JP7103202B2 - Image recognition device - Google Patents

Description

The present invention relates to an image recognition device.

Conventionally, in order to carry out automatic driving, an image recognition device that recognizes a lane from an image has been proposed (for example, Patent Document 1). Among the image recognition devices, an image recognition device that calculates the lane reliability of the candidate lane based on the pixel reliability calculated from the captured image and selects the candidate lane based on the lane reliability has been proposed. (Patent Document 1).

Special Table 2017-533482

By the way, when the pixel reliability is calculated from a plurality of captured images for the same object, the high and low pixel reliability may vary even for the same object depending on the type of the object and the shooting condition. When there are variations in pixel reliability, there is a problem that image recognition is difficult to perform properly and recognition accuracy is lowered.

The present invention has been made in view of the above circumstances, and a main object thereof is to provide an image recognition device capable of improving the accuracy of image recognition.

Means for solving the above problems are in an image recognition device that recognizes the course of the own vehicle based on the captured image, in an image acquisition unit that acquires a captured image from the photographing device at predetermined intervals, and in the acquired captured image. , An extraction unit that extracts the entropy for each pixel, a reliability calculation unit that calculates and outputs the pixel reliability for each pixel based on the entropy extracted by the extraction unit, and the reliability calculation unit outputs the image. A course recognition unit that recognizes a course based on the pixel reliability is provided, and the reliability calculation unit selects the calculated pixel reliability by referring to a time series of pixel reliability in a predetermined period. Outputs the selected pixel reliability.

Pixel reliability calculated by referring to the time series of pixel reliability even when a captured image whose shooting target is not clear or identification is difficult due to the surrounding conditions or the convenience of the shooting device. Select the degree. Therefore, the calculation accuracy of the pixel reliability can be improved, and the accuracy of image recognition can be improved.

A block diagram showing a driving support system. The schematic diagram which shows the mode of classification. The schematic diagram which shows the entropy. The schematic diagram which shows the occupied grid diagram. A flowchart showing an image recognition process.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. In each of the following embodiments, the same or equal parts are designated by the same reference numerals in the drawings.

The driving support system to which the image recognition device according to the present embodiment is applied will be described with reference to the drawings. The driving support system 10 is provided in the vehicle and executes driving support.

As shown in FIG. 1, the driving support system 10 is communicably connected to the monocular camera 21a as a photographing device. The monocular camera 21a is a camera in which an image sensor such as a CCD or COMS is used. The monocular camera 21a is arranged near the upper end of the windshield of the vehicle, for example, and captures (imaging) the surrounding environment including the road in front of the vehicle. The captured image captured by the monocular camera 21a is output to the image recognition device 20. Although the monocular camera 21a is provided, a plurality of cameras (compound eye cameras) may be provided.

The image recognition device 20 is a computer provided with a CPU, a ROM (memory), a RAM, an input / output interface, and the like. The image recognition device 20 has various functions. For example, the image recognition device 20 includes a function as an image acquisition unit 21, a function as a lane marking detection unit 22, and a function as an object detection unit 23. Further, the image recognition device 20 has a function as a classification unit 24, a function as an extraction unit 25, a function as a reliability calculation unit 26, a function as a feedback unit 27, and a function as a course recognition unit 28. , Equipped with. Various functions are realized by executing the program stored in the storage memory included in the image recognition device 20. The image recognition device 20 is configured to recognize the course on which the own vehicle can travel based on the captured image and output the recognition result. The various functions may be realized by electronic circuits that are hardware, or at least a part of them may be realized by software, that is, processing executed on a computer. Various functions of these image recognition devices 20 will be described later.

The image recognition device 20 is connected to the vehicle control device 30. The vehicle control device 30 is a computer (ECU) equipped with a CPU, ROM, RAM, an input / output interface, and the like, and has a function as a vehicle control unit 31.

The vehicle control unit 31 controls the steering wheel 41 (steering angle), accelerator 42 (acceleration amount), and brake 43 (braking amount) of the own vehicle, and controls the own vehicle based on the course recognized by the image recognition device 20. It is configured to run. More specifically, the vehicle control unit 31 maintains the lateral position of the own vehicle in the vehicle width direction at a predetermined position (for example, the center) in the course (the course in which the own vehicle can travel) recognized by the image recognition device 20. In this state, the traveling direction and speed of the own vehicle are controlled so that the own vehicle travels at a predetermined speed.

