted from a large number of images by machine learning is called a learning model. When searching / extracting a human candidate area by pattern recognition as in step 52, a learning model representing the outline of the human candidate area as shown in FIG. 4B is used as an example learning model. The outer shape of the learning model is often rectangular as in the example shown in FIG. 4B, but other shapes can also be used. In addition, as a learning model for searching / extracting a human candidate region, for example, a contour of a walking person other than the learning model representing the contour shape of an upright person as shown in FIG. A learning model that represents a shape, a learning model that represents the contour shape of a person riding a two-wheeled vehicle such as a bicycle, and the like may also be used.

  Pattern recognition by statistical machine learning is the evaluation value of the similarity between the learning model placed in an arbitrary position on the processing target image and the area where the placed learning model overlaps the learning model on the processing target image Is repeated by changing the arrangement position of the learning model on the processing target image and the relative size of the learning model with respect to the processing target image. As shown in FIG. 4A, the learning model is usually resized by reducing the size of the learning model in an area where an image portion of an object having a large distance from the vehicle 12 exists in the processing target image. Then, the size of the learning model is increased in an area where an image portion of an object having a small distance from the vehicle 12 exists.

  If the calculated evaluation value is equal to or greater than the threshold, the area overlapping with the learning model at that time is determined as the person candidate area, and the position and size of the area determined as the person candidate area are stored in the memory 24B. . For example, if the processing target image is the image shown in FIG. 4A, an area surrounded by a rectangle in FIG. 4C is extracted as a person candidate area from the processing target image as an example, The size and the like are stored in the memory 24B.

  In the next step 54, it is determined whether or not a person candidate region has been extracted from the processing target image by the processing in step 52. If this determination is negative, the person candidate area extraction process is terminated. Thereby, when no person is present around the vehicle 12 and the processing target image is an image in which no person area corresponding to the person exists, the load applied to the image processing unit 24 can be reduced. it can. If the determination in step 54 is affirmative, the process proceeds to step 56, and a single person candidate area is selected as a process target from the person candidate areas extracted from the process target image by the process in step 52.

In the next step 58, the edge strength and direction of each pixel in the person candidate area selected as the processing target in step 56 are calculated. That is, first, the edge strength in the vertical direction and the edge strength in the horizontal direction are calculated for each pixel in the person candidate area to be processed. For the calculation of the edge strength in the vertical direction and the horizontal direction, for example, the Previt operator in the vertical direction and the horizontal direction shown in FIG. 5A, the sobel operator in the vertical direction and the horizontal direction shown in FIG. The differential operator can be used. Next, for each pixel in the person candidate area to be processed, the calculated edge intensity Diff _V and edge edge Diff _H in the horizontal direction are substituted into the following expressions (1) and (2) to obtain the edge intensity. E and edge direction θ are respectively calculated, and the calculation result is stored in the memory 24B.
Edge strength E = √ ((Diff _V ) ² + (Diff _H ) ² ) (1)
Edge direction θ = tan ⁻¹ (Diff _V / Diff _H ) (2)
Note that the edge direction θ calculated by the expression (2) is the direction in which the edge intensity is maximum in the calculation target pixel, that is, the low density region and the high density region passing through the calculation target pixel as shown in FIG. The normal direction with respect to the boundary line is represented, and the edge strength E calculated by the expression (1) represents the edge strength in the edge direction θ in the pixel to be calculated. By the above calculation, for example, an edge image as shown in FIG. 7 is obtained from the person candidate region shown in FIG. 4C, for example. Note that the edge image shown in FIG. 7 is visualized by replacing the density of each pixel in the human candidate area shown in FIG. 4C with a value that decreases as the value of the edge strength E increases.

  In the next step 60 to step 68, based on the edge intensity E and the edge direction θ of each pixel in the processing target human candidate area calculated in step 58, a boundary between the low density area and the high density area is selected from the pixels. Pixels that are estimated to be located on the line (edge pixels: pixels that are shown at a low density in the edge image shown in FIG. 7) are extracted. That is, first, in step 60, the target pixel is selected from the pixels in the person candidate area to be processed. In the next step 62, based on the edge direction θ of the target pixel selected in step 60, an adjacent pixel that is present in the same direction as the target pixel's edge direction θ is selected for the target pixel, and the edge of the target pixel is selected. The intensity E is compared with the edge intensity E of the selected adjacent pixel.

