JP2019003343A - Driving support device and driving support method - Google Patents

Abstract

To provide a technique capable of determining the degree of risk even in a case where a mobile body is not detected.SOLUTION: A driving support device 20 comprises an input circuit 30 and an output circuit 36, and can be installed in a vehicle 100. The driving support device 20 divides a first image having a first pixel number to be inputted to the input circuit 30 into at least a road region and any other region than the road region. The driving support device 20 detects an edge portion that enters the road region from the other region than the road region in a first image. The driving support device 20 calculates the degree of risk on the basis of a second image including the edge portion and having a second pixel number that is less than the first pixel number. In a case where the calculated degree of risk is equal to or greater than a predetermined value, the output circuit 36 outputs a predetermined signal.SELECTED DRAWING: Figure 2

Description

  The present invention relates to a driving support device and a driving support method for determining a risk level from an image.

  The collision avoidance assistance device for a vehicle detects a moving body such as a pedestrian existing around the vehicle, and issues a warning to the driver when it is determined that there is a risk of a collision between the moving body and the vehicle. In order to improve the accuracy of predicting the risk of collision, the sidewalk boundary around the vehicle is also detected, and the relative distance between the moving body and the sidewalk boundary and the relative distance between the moving body and the vehicle are used. (For example, refer to Patent Document 1).

JP 2010-181928 A

  In Patent Document 1, when a moving body is detected, the moving body exceeds the sidewalk boundary by using the relative distance between the moving body and the sidewalk boundary and the relative distance between the moving body and the vehicle. Determine the risk of jumping out towards the vehicle. However, when a moving object hidden behind a shield such as a parked vehicle pops out, the risk of collision between the moving object and the vehicle is not determined when the moving object is not detected.

  This invention is made | formed in view of such a condition, The objective is to provide the technique which determines a danger, even when a moving body is not detected.

  In order to solve the above-described problem, a driving support device according to an aspect of the present invention is a driving support device that includes an input circuit and an output circuit and can be installed in a vehicle, and is a first pixel that is input to the input circuit. In the first image having a number, it is divided into at least a road region and a region other than the road region, and in the first image, an edge portion entering from the region other than the road region is detected in the road region, including the edge portion, and Based on the second image having the second number of pixels smaller than the first number of pixels, the risk level is calculated, and when the calculated risk level is equal to or greater than a predetermined value, the output circuit outputs a predetermined signal. .

  Another aspect of the present invention is a driving support method. This method is a driving support method that can be used in a driving support device that includes an input circuit and an output circuit, and that can be installed in a vehicle. In the first image that includes the first number of pixels that are input to the input circuit, A second pixel that is divided into a road region and a region other than the road region, detects an edge portion that has entered the road region from a region other than the road region in the first image, includes the edge portion, and is smaller than the first pixel number. The risk level is calculated based on the second image having a number, and when the calculated risk level is equal to or greater than a predetermined value, the output circuit outputs a predetermined signal.

  It should be noted that any combination of the above-described constituent elements and a conversion of the expression of the present invention between a method, an apparatus, a system, a recording medium, a computer program, etc. are also effective as an aspect of the present invention.

  According to the present invention, it is possible to determine the degree of risk even when a moving object is not detected.

It is a figure which shows the structure which looked at the vehicle interior of the vehicle which concerns on an Example from back. It is a figure which shows the structure of the driving assistance apparatus of FIG. FIGS. 3A to 3C are diagrams illustrating an outline of processing of the driving support apparatus of FIG. FIGS. 4A to 4B are diagrams showing another processing outline of the driving support apparatus of FIG. It is a flowchart which shows the output procedure by the driving assistance device of FIG. It is a figure which shows the structure of the driving assistance system which concerns on a modification.