The vehicle control unit 31 may be provided in the image recognition device 20, or conversely, the vehicle control device 30 may include all or a part of various functions of the image recognition device 20. .. Further, the vehicle control device 30 and the image recognition device 20 may be integrated.

Next, various functions of the image recognition device 20 will be described.

The image acquisition unit 21 acquires a captured image taken by the monocular camera 21a. The image acquisition unit 21 acquires captured images at predetermined intervals (for example, 100 ms).

The lane marking detection unit 22 detects a lane marking (for example, a white line) that divides the lane in which the own vehicle travels (hereinafter referred to as the own lane) from the captured image acquired by the image acquisition unit 21. Specifically, the pavement detection unit 22 extracts a change point of the contrast (edge intensity) between the lane dividing lane and the road surface as an edge candidate point based on the brightness of the captured image. Then, the candidate line of the lane marking is extracted from the series of the extracted edge candidate points. More specifically, the captured image is continuously processed at a predetermined sampling cycle, and a plurality of points whose brightness changes rapidly in the horizontal direction of the image are extracted as edge candidate points. Then, the extracted plurality of edge candidate points are subjected to a Hough transform to acquire a series of edge candidate points, and a plurality of candidate lines having the acquired series of edge candidate points as left and right contours are extracted.

Then, the lane marking detection unit 22 calculates the degree to which each of the plurality of candidate lines has a feature as a lane dividing lane (white line) at each edge candidate point, and the degree to which the feature is provided. Is detected as the lane marking that divides the lane. Of the detected lane markings, the lane marking detection unit 22 recognizes the left and right lane markings that are close to the own vehicle and are arranged so as to include the own vehicle as lane markings (white lines) that partition the own lane. .. The method of recognizing the lane marking is not limited to this, and may be arbitrarily changed. Further, the lane marking line is not limited to the white line, and may be a line indicating the boundary of the road (road surface) (for example, a curbstone, a guardrail, etc.).

The object detection unit 23 determines the type of the object existing in front of the own vehicle based on the acquired captured image and the dictionary information for object identification prepared in advance. The dictionary information for object identification is individually prepared according to the type of an object such as an automobile, a motorcycle, a pedestrian, or an obstacle on the road, and is stored in advance in a memory. As the dictionary information of the automobile, at least the dictionary information of the front pattern and the rear pattern is prepared. Furthermore, it is preferable that dictionary information is prepared for each of a plurality of vehicle types such as a large vehicle, an ordinary vehicle, a light vehicle, a motorcycle, a light vehicle including a bicycle, etc. as a pattern of the front or rear of the vehicle. ..

The object detection unit 23 determines the type of the object by collating the captured image with the dictionary information by pattern matching. Further, the object detection unit 23 acquires the position information (including the width of the object) of the object in the traveling direction of the own vehicle and the lateral direction substantially orthogonal to the traveling direction of the own vehicle based on the captured image and the dictionary information. The object recognition method is not limited to this, and may be arbitrarily changed. For example, a radar device may be provided to perform object recognition based on the detection information of the radar device.

The classification unit 24 carries out Semantic Segmentaion. That is, the classification unit 24 analyzes each pixel of the captured image based on its meaning (information on peripheral pixels) using a deep neural network (DNN), and classifies (categorizes) the image. The classification is also referred to as assigning a label to each pixel.

By classifying, specifically, white line (partition line), road surface (driving space), other vehicles, own vehicle, pedestrian, structure (building, etc.), tree, sky, etc. for each pixel Classes are associated. As a result, as shown in FIG. 2, the captured images are classified by class. FIG. 2 (a) is an original photographed image, and FIG. 2 (b) is a diagram showing the photographed images classified into classes. In FIG. 2B, the difference in class is illustrated by making the display color different.

The extraction unit 25 extracts the entropy of each pixel. Entropy is an average amount of information, and is a measure of information disorder, ambiguity, and uncertainty. The higher the entropy, the higher the uncertainty. The entropy calculation method is shown in the following mathematical formula (1), and the uncertainty of the image is shown in the mathematical formula (2).

By the way, when recognizing a course on which the own vehicle can travel safely, it is not necessary to recognize all the information included in the captured image, and it is sufficient to recognize only the information related to the course. That is, it is sufficient to extract the entropy of each pixel in a specific class related to the course. The specific class related to the course is, for example, a class such as a white line (partition line), a road surface, a vehicle, a pedestrian, etc., and a class such as a tree, the sky, etc. is not necessary in principle. The reason why there are vehicles and pedestrian classes is that if there are vehicles and pedestrians on the road surface, it is necessary to change the course. Therefore, the extraction unit 25 is configured to extract the entropy of each pixel in a specific class related to the course.