The selection of adjacent pixels in step 62 is, for example,
(1) Edge direction θ: 0 ° ± 22.5 ° or 180 ° ± 22.5 ° → Adjacent pixel: right adjacent pixel and left adjacent pixel
(2) Edge direction θ: 45 ° ± 22.5 ° or 225 ° ± 22.5 ° → Adjacent pixels: upper right and lower left pixels
(3) Edge direction θ: 90 ° ± 22.5 ° or 270 ° ± 22.5 ° → Adjacent pixel: upper adjacent pixel and lower adjacent pixel
(4) Edge direction θ: 135 ° ± 22.5 ° or 315 ° ± 22.5 ° → Adjacent pixels: Lower right and upper left pixels As described above, the angle range (0 to 360 °) of the edge direction θ is divided into four angles. Adjacent pixel selection conditions that determine the adjacent pixels to be selected for each angle range are defined, and the adjacent pixels are selected according to which of the four types of angle ranges the edge direction θ belongs to. Can be done.

  In the above condition, two pixels are selected as adjacent pixels regardless of the value of the edge direction θ. However, the present invention is not limited to this. For example, the individual angle range ± 22.5 ° is slightly expanded (for example, ± 30). If the edge direction θ is an angle corresponding to an overlap region that spans two types of angle ranges, the angle range is set to one of the two types of angle ranges. A total of four corresponding pixels may be selected as adjacent pixels, and the edge intensity E may be compared with the target pixel.

  In the next step 64, as a result of comparing the edge strength E in step 62, whether the edge strength E of the target pixel is maximum (whether the edge strength E of the target pixel is larger than the edge strength E of a plurality of adjacent pixels) or not. judge. If a pixel having an edge strength E equal to the target pixel exists among the plurality of adjacent pixels, a new adjacent pixel adjacent to the adjacent pixel having the edge strength E equal to the target pixel is selected as the above-mentioned adjacent pixel selection. The selection is performed using conditions, and the edge strength E of the selected new adjacent pixel and the target pixel is compared. If the determination in step 64 is negative, the process proceeds to step 68. If the determination in step 64 is affirmative, the process proceeds to step 66, and information indicating that the target pixel is an edge pixel is stored in the memory 24B. After that, the process proceeds to step 68.

  In step 68, it is determined whether or not the processing in steps 60 to 68 has been performed for all the pixels in the candidate human subject area to be processed. If the determination is negative, the process returns to step 60, and steps 60 to 68 are repeated until the determination of step 68 is affirmed. Thereby, all the edge pixels existing in the candidate human subject area to be processed are extracted.

  If the determination at step 68 is affirmative, the routine proceeds to step 70. In Step 70 to Step 86, the individual edge pixels extracted in Step 60 to Step 68 are sequentially selected as determination targets, and the line segment corresponding to the artificial object may be included in the person candidate region to be processed. Search for high artificial line segments (artificial linear and curved line segments). That is, first, in step 70, it is determined whether or not all edge pixels extracted in steps 60 to 68 have been selected as determination targets. Initially, this determination is denied and the routine proceeds to step 72, where a single edge pixel is extracted as an edge pixel to be determined from the edge pixels in the candidate human subject area extracted in steps 60 to 68. . In the next step 74, based on the edge direction θ of the determination target edge pixel selected in step 72, the determination target edge pixel is approximately 90 ° or −90 ° with respect to the determination target edge pixel edge direction θ. The adjacent pixels existing in the direction of are selected.