  Before describing the present invention in detail, an outline will be described. Embodiments of the present invention relate to a driving support apparatus that is mounted on a vehicle and determines a degree of danger from a captured image and notifies a driver. Here, in particular, the degree of danger of a pedestrian jumping out is determined from the image. One technique for executing processing on an image is a road terrain detection technique. In conventional road terrain detection techniques, only lane detection is performed on an image or only sidewalk boundary detection is performed on an image in many cases. Such terrain detection for a specific object can be performed at high speed. However, the detection accuracy of pedestrians, vehicles, buildings, buildings, etc. included in the image is insufficient. Moreover, in the prediction technique for predicting the pedestrian jumping out onto the roadway, when a pedestrian existing on the roadway or the sidewalk is detected, a risk level of whether the pedestrian jumps out into the own vehicle is determined. However, since this process is based on the assumption that a pedestrian has been detected in the image, the degree of risk cannot be determined when a pedestrian is not detected in the image, such as a pedestrian jumping out from a shield.

  In order to cope with this, the driving support apparatus according to the present embodiment performs the following four-stage processing on the image. In the first stage, semantic segmentation is performed on an image captured by the camera (hereinafter referred to as “first image”), thereby performing meaning for each pixel of the first image. As a result, a region is recognized for each person, other vehicle, road, building, traffic light, etc. included in the first image. As a result, labeling is performed for each category such as a person area, another vehicle area, and a road area.

  In the second stage, the road region in the first image that has been labeled is specified. When a moving body such as a person or another vehicle does not exist on the road, the road area is in contact with the building area and the sky area, and the boundary is generally indicated by a straight line or a smooth curve. On the other hand, when the moving body exists on the road, there is an edge portion where the person area and the other vehicle area enter the road area. Therefore, an edge part is detected in the second stage. This corresponds to detecting whether the contact surface changes discontinuously in the road region.

  In the third stage, for one edge portion, a plurality of types of images having different sizes while including the edge portion are extracted from the first image. For example, when four types of images having different sizes are extracted, these images are called a second image,..., A fifth image. The number of pixels of the second image,..., The fifth image is smaller than the number of pixels of the first image. Note that the second image,..., The fifth image may have different orientations as well as sizes. This process is performed for each edge portion.

  In the fourth stage, deep learning is executed by inputting each of the second image,..., And the fifth image at one edge portion into a neural network that has previously learned what the attention object is. Then, the risk of the edge portion is derived. For example, the degree of danger is learned so as to increase as there is an edge portion of a gap between the vehicle and the vehicle at risk of popping out a person and an edge portion of the person facing the roadway direction. Further, in the second image,..., The fifth image, the occupancy ratio of the edge portion in the image is different, so that the overall situation including the edge portion and the local situation near the edge portion are shown. Since these are used for deep learning, the determination accuracy is improved. This process is also performed for each edge portion. In these four stages of processing, it is possible to predict the jumping out of a person between vehicles parked on the road.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In addition, each Example described below is an example, and this invention is not limited by these Examples.

  FIG. 1 shows a configuration of the interior of a vehicle 100 according to the embodiment as viewed from the rear. A dashboard 10 extending left and right is disposed in front of the passenger compartment, and a windshield 12 is disposed above the dashboard 10. A ceiling 14 is arranged from the upper side of the windshield 12 to the rear side. A room mirror 16 is installed downward from a portion of the ceiling 14 on the windshield 12 side. Furthermore, a driving assistance device 20 is installed on the windshield 12 on the left side of the room mirror 16. The driving support device 20 may be installed on the right side of the rearview mirror 16 in the windshield 12.

  FIG. 2 shows a configuration of the driving support device 20. The driving support device 20 includes an input circuit 30, a storage device 32, a CPU (Central Processing Unit) 34, and an output circuit 36. The driving support device 20 is installed in the vehicle 100 as described above.

  The input circuit 30 includes an imaging device such as a camera, for example. The input circuit 30 is provided at a position where the front of the vehicle 100 can be imaged and captures a moving image. The moving image is composed of a plurality of first images that are sequentially captured. Each first image has a first number of pixels. In the following description, the processing for one first image will be described, but the same processing may be performed for the other first images. The input circuit 30 outputs the first image to the CPU 34. The input circuit 30 may include an interface such as a button or a touch panel for receiving an instruction from the user, and outputs the instruction received from the user to the CPU 34.