Further, when recognizing a course on which the own vehicle can travel safely, it is not necessary to recognize a course other than the course on which the own vehicle is scheduled to travel. The course on which the own vehicle is scheduled to travel can be estimated from vehicle information such as the destination and steering angle of the vehicle. Therefore, the extraction unit 25 is configured to extract the entropy of the pixels in the region of the planned course and in the vicinity thereof. In the present embodiment, the image recognition device 20 corresponds to an estimation unit that acquires vehicle information and estimates a planned course based on the vehicle information.

In addition, the entropy at the boundary tends to be extremely high in the captured image (see FIG. 3). FIG. 3 illustrates a region with high entropy. The boundary described here is a boundary between identification objects, for example, a boundary between a road surface and a structure, a boundary between a road surface and a white line, a boundary between a road surface and a sidewalk, a boundary between a tree and the sky, and the like. Is. If there is a pixel with extremely high entropy, it will be dragged by it and the overall entropy will increase, which will affect the recognition accuracy. Therefore, the extraction unit 25 is configured to calculate the entropy of the captured image excluding the vicinity of the boundary.

Further, in the present embodiment, the lane marking unit 22 is configured to detect the lane marking. Then, the object detection unit 23 detects an object such as a vehicle or a pedestrian. When these are detected, the extraction unit 25 is configured to extract the entropy of each pixel in the region including the detected target (lane marking, vehicle, pedestrian).

Next, the reliability calculation unit 26 will be described. The reliability calculation unit 26 calculates the pixel reliability for each pixel based on the entropy extracted by the extraction unit 25. Basically, the reliability calculation unit 26 calculates the pixel reliability of the pixel as the entropy of the pixel increases.

It is also known that the recognition accuracy differs for each class. For example, since the white line (partition line) has the same image pattern to some extent, the white line class can be said to be a class that is easily recognized. On the other hand, in the pedestrian class, it can be said that the class is difficult to recognize because the image pattern varies. Therefore, the reliability calculation unit 26 calculates the pixel reliability with different weighting according to the class associated with the pixel. For example, in the order of pavement> road surface> moving object (vehicle, pedestrian), the pixel reliability is different depending on the class so that the pixel reliability in the pavement class is most likely to be the highest. Is calculated.

In addition, normalization is performed according to the area in each class. That is, the larger the area, the higher the pixel reliability tends to be. Therefore, the reliability calculation unit 26 weights the pixels so that the larger the area of the class, the lower the pixel reliability.

After calculating the pixel reliability for each pixel in this way, the reliability calculation unit 26 calculates the calculation result on a flat surface (looking down at the vehicle from directly above) based on the position (height from the road surface and depression angle) of the monocular camera 21a. Convert to (figure). At that time, the pixel reliability is output (displayed) by the occupancy grid map. As a result, as shown in FIG. 4, on the occupied grid map, the portion having high pixel reliability is likely to be white, and the portion having low pixel reliability is likely to be black.

Further, when outputting the pixel reliability on the occupied grid map, the reliability calculation unit 26 selects the calculated pixel reliability with reference to the time series (history) of the pixel reliability, and the selected result. It is configured to output only. Specifically, the reliability calculation unit 26 determines the tendency from the time series of the pixel reliability, removes the pixel reliability that deviates from the tendency, and selects the pixel reliability along the tendency. For example, the reliability calculation unit 26 calculates an average value or a median value from a time series of pixel reliability in a predetermined period, and selects pixel reliability existing within a predetermined range centered on the average value or the median value. ,Output. On the other hand, pixel reliability that does not exist within a predetermined range is excluded. In this case, if it can be determined that the pixel reliability tends to increase or decrease, it may be excluded even when the pixel reliability deviates from these tendencies.

When referring to the time series of pixel reliability, it is necessary to consider that the pixels (areas) to be compared on the captured image also move as the own vehicle moves. How the pixel to be compared has moved may be specified by the speed and the traveling direction of the own vehicle.

Similarly, when referring to the time series of pixel reliability, it is considered that the pixel (area) to be compared may move on the captured image as the moving object (other vehicle, pedestrian, etc.) moves. There is a need to. In this case, it may be specified by the speed of the own vehicle, the traveling direction, and the speed and traveling direction of the moving body.

