JP2020035334A - Image processing apparatus - Google Patents

Abstract

To provide a technique that improves recognition precision of a target.SOLUTION: An image processing apparatus comprises a range setting part S310, an edge detection part S340, a distribution calculation part S370, a distribution pattern memory, and a target determination part S380. The range setting part sets as a target range a range in which presence of a target in a captured image acquired by an image acquisition part is estimated. The distribution calculation part calculates an edge distribution which is a distribution where edge degrees being numerals are arranged along a horizontal direction of the image, the edge degrees being obtained by adding the numbers of edge points calculated by the edge detection part. The distribution pattern memory memorizes at least one distribution pattern. The target determination part determines whether a target range represents a target corresponding to the distribution pattern according to respective similarities between the edge distribution calculated by the distribution calculation part and the at least one distribution pattern memorized in the distribution pattern memory.SELECTED DRAWING: Figure 4

Description

  The present disclosure relates to an image processing device.

  Patent Literature 1 discloses a technique for recognizing a target such as a pedestrian by pattern matching. According to the pattern matching, an area of a predetermined size of the image is extracted, and the extracted target area is divided into lattice cells. Then, for each of the cells divided in a grid pattern, a HOG feature value, which is a feature value obtained by converting the direction of the luminance gradient into a histogram, is calculated. Further, the HOG feature amount of each cell is compared with the HOG feature amount of a matching pattern that is stored in advance and is a pattern corresponding to a portion of a target to be recognized. If the HOG feature values match by collation, it is determined that the cell represents the target portion represented by the collation pattern. Then, it is determined that a target including a target portion corresponding to the matching pattern exists in a range including the cell in the image.

JP 2014-81826 A

  However, for example, when pattern matching using the HOG feature amount is performed for each of the target region including the human leg and the target region including the sign post, both the human leg and the sign post are positioned relative to the ground. Since there is a luminance gradient in the horizontal direction, the HOG feature amounts obtained by converting the direction of the luminance gradient into a histogram are similar.

  As a result, in the pattern matching using the HOG feature, it is not possible to distinguish a target region including a human leg from a target region including a sign post. There is a possibility that the recognition accuracy may be reduced by recognizing that the target region includes a pedestrian.

  One aspect of the present disclosure provides a technique for improving recognition accuracy of a target.

  One embodiment of the present disclosure is an image processing apparatus, and includes an image acquisition unit (S210), a range setting unit (S110, S310), an edge detection unit (S340), a distribution calculation unit (S370), and a distribution pattern. It includes a memory (64) and a target determination unit (S380). The image acquisition unit is configured to acquire a captured image (100) acquired by the imaging device (10) mounted on the vehicle. The range setting unit is configured to set, as a target range, a range in which a target in the captured image acquired by the image acquisition unit is assumed to be present. The edge detection unit is configured to detect, at least in a part of the target range set by the range setting unit, as an edge point that is a point representing a contour portion of the target along the horizontal direction in the image. . The distribution calculating unit calculates the edge degree, which is a numerical value obtained by adding the number of edge points calculated by the edge detecting unit in the vertical direction in the image, and distributes the edge degree along the horizontal direction in the image. It is configured to calculate a distribution. The distribution pattern memory is configured to store at least one distribution pattern that is a representative pattern of an edge distribution previously associated with each type of target. The target determination unit determines a target range corresponding to the distribution pattern according to the similarity between the edge distribution calculated by the distribution calculation unit and each of the at least one distribution pattern stored in the distribution pattern memory. It is configured so as to determine whether or not the information is represented.

  In the pattern matching using the HOG feature amount, the target area to be matched is divided into a plurality of cells, and the matching result is summed up for each of the divided cells, so that the entire target area matches the target target. It is determined whether or not. Therefore, if a part of each divided cell unit has a high degree of coincidence with the target to be compared, the result of the cell with the high degree of coincidence will affect the result of the entire target area. May be determined to match.