Regarding the selection of adjacent pixels in step 74, for example,
(1) Edge direction θ: 0 ° ± 30 ° or 180 ° ± 30 ° → Adjacent pixel: Upper and lower adjacent pixels
(2) Edge direction θ: 45 ° ± 30 ° or 225 ° ± 30 ° → Adjacent pixels: lower right and upper left pixels
(3) Edge direction θ: 90 ° ± 30 ° or 270 ° ± 30 ° → Adjacent pixels: right and left adjacent pixels
(4) Edge direction θ: 135 ° ± 30 ° or 315 ° ± 30 ° → Adjacent pixels: Upper right and lower left pixels As described above, the angle range (0 to 360 °) of the edge direction θ is divided into four types of angle ranges. The adjacent pixel selection condition that determines the adjacent pixels to be selected for each angle range is determined, and the adjacent pixels are selected according to which of the four types of angle ranges the edge direction θ belongs to. Can be done. Under the above conditions, an overlap region is provided in each angle range, and when the edge direction θ is an angle corresponding to the overlap region over two types of angle ranges, any of the two types of angle ranges is selected. A total of four pixels corresponding to these are selected as adjacent pixels.

  Note that the individual angle ranges under the above conditions are not limited to ± 30 °. If the overlap region is too wide, the processing time will be long. On the other hand, if the overlap region is too narrow, the extending direction of the artificial line segment to be searched in step 70 to step 86 is limited to 0 °, 45 °, etc. Therefore, the individual angle range ± 30 ° may be slightly widened (for example, ± 40 °), and the angle range of the overlap region extending over each angle range may be slightly enlarged.

  In step 76, it is determined whether or not the adjacent pixel selected in step 74 is an edge pixel. If the determination is negative, it is determined that the edge pixel to be determined is isolated from the other edge pixels and is not an edge pixel constituting an artificial line segment, and the process returns to step 70. In this case, another edge pixel is selected as the determination target edge pixel in step 72, and the subsequent processing is repeated.

  On the other hand, if the determination in step 76 is affirmative, the process proceeds to step 78 where the edge direction θ of the adjacent pixel selected in the previous step 74 is compared with the edge direction θ of the edge pixel to be determined, and the angular difference between the two is determined. Then, it is determined whether or not the edge is within an angle range corresponding to an artificial straight line segment that is substantially linearly continuous. As an angle range corresponding to an artificial straight line segment, for example, 0 ° ± 2 ° or a range before and after that can be applied. When this determination is affirmed, the edge pixel to be determined and the adjacent pixel (edge pixel) satisfying the determination in step 78 may be a pixel constituting a part of an artificial linear line segment corresponding to an artifact. Therefore, the process proceeds to step 82, and information on these pixels is stored in the memory 24B as information on pixels constituting a part of the artificial straight line segment.

  In the next step 86, the adjacent pixel (edge pixel) whose information is stored in the memory 24B is selected as the next edge pixel to be determined, and the process returns to step 74. Note that the determination in the previous step 78 is affirmed if at least one adjacent pixel that satisfies the determination condition in the step 78 exists among the plurality of adjacent pixels selected in the previous step 74. When there are a plurality of adjacent pixels that satisfy the determination condition of step 78, in step 86, a single adjacent pixel is selected as a determination target from among a plurality of adjacent pixels that satisfy the determination condition, and Information on the remaining adjacent pixels among the plurality of adjacent pixels satisfying the determination condition is stored in the memory 24B as determination target candidate pixels. This determination target candidate pixel is preferentially selected as the determination target edge pixel when step 72 is executed again later.

  With the above processing, the edge pixel to be determined selected in step 72 is an edge pixel that forms an artificial straight line segment (as shown as “edge pixel to be determined” in FIG. 8). In the case of the edge pixel located on the line segment), steps 74 to 78, 82, and 86 are repeated, so that the straight line segment corresponding to the initial determination target edge pixel is traced in a certain direction. (Traced), and as shown in FIG. 8 as “connection 1” to “connection 4”, it forms the same straight line segment as the edge pixel to be initially determined, and the edge to be initially determined Edge pixels existing on the tracking direction side with respect to the pixels are sequentially extracted, and information on the extracted edge pixels is sequentially stored in the memory 24B as information on the edge pixels constituting the same straight line segment. In the extracted individual edge image There is connected as an edge pixels constituting the same straight line.