When the pixel reliability is removed (not selected), the result is output assuming that the pixel reliability is the lowest in the region (pixel). Further, when the pixel reliability is not calculated due to reasons such as not being the region to be extracted for entropy, the result is similarly output assuming that the pixel reliability is the lowest in the region (pixel).

Next, the feedback unit 27 will be described. The feedback unit 27 is configured to return the result output by the reliability calculation unit 26 to the classification unit 24 and reflect it in the classification in the next cycle and thereafter. Specifically, the feedback unit 27 designates a region in which the pixel reliability is lower than a predetermined first threshold value as a region of interest in the classification unit 24.

When the area of interest is designated, the classification unit 24 classifies the pixels at a higher density than usual in the area of interest. That is, the classification unit 24 classifies the pixels in the region of interest by making the pixel size smaller (increasing the resolution) than usual.

The feedback unit 27 designates a region in which the pixel reliability is higher than a predetermined second threshold value as a stable region in the classification unit 24. When a stable region is designated, the classification unit 24 classifies pixels at a lower density than usual in the stable region. That is, the classification unit 24 classifies the pixels in the stable region by making the pixel size coarser (lowering the resolution) than usual.

Next, the course recognition unit 28 will be described. The course recognition unit 28 identifies a course on which the vehicle can travel stably based on the result output by the reliability calculation unit 26, that is, the pixel reliability for each pixel. Specifically, in the pixels to which the road surface classes are associated, the region where the pixel reliability is higher than the predetermined threshold value is specified as the path. At that time, it is desirable to specify the course along the planned course as much as possible.

The vehicle control unit 31 travels its own vehicle based on the course recognized by the course recognition unit 28. At that time, the vehicle control unit 31 may acquire the result output by the reliability calculation unit 26 and change various control amounts (control amounts of steering angle, accelerator, brake, etc.) based on the pixel reliability. .. For example, the vehicle control unit 31 may travel at a slower speed than usual when traveling in a region where the pixel reliability is lower than the normal value. Further, when the pixel reliability is lower than a predetermined value determined in advance, the traveling may be stopped.

Further, for example, when the recognized path includes a region where the pixel reliability is lower than the normal value, the vehicle control unit 31 has a margin in time and distance before entering the region. If so, the determination of whether to drive the vehicle and the control amount may be withheld. Then, the history of the pixel reliability in the region may be acquired over a predetermined period, and whether or not the vehicle is driven and the control amount may be determined with reference to the history. That is, the determination may be made after waiting for the pixel reliability in the region to become higher than the normal value.

Next, the image recognition process executed by the image recognition device 20 at predetermined intervals will be described with reference to FIG. First, the image recognition device 20 as the image acquisition unit 21 acquires a captured image (step S101).

Next, the image recognition device 20 as the lane marking detection unit 22 detects the lane marking based on the acquired captured image (step S102). Further, the image recognition device 20 as the object detection unit 23 detects an object based on the acquired captured image (step S103).

Next, the image recognition device 20 as the classification unit 24 classifies each pixel based on the acquired captured image (step S104). At that time, when the region of interest was designated by the feedback unit 27 in the previous cycle, the image recognition device 20 classifies the pixels in the region of interest by making the pixel size smaller than usual. When a stable region is specified, the pixel size is made coarser than usual in the stable region to classify the pixels.

Next, the image recognition device 20 as the extraction unit 25 calculates the entropy for each pixel (step S105). In step S105, the image recognition device 20 is a specific class related to the course, and extracts the entropy of the pixels by narrowing down to the area within or near the area of the planned course. At that time, the image recognition device 20 extracts the entropy by excluding the pixels near the boundary in the captured image.

Further, when the lane marking is detected by the lane marking detection unit 22, the entropy of the pixel in the region including the detected lane marking is extracted. When an object is detected by the object detection unit 23, the entropy of the pixels in the region including the detected object is extracted.

Next, the image recognition device 20 as the reliability calculation unit 26 calculates and outputs the pixel reliability for each pixel based on the entropy (step S106). At that time, the reliability calculation unit 26 calculates the pixel reliability with different weighting according to the class associated with the pixel. In addition, normalization is performed according to the area in each class. Further, the image recognition device 20 outputs (displays) the pixel reliability of the calculation result on the occupied grid map. At that time, the image recognition device 20 selects the calculated pixel reliability with reference to the time series of the pixel reliability, and outputs only the selected result on the occupied grid map.