  According to the configuration described above, the distribution calculation unit calculates an edge distribution, which is a distribution in which a numerical value obtained by adding in the vertical direction in an image is arranged along the horizontal direction in the image. Thereby, at least in the vertical direction, the edge distribution in a wider range is calculated as compared with the case where the determination is made on a cell-by-cell basis, and the calculated edge distribution is compared with a predetermined distribution pattern, so that the target is determined. recognize.

Since the comparison can be made based on the detection results in a wider range than when the determination is made on a cell-by-cell basis, even when the degree of coincidence of a part corresponding to a cell is high, the influence on the whole can be suppressed.
As a result, the accuracy of recognizing the target represented by the target range can be improved.

  Note that the reference numerals in parentheses described in this column and in the claims indicate a correspondence relationship with specific means described in the embodiment described below as one aspect, and the technical scope of the present disclosure will be described. It is not limited.

It is a block diagram showing the composition of an image recognition system. It is a flowchart of a pedestrian recognition process. It is a flowchart of an estimation process. It is a flowchart of a recognition process. FIG. 3 is a diagram illustrating an example of a captured video and an estimated range and a range of interest in the captured image. It is a figure showing the image of the 1st attention range. FIG. 5 is a diagram illustrating an edge image of a first attention range. FIG. 6 is a diagram illustrating an edge distribution of a first attention range. It is a figure showing the image of the 3rd attention range. It is a figure showing the edge image of the 3rd attention range. It is a figure showing the edge distribution of the 3rd attention range.

Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.
[1. Constitution]
The image recognition system 1 shown in FIG. 1 is mounted on a vehicle and includes a camera module 10, a control device 20, a brake actuator 30, a display device 40, a speaker 50, and an image processing device 60. Hereinafter, a vehicle on which the image recognition system 1 is mounted is referred to as a host vehicle.

  The camera module 10 is, for example, a camera module of a CCD camera that outputs a color image. The camera module 10 is provided, for example, in the cabin of the host vehicle to capture an image of the front of the host vehicle.

  Further, the camera module 10 is not limited to the one that images the front of the own vehicle, but may be any one that images the surroundings of the own vehicle, and the position where the camera module 10 is installed may not be in the vehicle interior.

The brake actuator 30 is an actuator that adjusts the braking force of the own vehicle.
The display device 40 is a device that is installed in the cabin of the host vehicle and displays an image or the like on a display screen of the display device 40.

The speaker 50 is a device that is installed in the cabin of the vehicle and outputs sound.
The control device 20 controls the brake actuator 30, the display device 40, or the speaker 50 according to the result of the pedestrian recognition output from the image processing device 60.

  The control of the brake actuator 30 refers to brake control for increasing the braking force of the own vehicle, for example, when it is determined that a pedestrian exists within a predetermined range based on the position of the own vehicle.

  The control of the display device 40 is, for example, when a captured image around the own vehicle is displayed on the display device 40 and a pedestrian range that is a range where it is determined that a pedestrian exists in the captured image exists. Control for highlighting and displaying the range. In addition, the control of the display device 40 is not limited to the control of performing the highlighted display, and may be any control as long as it performs a display that calls attention to the driver of the own vehicle. Further, when there is an object range that is a range determined to include an object present in the captured image, the range may be emphasized and displayed. In this case, the display of the pedestrian range and the display of the object range may be displayed in different display modes.

  The control of the speaker 50 means, for example, a control that emits an alarm sound and notifies the driver of the own vehicle when it is determined that a pedestrian is present around the own vehicle captured by the camera module 10.

  The control device 20 may perform only one of the control of the brake actuator 30, the control of the display device 40, and the control of the speaker 50, or may perform a plurality of controls.

  The image processing device 60 includes a microcomputer having a CPU 61 and a semiconductor memory such as a RAM or a ROM (hereinafter, a processing memory 62), a matching pattern memory 63, and a distribution pattern memory 64. Each function of the image processing device 60 is realized by the CPU 61 executing a program stored in a non-transitional substantial recording medium. In this example, the processing memory 62 corresponds to a non-transitional substantial recording medium storing a program. When this program is executed, a method corresponding to the program is executed. Note that the image processing device 60 may include one microcomputer, or may include a plurality of microcomputers.