  The storage device 32 stores various data. When the driving support device 20 has a drive recorder function, the storage device 32 stores moving images by sequentially receiving and storing the first images from the input circuit 30 via the CPU 34.

  The CPU 34 executes various computer programs. The CPU 34 may be connected to a memory for storing various computer programs. Here, the processing by the computer program executed in the CPU 34 will be described in the order from the first stage to the fourth stage.

(1) First Stage In the first stage, a first image from the input circuit 30 is received. FIGS. 3A to 3C show an outline of processing of the driving support device 20. FIG. 3A shows the first image 40. As an example, the first building 58 and the second building 60 are built on both sides of the road 56, and the sky 64 is disposed between the first building 58 and the second building 60. Further, the other vehicle 54 is parked on the first building 58 side on the road 56, and the first person 50 and the second person 52 exist in the vicinity of the other vehicle 54. Further, a traffic light 62 is installed on the road 56. FIGS. 3B to 3C will be described later, and the description returns to FIG.

  In the first stage, semantic segmentation is performed on the first image 40. Semantic segmentation is also called image segmentation. This is a process of labeling the meanings indicated by the pixels for each of the first number of pixels provided in the first image 40. This is equivalent to labeling which class of object each pixel belongs to, which category.

  FIG. 3B shows a first image 40 obtained by performing semantic segmentation on the first image 40 in FIG. The other vehicle 54, the road 56, the traffic light 62, and the sky 64 in FIG. 3A are labeled as a vehicle area 72, a road area 74, a traffic light area 78, and an empty area 80, respectively. Also, the first person 50 and the second person 52 in FIG. 3A are labeled as a person area 70, and the first building 58 and the second building 60 in FIG. 3A are labeled as a building area 76. . The first image 40 shown in FIG. 3 (b) recognizes the object included in the first image 40 shown in FIG. 3 (a), classifies the object for each category, and sets the color for each classified category. It can be said that it was generated by changing the space. FIG. 3C will be described later, and the processing returns to FIG.

  That is, in the first stage, the first image 40 is divided into at least a road area 74 and an area other than the road area 74. For such semantic segmentation, i.e., classification of the road area 74 and the area other than the road area 74, for example, Fully Convolutional Neural Network is used.

(2) Second Stage When there is no object on the road 56, the boundary of the road region 74 in the labeled first image 40 (hereinafter also referred to as “first image 40”) is a straight line or a smooth curve. Composed. On the other hand, when there is a person who is hidden in the first person 50, the second person 52, the other vehicle 54, and the other vehicle 54, which may cause danger, on the road 56, the road area 74 is caused by at least one of them. The boundary changes to a distorted shape. The distorted shape at the boundary of the road region 74 is called an “edge portion”. In the second stage, an edge portion that enters the road area 74 from an area other than the road area 74 is detected. The edge portion is detected by specifying a portion where the shape of the line is greatly changed in a predetermined section of the line constituting the boundary of the road region 74. For example, an edge portion is detected in a portion where the change in the differential value of a line in a predetermined section is larger than a threshold value.

  FIG. 3C shows the result of detecting the edge portion 90 for the first image 40 in FIG. A plurality of edge portions 90 may be detected, and here, five edge portions 90 are detected. Thus, when detecting the edge part 90, the exact positional information on whether the person area 70 or the vehicle area 72 is in contact with the road area 74 is needed. In object detection, since it is unclear to which object the pixel belongs, it cannot be accurately determined whether the person area 70 or the vehicle area 72 is in contact with the road area 74. In order to perform the second stage, it is necessary to label the first image 40 in the first stage. Returning to FIG.