Further, the image recognition device 20 as the feedback unit 27 determines that the result output by the reliability calculation unit 26 is reflected in the classification in the next cycle and thereafter (step S107). That is, the image recognition device 20 determines that a region in which the pixel reliability is lower than the predetermined first threshold value is designated as the region of interest and is reflected in the classification in the next cycle and thereafter. Further, the image recognition device 20 determines that a region in which the pixel reliability is higher than a predetermined second threshold value is designated as a stable region and is reflected in the classification in the next cycle and thereafter.

After that, the image recognition device 20 as the course recognition unit 28 identifies the course on which the vehicle can travel stably based on the calculated pixel reliability for each pixel, and outputs it to the vehicle control device 30 (step S108). .. Then, the image recognition process is terminated. The vehicle control device 30 controls the traveling of the own vehicle according to the course recognized by the image recognition device 20.

According to the above-described embodiment described in detail above, the following excellent effects can be obtained.

Even if the shooting target (road surface, white line, etc.) has acquired an unclear captured image due to the surrounding conditions or the convenience of the monocular camera 21a, or even if a captured image that is difficult to recognize has been acquired. The reliability calculation unit 26 selects the calculated pixel reliability with reference to the time series of pixel reliability. Therefore, it is possible to exclude a temporary outlier of the pixel reliability due to a change in the environment or the like, and it is possible to improve the calculation accuracy of the pixel reliability. Therefore, it is possible to accurately recognize the accuracy of image recognition, that is, the course on which the own vehicle can travel stably.

Since the extraction unit 25 extracts the entropy of each pixel in a specific class related to the course of the own vehicle, the calculation efficiency is improved, and as a result, the calculation accuracy can be improved. That is, it is possible to omit the image recognition of the useless area and sufficiently perform the image recognition of the necessary area, so that the recognition accuracy can be improved. In addition, the calculation burden can be reduced.

Depending on the class, there are classes that are easy to recognize because the pattern is fixed to some extent. On the contrary, there are some classes that are difficult to recognize. For example, the white line is a class that is easier to recognize than moving objects and trees. Therefore, the reliability calculation unit 26 is configured to weight the pixel reliability for each class. Thereby, the calculation accuracy of the pixel reliability can be improved.

When determining the reliability for each class by the sum of the pixel reliability for each class, the area of the class has an effect. Therefore, by normalizing the pixel reliability based on the area of each class, it is possible to make a judgment based on the pixel reliability regardless of the size of the area. Thereby, the accuracy can be improved.

At the boundary between objects or the boundary between the white line and the road surface, the entropy becomes high, which may cause misrecognition. Therefore, the extraction unit 25 is configured to extract the entropy by removing the boundary in the captured image. As a result, erroneous recognition can be suppressed and the calculation accuracy of the pixel reliability can be improved.

The extraction unit 25 is configured to extract entropy for pixels in a region including an object (white line, vehicle, pedestrian, etc.) detected by the lane marking detection unit 22 or the object detection unit 23, and calculate the pixel reliability. Has been done. Since the target to be detected (recognized) is narrowed down by a plurality of methods in this way, the recognition accuracy can be further improved. In addition, the area for which the pixel reliability should be calculated can be reduced, and the calculation efficiency can be improved.

The extraction unit 25 is configured to extract the entropy of the pixels in the region including the planned course. As a result, the own vehicle can be recognized only in the area that needs to be recognized in order to recognize the course on which the vehicle can travel safely, and the recognition accuracy can be further improved.

The classification unit 24 classifies each pixel by making the pixel size smaller than usual for the region of interest designated by the feedback unit 27 that the pixel reliability is lower than the threshold value. Therefore, the calculation accuracy of the pixel reliability can be improved.

On the other hand, in the stable region where the pixel reliability is specified to be higher than the threshold value, it is expected that the recognition accuracy will be sufficiently high thereafter. When detailed recognition is performed again in such a region, the calculation efficiency becomes poor. Further, when the pixel reliability is low in such a stable region, there is a high possibility that some kind of erroneous recognition has occurred. Therefore, the pixel size is made coarser than usual with respect to the stable region, and each pixel is classified. As a result, it is possible to suppress the fluctuation of the pixel reliability and improve the calculation accuracy of the pixel reliability.

(Other embodiments)
The present invention is not limited to the above embodiment, and various modifications can be made within the scope of the gist of the present invention. In the following, the parts that are the same or equal to each other in each embodiment are designated by the same reference numerals, and the description thereof will be incorporated for the parts having the same reference numerals.

-In the above embodiment, it is not necessary to weight the pixel reliability for each class.