  The technique for realizing each function of the image processing apparatus 60 is not limited to software, and some or all of the functions may be realized using one or a plurality of hardware. For example, when the above function is implemented by an electronic circuit that is hardware, the electronic circuit may be implemented by a digital circuit, an analog circuit, or a combination thereof.

  Note that the image processing device 60 receives power supply from the own vehicle while the ignition switch of the own vehicle is on. Further, the image processing device 60 performs a process of outputting a detection result to the control device 20. The detection result here includes, for example, whether or not a pedestrian is present, and may include information on the position of the detected pedestrian.

  The matching pattern memory 63 is a memory that stores a matching pattern. The matching pattern is a rectangular pattern corresponding to a part of the pedestrian, and represents a predetermined HOG feature amount. The HOG feature amount is a histogram representing the gradient strength of the luminance gradient for each gradient direction of the luminance gradient.

  The distribution pattern memory 64 is a memory that stores a distribution pattern. The distribution pattern is a representative pattern of the edge distribution of at least a part of the target preset for each type of target. The edge distribution is a distribution in which the edge degrees are arranged along the horizontal direction in the captured image. The edge degree is a value calculated by adding the number of edge points along the vertical direction in the captured image.

  That is, the distribution pattern is a representative pattern of the edge distribution stored in advance for each type of target in order to compare with a part of the edge distribution of the target detected in the captured image.

  Further, the distribution pattern includes an object pattern. The object pattern is a distribution pattern representing at least a part of an object that is a target other than a pedestrian. The object pattern includes, for example, a pole pattern which is an object pattern representing a part of a pole which is a pillar-shaped body erected on a road.

Note that the camera module 10 corresponds to an imaging device.
[2. processing]
Next, the pedestrian recognition process executed by the CPU 61 will be described with reference to the flowchart in FIG. The pedestrian recognition device is repeatedly executed while the power of the image processing device 60 is on.

  In S110, the CPU 61 performs an estimation process. Although the details of the estimation processing here will be described later, a range in which a pedestrian is estimated to be a pedestrian in a target range which is a range representing a target existing in an image captured by the camera module 10 is referred to as an estimation range. Means a process of extracting as When a plurality of estimation ranges exist in the captured image, a plurality of estimation ranges existing in the captured image are extracted.

In S120, the CPU 61 determines whether or not an estimated range exists in the captured image, that is, whether or not one or more estimated ranges have been extracted by the estimation process, in the estimation process in S110.
If it is determined in S120 that an estimation range exists, that is, if it is determined that one or more estimation ranges have been extracted, the CPU 61 shifts the processing to S130.

On the other hand, if it is determined in S120 that no estimation range exists, that is, if no estimation range has been extracted, the CPU 61 ends the pedestrian recognition process.
In S130, the CPU 61 performs a recognition process. Although the details of the recognition processing here will be described later, the recognition processing indicates that each of the estimated ranges extracted in S110 is a pedestrian range that represents a pedestrian, or represents a region other than a pedestrian. This is a process of determining whether the object range is an object range and outputting a pedestrian range and an object range.

S110 corresponds to the processing as the range setting unit.
<Estimation processing>
Next, details of the estimation processing executed by the CPU 61 in S110 of the pedestrian recognition processing will be described with reference to FIG.

In S210, the CPU 61 obtains an image captured by the camera module 10.
In S220, the CPU 61 divides the captured image acquired in S210 into rectangular cells having a preset number of pixels, and calculates a HOG feature amount for each cell. Here, the number of pixels of the cell to be set is set to, for example, a range of 16 pixels × 16 pixels.