(3) Third Stage In the third stage, for each detected edge portion 90, a plurality of images including the edge portion 90 are generated. In order to explain this process, FIGS. 4A to 4B are used here. FIGS. 4A and 4B show another processing outline of the driving support device 20. FIG. 4A corresponds to the first image 40 similar to FIG. As illustrated, a vehicle area 72 is disposed on the left side of the road area 74, and a person area 70 is disposed in the vicinity of the vehicle area 72. Here, for the sake of clarity, the person area 70 disposed on the far side of the vehicle area 72 is referred to as a first person area 70a, and the person area 70 disposed on the near side of the vehicle area 72 is referred to as a second person area. 70b. Furthermore, the edge part 90 detected at the boundary between the first person area 70a and the road area 74 is indicated as a first edge part 90a. Edge portions 90 detected at the boundary between the vehicle region 72 and the road region 74 are indicated as a second edge portion 90b and a third edge portion 90c. The edge part 90 detected at the boundary between the second person area 70b and the road area 74 is indicated as a fourth edge part 90d.

  In the third stage, as shown in FIG. 4B, the second image 42 including the first edge portion 90 a and having the second number of pixels smaller than the first number of pixels is extracted from the first image 40. In the third stage, the third image 44 to the fifth image 48 are also extracted from the first image 40. The third image 44 has a third number of pixels, the fourth image 46 has a fourth number of pixels, and the fifth image 48 has a fifth number of pixels. These include a first edge portion 90a. Further, the third to fifth pixel numbers are smaller than the first pixel number, but they are different from each other. That is, in the third stage, the fifth image 48 is cut out from the second image 42 having a different number of pixels while including the first edge portion 90a.

  Of the second image 42 to the fifth image 48, an image with a small number of pixels has a large occupation ratio of the first edge portion 90 a in the image, and thus a local situation is indicated. On the other hand, among the second image 42 to the fifth image 48, an image having a large number of pixels has a smaller occupation ratio of the first edge portion 90a in the image, and thus the overall situation is indicated. Here, four types of images of the fifth image 48 are extracted from the second image 42 with respect to the first edge portion 90a, but are not limited to four types. Further, in the third stage, the same processing is performed for the second edge portion 90b to the fourth edge portion 90d. That is, the extraction of the fifth image 48 from the second image 42 is performed for each edge portion 90. Returning to FIG.

(4) Fourth stage Before executing the fourth stage, learning is performed based on images of various situations including the edge portion 90 and supervised data that is the degree of risk associated with the situation. A learning result is obtained by processing. Specifically, by inputting an image to a neural network, the weight parameters of the neural network are adjusted as learning results so that supervised data is obtained as an output. Here, the degree of danger is an index indicating the probability of a collision. For example, there is an edge portion 90 of a gap between a vehicle and a person at risk of popping a person, and an edge portion 90 of a person facing a roadway direction. , Raised.

  Subsequently, in the fourth stage, based on each of the second image 42 to the fifth image 48 with respect to the first edge portion 90a and the learning result, a neural network is used to apply to the first edge portion 90a. Calculate the risk. Specifically, each of the second image 42 to the fifth image 48 for the first edge portion 90a is input to a learned neural network, and the degree of risk output from the neural network is acquired. Here, the first edge portion 90a is detected in the first person area 70a, the foot is facing the road area 74, and the first person area 70a is hidden in the vehicle area 72 and is in a position that is difficult to see. High risk is used as a criterion element. The degree of risk is calculated based on the stipulation defined for each element, and the degree of risk is quantified, for example, “80”.

  Such processing is performed for each of the second edge portion 90b to the fourth edge portion 90d. The second edge portion 90b on the back side is detected in the vehicle region 72 with respect to the second edge portion 90b, and it is difficult to grasp the pedestrian jumping out of the first person region 70a existing in the back of the vehicle region 72. Larger is used as a criterion element. Therefore, the degree of danger is quantified as “60”. With respect to the third edge portion 90c, the third edge portion 90c on the near side in the vehicle region 72 is detected, and it is easy to grasp the pedestrian jumping out of the second person region 70b existing in front of the vehicle region 72. Smallness is used as a criterion element. Therefore, the degree of danger is quantified as “10”.