  When the position of the edge pixel to be determined reaches one end of the straight line segment, the determination in step 76 is once denied and the process returns to step 70, and then the straight line with the original edge pixel to be determined interposed therebetween. Adjacent edge pixels (edge pixels stored in the memory 24B as determination target candidate pixels) existing on the side opposite to the original tracking direction of minutes are selected as the determination target edge pixels in step 72, and steps 74 to 78, By repeating 82 and 86 again, the straight line segment corresponding to the original edge pixel to be determined is traced (traced) in the direction opposite to the original tracking direction. As shown by +3 '', it forms the same straight line segment as the original edge pixel to be judged, and exists on the opposite side of the original tracking direction with respect to the edge pixel to be initially judged. Edge pixel All the edge pixels constituting the same straight line segment are extracted in order, and the information of the extracted edge pixels is sequentially stored in the memory 24B as the information of the edge pixels constituting the same straight line segment. Will be concatenated.

  Further, when the determination in the previous step 78 is negative (when the angle difference between the edge direction θ of the adjacent pixel selected in step 74 and the edge pixel to be determined is out of the angle range corresponding to the straight line segment) Shifts to step 80, the angle difference between the edge direction θ between the adjacent pixel selected in step 74 and the edge pixel to be determined is a continuous edge, and the variation in curvature in the continuous edge is within a predetermined range with respect to the reference value. It is determined whether or not the angle is within an angle range corresponding to an artificial curved line segment that is changing within the range. As an angle range corresponding to an artificial curved line segment, for example, 5 ° ± 2 ° or a range before and after that can be applied.

  If the determination in step 80 is affirmed, the edge pixel to be determined and the adjacent pixel (edge pixel) satisfying the determination in step 80 are pixels constituting a part of an artificial curved line segment corresponding to an artifact. Since there is a high possibility, the process proceeds to step 84, and information on these pixels is stored in the memory 24B as information on pixels constituting a part of the artificial curved line segment. In the next step 86, the adjacent pixel (edge pixel) whose information is stored in the memory 24B is selected as the next edge pixel to be determined, and the process returns to step 74. The determination in step 80 is also affirmed if there is at least one adjacent pixel that satisfies the determination condition in step 80 among the plurality of adjacent pixels selected in previous step 74. In step 86, even when there are a plurality of adjacent pixels satisfying the determination condition of step 80, a single adjacent pixel is selected as a determination target from among the plurality of adjacent pixels satisfying the determination condition. For the remaining adjacent pixels among the plurality of adjacent pixels that satisfy the determination condition, information is stored in the memory 24B as determination target candidate pixels.

  If the determination-target edge pixel selected in step 72 is an edge pixel that constitutes an artificial curved line segment (edge pixel located on the artificial curved line segment) by the above processing, step By repeating 74 to 80, 84, and 86, the curved line segment corresponding to the edge pixel of the initial determination target is traced (traced) in a certain direction, and the same straight line as the edge pixel of the initial determination target Edge pixels that are included in the tracking direction with respect to the edge pixel of the original determination target are extracted in order, and the extracted edge pixel information constitutes the same curved line segment. By sequentially storing in the memory 24B as pixel information, the extracted individual edge pixels are connected as edge pixels that constitute the same curved line segment.

  When the position of the edge pixel to be determined reaches one end of the straight line segment, the determination in step 76 is once denied and the process returns to step 70, and then the curved line segment is sandwiched with the edge pixel to be determined initially. Adjacent pixel (edge pixel stored in the memory 24B as a determination target candidate pixel) existing on the opposite side to the initial tracking direction is selected as the determination target edge pixel in step 72, and steps 74 to 80, 84 are performed. , 86 are repeated again, the curved line segment corresponding to the initial determination target edge pixel is traced (traced) in the direction opposite to the initial tracking direction, and the same curved line shape as the initial determination target edge pixel is obtained. Edge pixels constituting the line segment and existing on the opposite side of the initial tracking direction with respect to the edge pixel to be initially determined are sequentially extracted, and information on the extracted edge pixel is stored in the memory 2. By sequentially storing in 4B, all edge pixels constituting the same curved line segment are connected.