  Note that the HOG feature amount is obtained as follows. That is, assuming that the abscissa of the cell is x, the ordinate is y, and the luminance at the coordinates (x, y) is L (x, y), first, a gradient intensity m representing the magnitude of the luminance gradient of each cell. (X, y) and a gradient direction θ (x, y) representing the direction of the luminance gradient of each divided region are calculated.

  Then, a brightness gradient histogram is created by dividing 0 ° to 180 ° in the m direction and summing the gradient intensities of the divided regions in which the gradient direction can be considered to be the same direction as the gradient intensity in that direction. I do. The m-dimensional value represented by the luminance histogram is the HOG feature amount.

In S230, the CPU 61 determines whether or not an uncollated pattern, which is a collation pattern not selected in S240 described later, exists in the captured image acquired in S210.
If the CPU 61 determines that an unmatched pattern exists, the process proceeds to S240.

  By the processing from S240 to S270, the area in the captured image is collated with each collation pattern stored in the collation pattern memory 63, and an estimated range representing a pedestrian present in the captured image is extracted.

In S240, the CPU 61 selects one matching pattern from the unmatched patterns.
In S250, the CPU 61 determines whether or not there is an unselected search size (hereinafter, referred to as an unselected size) that has not been selected in S260 described later among a plurality of search sizes set for the search area. Here, the search area is a rectangular area represented by n cells × n cells where n is a natural number. As the size of the search area, a plurality of types are set according to the size of the target to be searched which is reflected in the captured image.

  If the CPU 61 determines in step S250 that there is no unselected size, the process returns to step S230, and performs the subsequent processing. That is, the process from S260 to S270 is repeatedly performed while changing the matching pattern.

On the other hand, if the CPU 61 determines in step S250 that there is an unselected size, the process proceeds to step S260.
In S260, the CPU 61 selects any one of the unselected sizes. Hereinafter, the search size selected in S260 is also referred to as a selected size.

  In S270, the CPU 61 extracts an estimation range. The extraction of the estimation range is specifically performed as follows. That is, the entire captured image is scanned while shifting the search area of the selected size by one cell. At this time, the sum of the HOG features of all the cells included in the search area is used as the HOG feature of the search area, and the HOG feature of the search area is compared with the HOG feature of the matching pattern to determine the similarity. calculate. The position of the search area having a high degree of similarity is extracted as an estimated range in which a pedestrian represented by the matching pattern may exist.

When the scanning of the entire captured image is completed, the CPU 61 shifts the processing to S250. That is, the CPU 61 changes the selected size and repeatedly executes the same processing.
If it is determined in S230 that there is no unmatched pattern, the estimation process ends.

Note that S210 corresponds to the processing as the image acquisition unit.
<Recognition processing>
Next, the details of the recognition processing executed by the CPU 61 in S130 of the pedestrian recognition processing will be described with reference to FIG.

  In S310, the CPU 61 acquires all the estimation ranges extracted in S270 of the estimation processing. Here, the estimation range is the range in which the pedestrian is estimated in the target range, as described above.

  In the processing from S320 to S400, it is determined whether or not the attention range, which is at least a part of each of the estimation ranges, matches the object pattern, so that the estimation range represents a pedestrian or a target other than the pedestrian. It is determined whether or not the object is represented.

In S320, the CPU 61 determines whether or not there is an unprocessed range, which is an estimation range not selected in S330 described later, among all the estimation ranges.
When determining that the unprocessed range exists, the CPU 61 shifts the processing to S330.

In S330, the CPU 61 selects one estimated range from the unprocessed ranges and sets it as the selected range.
In S340, the CPU 61 detects, as an edge point, a point in which the absolute value of the luminance difference between the pixels adjacent in the horizontal direction in the range selected in S330 is equal to or greater than a predetermined luminance threshold value. . The luminance threshold is set to a value corresponding to the luminance difference of the outline of the target.

In S350, the CPU 61 determines whether or not there is an uncalled pattern that has not been called in S360 described later among the distribution patterns.
In S360, the CPU 61 calls one of the object patterns stored in the distribution pattern memory 64. The object pattern called in S360 is called from among the object patterns that have not been called yet for the estimation range selected in S330.