-In the above embodiment, it is not necessary to normalize the pixel reliability based on the area of each class.
-In the above embodiment, the entropy near the boundary is not extracted, but the entropy near the boundary may be extracted.

-In the above embodiment, the lane marking detection unit 22 may not be provided. Similarly, the object detection unit 23 may not be provided.

-In the above embodiment, the planned course is estimated and the area is narrowed down so as to extract the entropy in the area including the planned course, but it is not necessary to narrow down the area in that way.

-In the above embodiment, the feedback unit 27 may not be provided. That is, it is not necessary to perform feedback control.

-In the above embodiment, the feedback unit 27 may provide feedback to the reliability calculation unit 26. For example, the stable region designated as having a pixel reliability higher than the threshold value may be weighted so that the pixel reliability becomes higher after the next cycle. As a result, the calculation efficiency can be improved and the occurrence of erroneous determination can be suppressed.

-In the above embodiment, the feedback unit 27 reflects the result in the next cycle or later, but it may be reflected in the result in the current cycle. That is, the same captured image may be classified by different pixel sizes, entropy extraction, pixel reliability calculation, and the like.

-In the above embodiment, the image recognition device 20 outputs the pixel reliability on the occupied grid map, but the output format may be arbitrarily changed. The calculation result may be output (displayed) on the saliency map. This makes it easier to use the calculation results in other systems.

-In the above-described embodiment and modification, the control unit and the method thereof described in the present disclosure constitute a processor and a memory programmed to execute one or a plurality of functions embodied by a computer program. It may be realized by a dedicated computer provided by. Alternatively, the controls and methods thereof described in the present disclosure may be implemented by a dedicated computer provided by configuring the processor with one or more dedicated hardware logic circuits. Alternatively, the control unit and method thereof described in the present disclosure may be a combination of a processor and memory programmed to perform one or more functions and a processor composed of one or more hardware logic circuits. It may be realized by one or more dedicated computers configured. Further, the computer program may be stored in a computer-readable non-transitional tangible recording medium as an instruction executed by the computer.

20 ... image recognition device, 21 ... image acquisition unit, 21a ... monocular camera, 25 ... extraction unit, 26 ... reliability calculation unit, 28 ... course recognition unit.

Claims (9)

In the image recognition device (20) that recognizes the course of the own vehicle based on the captured image,
An image acquisition unit (21) that acquires a photographed image from the photographing device (21a) at predetermined intervals, and
An extraction unit (25) that extracts the entropy of each pixel in the acquired captured image, and
A reliability calculation unit (26) that calculates and outputs the pixel reliability for each pixel based on the entropy extracted by the extraction unit, and
A course recognition unit (28) that recognizes a course based on the pixel reliability output by the reliability calculation unit is provided.
The reliability calculation unit is an image recognition device that refers to a time series of pixel reliability in a predetermined period, selects the calculated pixel reliability, and outputs the selected pixel reliability. It is equipped with a classification unit (24) that analyzes each image and classifies each pixel.
The image recognition device according to claim 1, wherein the extraction unit extracts the entropy of each pixel in a specific class related to the course of the own vehicle. The image recognition device according to claim 2, wherein the reliability calculation unit weights pixel reliability for each class. The image recognition device according to claim 2 or 3, wherein the reliability calculation unit normalizes the pixel reliability based on the area of each class. A feedback unit (27) for designating an area based on the pixel reliability calculated by the reliability calculation unit is provided.
Of claims 2 to 4, the classification unit classifies each pixel by making the pixel size smaller than usual for a region of interest designated by the feedback unit as having a pixel reliability lower than a threshold value. The image recognition device according to any one of the items. The image recognition device according to any one of claims 1 to 5, wherein the extraction unit extracts entropy from the captured image excluding boundaries between objects. It is equipped with detection units (22, 23) that detect the target based on the image pattern of the captured image.
The image recognition device according to any one of claims 1 to 6, wherein the extraction unit extracts entropy for pixels in a region including a target detected by the detection unit. Equipped with an estimation unit (20) that acquires vehicle information and estimates the planned course of the vehicle based on the vehicle information.
The image recognition device according to any one of claims 1 to 7, wherein the extraction unit extracts entropy for pixels in a region including a planned course. A feedback unit that specifies an area based on the pixel reliability calculated by the reliability calculation unit is provided.
The image recognition device according to any one of claims 1 to 8, wherein weighting is performed so that the pixel reliability is high in the stable region designated by the feedback unit as having a pixel reliability higher than the threshold value. ..

Images