  By the above-described processing, artificial straight line segments and curved line segments existing in the candidate human candidate region to be processed are searched, and artificial straight line segments and curves are searched for in the candidate human candidate region to be processed. If a line segment exists, the straight line segment or the curved line segment is extracted from the candidate person area to be processed. Steps 54 to 86 correspond to the artificial line segment detecting means according to the present invention (more specifically, the artificial line segment detecting means according to claims 2, 4 and 6). When all the edge pixels in the person candidate region to be processed are selected as determination targets, the determination in step 70 is affirmed and the process proceeds to step 88, where the artificial straight line segments extracted in steps 70 to 86 are Based on the number and size of the curved line segments, it is determined whether or not the candidate human subject area to be processed is an artifact area corresponding to an artifact. The determination in step 88 is, for example, a first condition that “the number of extracted straight line segments is equal to or greater than a first threshold value and the length of the extracted straight line segments is equal to or greater than a second threshold value” It can be performed by determining each of the second conditions that the number of curved line segments is equal to or greater than the third threshold and the length of the extracted curved line segments is equal to or greater than the fourth threshold. When at least one of the two conditions is satisfied, affirmation can be made.

  If the determination in step 88 is affirmative, an artificial straight line segment or curved line segment is clearly present in the processing target person candidate region, so that the processing target human candidate region is an artifact. It can be determined that this is a corresponding artifact area (not a person area corresponding to a person). For this reason, if the determination in step 88 is affirmed, the process proceeds to step 90, and after the process target person candidate area is excluded from the output targets, the process proceeds to step 92. If the determination in step 88 is negative, an artificial straight line segment or curved line segment does not exist in the person candidate area to be processed, or an artificial straight line segment or curve Even if there is a line segment, it is a small number or a short length, so it is determined that the person candidate area to be processed is not an artifact area corresponding to an artifact (highly likely to be a person area corresponding to a person) it can. Therefore, if the determination in step 88 is negative, step 90 is skipped and the process proceeds to step 92.

  In step 92, it is determined whether or not a person candidate region that has not been selected as a processing target in step 56 exists in the person candidate regions extracted from the processing target image in step 52. If the determination is affirmative, the process returns to step 56, and steps 56 to 92 are repeated until the determination of step 92 is negative. As a result, the processing from step 58 onward is performed for all the human candidate regions extracted in step 52, and as a result, an artificial straight line segment or curved line segment in the processing target image becomes clear. Only the person candidate area existing in the area excluding the existing area remains without being excluded. If the determination in step 92 is negative, the process proceeds to step 94. Among the person candidate areas extracted in step 52, information on the person candidate areas remaining without being excluded from the output targets in step 90 is displayed in the person area. The information is output as information of a certain area with high accuracy, and the person candidate area extraction process is terminated. Steps 88 to 92 described above correspond to the person candidate region extracting means (more specifically, the person candidate region extracting means described in claim 2) according to the present invention together with step 52 described above.

  In addition, as an output destination of the processing result of the person candidate area extraction process, for example, a part of a vehicle control system mounted on the vehicle 12 together with the object detection device 10 is configured, and the positions and distances of persons existing around the vehicle 12 A control unit that performs predetermined control (for example, control for deploying an airbag for alleviating the impact to the pedestrian when the host vehicle is predicted to collide with the pedestrian) It is done. In the present embodiment, among the candidate human regions extracted by pattern recognition, the human candidate region in which an artificial line segment is clearly present is excluded from the output target, and thus is a region corresponding to a person. A region with high accuracy can be extracted and output, and the accuracy of the predetermined control can be improved even when the processing result of the person candidate region extraction process is received and the predetermined control is performed.

  Further, in the present embodiment, a region with high accuracy that is a region corresponding to a person can be extracted from a visible image, and since it is not necessary to use an infrared image, the configuration of the object detection device 10 can be simplified. .

  In the above description, an embodiment has been described in which all the artificial straight line segments and curved line segments existing in the candidate human subject region to be processed are searched for and extracted (regardless of the direction in which the line segment extends, etc.). The invention is not limited to this, and only artificial straight line segments (vertical straight line segments) extending substantially along the vertical direction of the processing target image corresponding to the vertical direction in real space. Search and extraction may be performed as a target.