  In S370, the CPU 61 detects in S340 each of the boundaries of the pixel rows arranged in the vertical direction in the predetermined range of interest, which is at least a part of the selection range selected in S330. The number of edge points thus calculated is calculated as an edge degree. Then, a distribution obtained by arranging the calculated edge degrees in the horizontal direction corresponding to the position of the boundary of the pixel row is derived as an edge distribution.

  As the attention range, a range corresponding to the object pattern called in S360 in the selection range is set. Here, the range of interest set corresponding to the object pattern is set in advance so that a part of the target represented by the object pattern is easily recognized as a part of the pedestrian.

In S380, the CPU 61 determines whether or not the calling pattern, which is the object pattern called in S360, matches the edge distribution calculated in S370.
Here, the coincidence between the call pattern and the edge distribution is not limited to a perfect coincidence, but may be determined to coincide if the similarity is equal to or greater than a predetermined threshold. The similarity may be determined by comparing the characteristics of the distribution of the call pattern with the characteristics of the distribution of the edge distribution. Here, the characteristics of the distribution include, for example, the number of peaks, the extent of the distribution, the positions of the peaks, and the like. Here, the peak refers to, for example, a portion where the edge degree is maximum or a portion where the edge degree is maximum. For example, the half width of the peak may be used as the spread of the distribution. For example, the distribution similarity may be regarded as a match if the edge distribution and the distribution pattern have the same number of edge degree peaks and the spread of the edge degree distribution is similar.

If the CPU 61 determines in step S380 that the edge distribution matches the calling pattern, the process proceeds to step S390.
In S390, the CPU 61 recognizes that the selection range selected in S330 is the object range representing the object, returns to S320, and executes the subsequent processing.

On the other hand, if the CPU 61 determines in step S380 that the edge distribution does not match the calling pattern, the process proceeds to step S350.
Returning to S350, if it is determined that there is no uncalled pattern, the process proceeds to S400.

In S400, the CPU 61 recognizes that the selection range selected in S330 is a pedestrian range representing a pedestrian, returns to S320, and executes the subsequent processing.
As described above, by the processes from S350 to S400, the attention range which is a part of the selection range selected in S330 is compared with the object pattern which is a pattern representing a part of the object, and the selection range represents the object. It is determined whether there is.

If it is determined in S320 that there is no unprocessed area, the process proceeds to S410.
In S410, the CPU 61 outputs the pedestrian range and the object range, and ends the estimation processing.

By the above-described recognition processing, the CPU 61 outputs an estimated range other than the object range among the estimated ranges as the pedestrian range.
Note that S310 corresponds to processing as a range setting unit, S360 corresponds to processing as a range extraction unit, S340 and S370 correspond to processing as an edge detection unit, and S370 corresponds to processing as a distribution calculation unit. I do. In addition, S380 corresponds to processing as a target determination unit, and S400 corresponds to processing as a pedestrian determination unit.

[3. Operation example]
An example of image processing when the pedestrian recognition processing is performed will be described with reference to the drawings.
As shown in FIG. 5, for example, as a captured image, a first pedestrian 101 who is a stationary pedestrian, a second pedestrian 102 walking across the front of the own vehicle, and a pole 103 are included in the image. The captured image 100 included is taken as an example. Note that the captured image 100 is not limited to the image shown in FIG. 5 and may be various images obtained by capturing the surroundings of the own vehicle.

  When the estimation process is performed on the captured image 100 in S110 of the pedestrian recognition process, for example, the first estimation range 101a, the second estimation range 102a, and the third estimation range 103a in FIG. 5 are extracted as the estimation ranges.

Here, when extracting the estimation range using the HOG feature amount, not only the first pedestrian 101 and the second pedestrian 102 but also the pole 103 may be extracted as the estimation range.
The pole shape of the pole included in the attention range, which is a part of the estimation range, has a shape that is longitudinal in the vertical direction with respect to the image, similar to the human leg portion, and has brightness in the horizontal direction with respect to the image. It has a gradient. For this reason, the pole-like shape of the pole may be recognized as a human leg.

That is, when a pedestrian is detected by comparing the HOG feature amounts, not only a pedestrian but also a three-dimensional object installed on a road such as a pole may be extracted.
Next, in S130, a recognition process is performed for each estimation range. In the recognition process, when the first estimation range 101a is selected in S330, the first estimation range 101a is collated with the object pattern.

In S360, the calling of the object pattern is executed. In S340, an edge point is detected for the selected range.
The detection of an edge point is performed as follows. That is, an edge image as shown in FIG. 7 is created for the selection range as shown in FIG. The edge image referred to here is an image representing an edge point at which the absolute value of the horizontal luminance difference is equal to or greater than a predetermined luminance threshold. In FIG. 7, a plurality of white points Pe1 represent edge points.

  By adding the number of edge points along the vertical direction using the edge image shown in FIG. 7, the edge degree for each boundary of the pixel row in the selection range is calculated. By arranging the calculated number of edge points for each pixel row boundary in the horizontal direction with respect to the image, an edge distribution representing the edge degree at each pixel row boundary of the attention range as shown in FIG. 8 is derived. Is done.

  In S380, it is determined whether or not the edge distribution derived in S370 matches an object pattern representing a part of a pole which is a call pattern. As a feature of the object pattern of the pole, for example, if it is stored that the peak of the edge degree of the edge distribution is two, the edge distribution derived in S370 has three or more peaks. It is determined that it does not match the object pattern representing the pole.

As described above, when it is determined that the first attention range 101b which is a part of the first estimation range 101a does not match any of the object patterns, it is recognized as a range representing a pedestrian.
Similar to the case of the first estimation range 101a, the second estimation range 102a is compared with the object pattern in the second attention range 102b, and it is determined whether the second estimation range 102a is a range representing a pedestrian.

  On the other hand, when the third estimation range 103a is selected in S330 in the recognition processing, the matching with the object pattern is performed in the third attention range 103b as in the case of the first estimation range 101a.

  Here, by detecting the edge point in S340, a point Pe2 representing an edge point as shown in FIG. 10 is formed with respect to the third attention range 103b as shown in FIG. 9 which constitutes a part of the third estimation range 103a. The obtained edge image is obtained. When the edge distribution is obtained using the edge image shown in FIG. 10 as in the case of the first range of interest 101b, an edge distribution representing the edge degree for each boundary of the pixel row is derived as shown in FIG. .

  In S380, it is determined whether or not the edge distribution derived in S370 matches an object pattern representing a part of a pole which is a call pattern. Similarly, if the feature of the object pattern of the pole is stored that the peak of the edge degree of the edge distribution is two, the edge distribution derived in S370 has two peaks. It is determined that the feature matches the feature of the object pattern representing the pole.

As a result, the third estimation range 103a is determined to match the object pattern representing the pole, and is recognized as the object range representing the object.
If it is determined that the recognition processing has been performed on the first estimation range 101a, the second estimation range 102a, and the third estimation range 103a that are all the estimation ranges in the captured image 100, the S first estimation range is determined in S410. A recognition result indicating that the range 101a and the second estimation range 102a are pedestrian ranges and the third estimation range 103a is an object range is output.

[4. effect]
According to the embodiment described in detail above, the following effects can be obtained.
(1) According to the above embodiment, an edge distribution which is a distribution of the edge degree indicating the position of the contour of the target in the cell, and a distribution which is a distribution of the edge degree indicating the position of the contour of the predetermined target in the cell. The pattern is compared. For this reason, the target is determined using the position of the contour on the target as the determination material, as compared with the case where the HOG feature amounts of the cell and the matching pattern are compared.

As a result, the accuracy of recognizing the target represented by the target range can be improved.
For example, according to the above-described embodiment, each target can be identified even if it is a target such as a pedestrian and a pole that cannot be identified by pattern matching using the HOG feature amount.