  Specifically, searching and extracting only the straight line segment in the vertical direction is, for example, a pixel in which the edge direction θ is approximately 90 ° or −90 ° with respect to the vertical direction in step 60 to step 68 in FIG. Only pixels (edge direction θ pixels that may be edge pixels constituting a vertical straight line segment) are to be determined, and edge pixels are extracted by comparing the edge intensity E with the adjacent pixels. An example of the result is shown in FIG. 9A), and only the search and extraction of the straight line segment is performed in Step 70 to Step 86 in FIG. 3 (when the determination in Step 78 is negative, the process returns to Step 70, (Steps 80 and 84 are not performed). Note that FIG. 9B shows an example of a vertical straight line segment extracted by searching a straight line segment for the edge pixel extraction result shown in FIG. 9A.

  In an image captured from the vehicle, most of the artifacts that may appear in the image are artifacts that generate a straight line segment extending substantially along the vertical direction in the image. By performing the search / extraction only on the straight line segment in the direction, it is possible to simplify the person candidate area extraction process while avoiding the decrease in the extraction accuracy of the area corresponding to the person. The above aspect corresponds to the invention described in claim 5.

[Second Embodiment]
Next, a second embodiment of the present invention will be described. Since the second embodiment has the same configuration as that of the first embodiment, the same reference numerals are given to the respective parts, description of the configuration is omitted, and the operation of the second embodiment will be described below with reference to FIG. As for the person candidate area extraction process according to the second embodiment, only the parts different from the person candidate area extraction process (FIG. 3) described in the first embodiment will be described. In addition, the person candidate area | region extraction process which concerns on 2nd Embodiment is a process to which the image processing method of Claim 10 was applied.

  In the human candidate region extraction processing according to the second embodiment, after the image data of the processing target image is read from the storage unit 24C and stored in the memory 24B in step 50, all pixels of the processing target image are stored in the next step 59. The edge strength and the edge direction are respectively calculated for. Note that the calculation of the edge strength and the edge direction itself can be performed in the same manner as Step 58 of the person candidate region extraction process (FIG. 3) according to the first embodiment.

  In the next steps 61 to 68, based on the edge strength E and the edge direction θ of each pixel calculated in step 58, the same as steps 60 to 68 of the human candidate region extraction process (FIG. 3) according to the first embodiment. Then, edge pixels are extracted from each pixel. Steps 61 to 68 of the human candidate region extraction process according to the second embodiment are performed by selecting a target pixel from all the pixels of the processing target image in step 61. The selection is different from Step 60 to Step 68 of the human candidate region extraction processing (FIG. 3) according to the first embodiment in that edge pixels are extracted from the entire region of the processing target image. As a result, when the processing target image is an image as shown in FIG. 11A as an example, the above-described Step 61 to Step 68 are used to display all of the processing target images as shown in FIG. 11B as an example. Edge pixels are extracted for the region.

  In the next Step 70 to Step 86, the individual edge pixels extracted in Step 61 to Step 68 are determined in the same manner as Step 70 to Step 86 of the person candidate region extraction process (FIG. 3) according to the first embodiment. The artificial straight line segment and the curved line segment are searched in order as the target, and in the person candidate region extraction process according to the second embodiment, the entire region of the processing target image is extracted in step 61 to step 68. Since edge pixels are extracted as targets, Step 70 to Step 86 of the person candidate region extraction processing according to the second embodiment are also determined in Step 73 from among all edge pixels in the processing target image. By selecting a pixel, the human candidate according to the first embodiment is searched for and extracted as a straight line segment and a curved line segment for the entire region of the processing target image. It differs from the step 70 to step 86 in the region extraction processing (FIG. 3).

  Thus, when an edge pixel is extracted as shown in FIG. 11B as an example from a processing target image as shown in FIG. 11A as an example, the steps 70 to 86 described above are used as examples. As shown in FIG. 12, artificial straight line segments and curved line segments are extracted for the entire region of the processing target image. It should be noted that steps 59 to 86 in the person candidate region extraction process according to the second embodiment are the artificial line segment detection means according to the present invention (more specifically, the artificial line segment detection according to claims 3, 4 and 6). Means).

  Further, in the human candidate region extraction processing according to the second embodiment, the search / extraction of the artificial straight line segment and the curved line segment for the entire region of the processing target image is completed, so that If the determination is affirmative, the process proceeds to step 96, and based on the result of the extraction of the artificial straight line segment and the curved line segment in step 70 to step 86, the first region is processed for the entire region of the processing target image. By performing the same determination as in step 88 of the human candidate region extraction process (FIG. 3) according to the embodiment, an artificial straight line segment or a curved line segment clearly exists in the entire region of the processing target image. Search for human candidate areas, excluding areas determined from all areas of the image to be processed (areas where artificial straight line segments or curved line segments exist clearly) Set as a range.

  Thereby, when the straight line segment and the curved line segment shown in FIG. 12 are extracted from the entire region of the processing target image, the artificial linear line segment shown in FIG. Or a curved line segment and its surrounding area (area shown with high density in FIG. 13) are determined as areas where an artificial straight line segment or a curved line segment is clearly present, and Of these, the range excluding the region (the range indicated by the low density in FIG. 13) is set as the search range of the person candidate region.

  In the next step 98, the person candidate area is searched for within the search range set in step 96 in the processing target image. This person candidate area search can be performed by applying pattern recognition in the same manner as in step 52 of the person candidate area extraction process (FIG. 3) according to the first embodiment. As described above, in the human candidate region extraction processing according to the second embodiment, the human candidate region is searched within the search range excluding the region where the artificial linear component or the curved line segment is clearly present. As a result, it is possible to extract a region having a high probability of being a human region as the human candidate region, and to shorten the processing time required for searching for the human candidate region. In the next step 100, the information on the person candidate area extracted in step 98 is output as information on the person area having a high probability, and the person candidate area extracting process is terminated. Steps 96 and 98 described above correspond to the person candidate area extracting means (more specifically, the person candidate area extracting means according to claim 3) according to the present invention.

  Also in the second embodiment, only the artificial straight line segment (vertical straight line segment) extending substantially along the vertical direction of the processing target image is searched and extracted, and the vertical straight line is searched. An area where the line segment is clearly present may be excluded from the search range of the person candidate area. Also in this case, it is possible to simplify the person candidate area extraction process while avoiding a decrease in the extraction accuracy of the area corresponding to the person. The above aspect also corresponds to the invention described in claim 5.

  In the above description, after the edge pixel is extracted, the determination-target edge pixel and the adjacent edge pixel are identical to each other based on the angular difference in the edge direction θ between the determination-target edge pixel and the adjacent edge pixel. It is determined whether or not the edge pixel corresponds to the curved line segment, and if the determination is affirmative, the adjacent edge pixel is repeatedly selected as the new edge pixel to be determined. Although the mode of extracting a straight line segment or a curved line segment by tracking the minute has been described, the present invention is not limited to this, and the edge pixel is extracted and then the extracted edge pixel is huffed. A linear line segment or a curved line segment may be extracted by applying (Hough) transformation. This aspect corresponds to the seventh aspect of the invention.

  In the above description, the person candidate region extraction program corresponding to the image processing program according to the present invention is stored (installed) in advance in the storage unit 24C of the image processing unit 24. However, the image processing program according to the present invention is described above. Can be provided in a form recorded on a recording medium such as a CD-ROM or a DVD-ROM.

It is a block diagram which shows schematic structure of the object detection apparatus which concerns on this embodiment. The imaging range by a visible light camera is shown, (A) is a side view, and (B) is a plan view. It is a flowchart which shows the person candidate area | region extraction process which concerns on 1st Embodiment. (A) is a search for a person candidate area by pattern recognition, (B) is an example of a learning model (pattern) used for the search, and (C) is an image diagram showing an example of an extracted person candidate area. It is an image figure which shows an example of the differential operator used for the calculation of edge strength. It is an image figure for demonstrating the calculation of the edge intensity | strength and edge direction of a calculation object pixel. It is an image figure which shows an example of the edge extraction result with respect to a person candidate area | region. It is an image figure for demonstrating tracking of an edge pixel (connection of the edge pixel which forms an artificial line segment). It is an image figure which shows an example of the process result in the case of extracting only the edge extended in the vertical direction. It is a flowchart which shows the person candidate area | region extraction process which concerns on 2nd Embodiment. (A) is an example of a processing target image, and (B) is an image diagram showing an example of an edge extraction result for the image of (A). It is an image figure which shows an example of the detection result of a linear line segment and a curve. It is an image figure which shows an example of the search area | region of a person candidate area | region set based on the detection result of a straight line segment and a curve.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Object detection apparatus 12 Vehicle 14 Visible light camera 24 Image processing part 24A CPU
24B memory 24C storage unit