[5. Other Embodiments]
Although the embodiments of the present disclosure have been described above, the present disclosure is not limited to the above embodiments, and can be implemented with various modifications.

  (1) In the estimation processing, whether or not a pedestrian is estimated is based on whether or not the HOG feature quantities match. However, in the estimation processing, the feature quantity used for estimating whether or not a pedestrian is HOG is used. The feature amount is not limited to the feature amount, and may be a feature amount corresponding to a feature of each cell representing a pedestrian. Although the estimation process is performed by the method shown in FIG. 3, the present invention is not limited to such a method. If an estimation range that is estimated to represent a pedestrian is extracted, , And other well-known methods such as template matching.

  (2) In the above embodiment, the recognition range is performed after the estimation range is extracted by the estimation process. However, the extraction of the estimation range may be omitted. In this case, a captured image may be acquired, and an arbitrary range among the acquired imaging ranges may be sequentially extracted as an estimated range, and recognition processing may be performed.

  (3) In the above embodiment, a distribution pattern of a part of a pole installed on a road is given as an example of the object pattern. However, an object that is easily recognized as a pedestrian is not limited to a part of the pole, and may be a tree. A pattern representing a part of another object, such as the outline shape of the object, may be stored.

  (4) Also, as an example of the object pattern, a part of an object recognized as a pedestrian's leg is given as an example. However, any part that can be easily recognized as a pedestrian is not limited to the leg, but may be a head or the like. May be included as an object pattern. Further, as the distribution pattern to be compared with the attention range, a distribution pattern corresponding to the contour shape of the head may be set. Specifically, for example, a distribution pattern in which the edge degree at the center is high and the edge degree at the periphery is low along the horizontal direction may be set.

  (5) In the above embodiment, the edge degree is obtained by calculating the number of edge points for each boundary of the pixel row arranged in the vertical direction in the selection range. The edge degree is obtained by such a method. The edge degree is not limited to the calculated one, but may be calculated by, for example, multiplying the luminance difference between the edge points for each edge point and adding them. The calculation of the edge degree may be performed by adding a value obtained by using a Sobel filter to the luminance value of the image for each boundary of the pixel row arranged in the vertical direction.

  (6) In the above-described embodiment, in the pedestrian recognition processing, if the estimated range does not exist in the captured image, the pedestrian recognition processing ends, but the present invention is not limited to such processing. For example, a process of notifying that a pedestrian does not exist in the captured image may be performed.

  (7) Furthermore, although the object pattern is stored in the distribution pattern memory 64, the object pattern stored in the distribution pattern memory 64 may be updated. For example, a part of an object that is recognized as a pedestrian in the estimation processing but is determined not to be a pedestrian by the recognition processing is selected, and a part of the selected object is added to the distribution pattern memory 64 as an object pattern. You may have the structure which performs.

  (8) Further, a sensor or a sonar for detecting a target may be provided. Further, the range of the target detected by the sensor or the sonar may be extracted as the estimated range, and the recognition processing may be executed.

  (9) A sensor may be provided for detecting the direction in which the vehicle has moved and the amount of movement of the vehicle. The moving amount of the target in the image may be calculated from the moving direction and the moving amount of the own vehicle detected by the sensor, and the target moving with the movement may be detected.

  (10) Further, the camera module may be, for example, a camera module having a wide-angle lens. In this case, the image processing device may include a conversion unit that converts a wide-angle image into a normal image when acquiring an image. Note that a normal image is not an image captured by a wide-angle lens or the like, but an image similar to that when viewed with normal human eyes.

(11) In the above embodiment, the distribution pattern includes the object pattern representing the object, but may further store a pedestrian pattern representing the pedestrian.
(12) Also, in the above embodiment, in the recognition process, object patterns are sequentially called, and when there is no longer an object pattern that has not been called, it is recognized that the estimated range represents a pedestrian. It is not limited to such steps. For example, on the contrary, the pedestrian patterns may be sequentially called, and when there is no pedestrian pattern that has not been called, the estimation range may be recognized as representing the object.

  (14) In addition, both the object pattern and the pedestrian pattern may be called, and among the called patterns, a pattern having a high degree of coincidence may be recognized as a part of the target represented by the selection range.

  (15) In the above embodiment, the recognition processing is used for pedestrian recognition, but is not limited to such processing. For example, it may be used for recognition of a target other than a pedestrian. Further, the estimation process does not have to be performed. When the estimation process is not performed, a predetermined range in the captured image may be sequentially extracted as the estimation range or the attention range.

  (16) In the above embodiment, the pedestrian range and the object range are output in the recognition processing. However, the recognition result need not include both the pedestrian range and the object range. For example, the pedestrian range may be output without outputting the object range as a recognition result.

  (17) A plurality of functions of one component in the above embodiment may be realized by a plurality of components, or one function of one component may be realized by a plurality of components. . Also, a plurality of functions of a plurality of components may be realized by one component, or one function realized by a plurality of components may be realized by one component. Further, a part of the configuration of the above embodiment may be omitted. Further, at least a part of the configuration of the above-described embodiment may be added to or replaced with the configuration of another above-described embodiment. Note that all aspects included in the technical idea specified by the language described in the claims are embodiments of the present disclosure.

  (18) In addition to the above-described image processing apparatus, a system including the image processing apparatus as a component, a program for causing a computer to function as the image processing apparatus, and a non-transitional actual recording such as a semiconductor memory storing the program. The present disclosure can be realized in various forms such as a medium and an image processing method.

  DESCRIPTION OF SYMBOLS 1 ... Image recognition system, 10 ... Camera module, 20 ... Control device, 30 ... Brake actuator, 40 ... Display device, 50 ... Speaker, 60 ... Image processing device, 61 ... CPU, 62 ... Processing memory, 63 ... Matching pattern memory 64, distribution pattern memory, 100, captured image, 101, first pedestrian, 101a, first estimation range, 101b, first attention range, 102, second pedestrian, 102a, second estimation range, 102b, second Second attention range, 103 ... pole, 103a ... third estimated range, 103b ... third attention range.

Claims (4)

An image acquisition unit (S210) configured to acquire a captured image (100) acquired by an imaging device (10) mounted on a vehicle;
A range setting unit (S110, S310) configured to set, as a target range, a range in which the target in the captured image acquired by the image acquisition unit is assumed to be present;
An edge detection unit configured to detect as an edge point that is a point representing a contour portion of the target along a horizontal direction in the image in at least a part of the target range set by the range setting unit. (S340),
Calculate an edge distribution, which is a distribution obtained by adding the number of the edge points calculated by the edge detection unit in the vertical direction of the image in a vertical direction in the image, and a distribution of the edge degrees arranged along the horizontal direction of the image. A distribution calculation unit (S370) configured to perform
A distribution pattern memory (64) configured to store at least one distribution pattern that is a representative pattern of the edge distribution previously associated with at least a part of the target for each type of target;
The target range represents a target corresponding to the distribution pattern according to the similarity between the edge distribution calculated by the distribution calculation unit and each of the at least one distribution pattern stored in the distribution pattern memory. A target determining unit (S380) configured to determine whether the object is a target or not;
An image processing apparatus comprising: The image processing device according to claim 1,
The image processing device, wherein the distribution pattern memory stores the distribution pattern so as to include an object pattern that is a pattern representing an object that is a target other than a pedestrian. The image processing device according to claim 2,
The range setting unit is configured to set the target range estimated to represent a pedestrian as an estimated range,
A pedestrian determination unit configured to determine, by the target determination unit, an estimation range other than the estimation range determined to represent the object in the estimation as a range representing a pedestrian; An image processing apparatus further comprising (S130, S400). The image processing apparatus according to claim 2 or 3, wherein:
The image processing device, wherein the object pattern includes a pole pattern that is a pattern representing a pole that is a columnar body installed on a road.