Claims (11)

Artificial line segment detection means for detecting an artificial line segment existing in the visible image to be processed;
Extract a candidate human area that is estimated to be equivalent to a person that exists in an area excluding an area where an artificial line segment detected by the artificial line segment detection unit exists in the visible image to be processed. A person candidate area extracting means;
An image processing apparatus. The artificial line segment detection means searches for the person candidate area for the entire area of the visible image to be processed by the person candidate area extraction means, and then extracts the person candidate area extracted by the search. Detecting the artificial line segment as a target,
The person candidate area extraction means searches the person candidate area for the entire area of the visible image to be processed, and after the artificial line segment is detected by the artificial line segment detection means, The image processing apparatus according to claim 1, wherein a human candidate area in which the artificial line segment is detected in the area by the artificial line segment detecting unit is excluded from human candidate areas extracted from the search. The artificial line segment detection means detects the artificial line segment for the entire region of the processing target visible image,
The person candidate region extraction unit searches for the person candidate region for a region excluding the region in which the artificial line segment detected by the artificial line segment detection unit exists in the processing target visible image. The image processing apparatus according to claim 1.   The artificial line segment detecting means is configured to detect, as the artificial line segment, a straight line segment in which edges in the visible image to be processed are substantially linearly continuous, or an edge in the visible image to be processed. The image processing apparatus according to claim 1, further comprising: a curved line segment in which the variation of curvature in a continuous edge transitions within a predetermined range with respect to a reference value.   The artificial line segment detecting means, as the artificial line segment, is a straight line extending substantially along the vertical direction in the processing target visible image corresponding to the vertical direction in the real space among the straight line segments. The image processing apparatus according to claim 4, wherein only the line segment is detected.   The artificial line segment detection unit extracts an edge pixel corresponding to an edge in the image to be processed by differentiating the visible image to be processed, and then extracts a single edge pixel from the extracted edge pixels. Whether or not another edge pixel exists in the vicinity of the edge pixel to be processed and within the angle range corresponding to the linear component or the curved component with respect to the edge pixel to be processed. If the corresponding edge pixel exists, the linear line configured by connecting the edge pixels as the artificial line segment is repeated by repeating the process of selecting the corresponding edge pixel as a new processing target. The image processing apparatus according to claim 4, wherein a minute or the curved line segment is detected.   The artificial line segment detection means extracts an edge pixel corresponding to an edge in the image to be processed by differentiating the visible image to be processed, and then applies a Hough transform to the extracted edge pixel. The image processing apparatus according to claim 4, wherein the straight line segment or the curved line segment configured by connecting edge pixels is detected as the artificial line segment.   The image processing apparatus according to claim 1, wherein the visible image to be processed is a moving image obtained by imaging the periphery of the vehicle by an imaging unit mounted on the vehicle. Search for a human candidate area that is estimated to correspond to a person for the entire area of the visible image to be processed,
Detecting the artificial line segment for the human candidate area extracted in the search,
By excluding the human candidate area in which the artificial line segment is detected in the human candidate area extracted in the search from the human candidates, there is an artificial line segment in the visible image to be processed. An image processing method for extracting a candidate person area existing in an area excluding the existing area. Artificial line segments are detected for the entire region of the visible image to be processed,
By searching for a candidate human region that is estimated to correspond to a person in an area excluding an area where the artificial line segment exists in the visible image to be processed, among the visible images to be processed An image processing method for extracting a human candidate area existing in an area excluding an area where an artificial line segment exists. Computer
Artificial line segment detection means for detecting an artificial line segment existing in a visible image to be processed;
And a human candidate area estimated to be equivalent to a person existing in an area excluding an area where the artificial line segment detected by the artificial line segment detection unit is present in the visible image to be processed. An image processing program for functioning as a person candidate area extracting means for extraction.