  With respect to the fourth edge portion 90d, the fourth edge portion 90d is detected in the second person area 70b, the foot is facing the road area 74, and the second person area 70b is hidden in the vehicle area 72. It is used as a criterion element because it is in a position that is easy to see and has low risk. Therefore, the degree of danger is quantified as “50”. In the fourth stage, for example, the threshold value is set to “60”. The risk level for each edge portion 90 is compared with a threshold value, and a risk level equal to or greater than the threshold value is determined as dangerous. In the case described above, when the threshold value is set to “60”, “80” “pedestrian hidden in the car” and “60” “vehicle shadow where a pedestrian easily jumps out” are determined as dangerous. When it is determined as dangerous in the fourth stage, the determination result is notified to the output circuit 36.

  The output circuit 36 includes at least a speaker that emits sound, for example. When the notification from the CPU 34 is received, that is, when the calculated degree of danger is greater than or equal to a threshold value, the output circuit 36 outputs a sound for notifying the driver of the danger. The sound may be for simply informing the danger, or may be for specifically informing the contents of the danger. In the latter case, the content of the risk level that is equal to or greater than the threshold value is output as sound. Further, the output circuit 36 may be a display device, and executes display for notifying danger.

  This configuration can be realized in terms of hardware by a CPU, memory, or other LSI of any computer, and in terms of software, it can be realized by a program loaded in the memory, but here it is realized by their cooperation. Draw functional blocks. However, it will be understood by those skilled in the art that these functional blocks can be realized in various forms by hardware alone or by a combination of hardware and software.

  The operation of the driving support apparatus 20 having the above configuration will be described. FIG. 5 is a flowchart showing an output procedure by the driving support device 20. The first image 40 is input to the CPU 34 (S10). The CPU 34 performs labeling on the first image 40 by executing semantic segmentation (S12). If the edge portion 90 entering the road area 74 is not detected (N in S14), the process returns to step 10. If the edge portion 90 that has entered the road area 74 is detected (Y in S14), the CPU 34 extracts the fifth image 48 from the second image 42 for each edge portion 90 (S16). The CPU 34 calculates the degree of risk by inputting each of the second image 42 to the fifth image 48 to the neural network (S18). If the degree of danger is not greater than or equal to the threshold value (N in S20), the process returns to step 10. If the degree of risk is greater than or equal to the threshold (Y in S20), the output circuit 36 outputs that it is dangerous (S22), and returns to Step 10.

  Next, a modified example will be described. In the embodiment, the CPU 34 included in the driving support device 20 executes processing from the first stage to the fourth stage. However, the driving support device 20 may perform at least one of the first to fourth steps in the server by wirelessly communicating with a server installed outside the vehicle 100. FIG. 6 shows a configuration of a driving support system 200 according to a modification. The driving support system 200 includes a driving support device 20, a network 22, and a server 24. The driving support device 20 includes a communication device 38 in addition to the configuration of FIG.

  The communication device 38 is connected to the CPU 34, includes an antenna unit 39, and has a wireless communication function via the antenna unit 39. An example of wireless communication is cellular phone communication, but is not limited to this. The communication device 38 can communicate with the server 24 via the antenna unit 39 of the driving support device 20, the antenna unit 28 of the base station 26, and the network 22.

  The server 24 executes at least one of the processes from the first stage to the fourth stage. That is, the server 24 executes a part of processing performed in the CPU 34. In order to realize the sharing of processing between the server 24 and the CPU 34, communication between the server 24 and the communication device 38 is executed. For example, when the server 24 executes all the processes from the first stage to the fourth stage, the communication device 38 transmits the first image 40 to the server 24, and the server 24 transmits the determination result to the communication device 38.

  According to the present embodiment, since the edge portion entering from the area other than the road area is detected in the road area divided in the first image, the person area or vehicle area in contact with the road area can be specified. Moreover, since the neural network is used, the accuracy of the division into the road area and the area other than the road area can be improved. In addition, since the second image including the detected edge portion and having a smaller size than the first image is extracted from the first image, the person image or the vehicle region in contact with the road region is surrounded by the second image. A focused second image can be generated. Further, since the risk level indicating the probability of collision is calculated based on the second image focusing on the person area or the vehicle area in contact with the road area, it is possible to determine the risk level even when no moving object is detected. . In addition, since the second image focusing on the person area or the vehicle area in contact with the road area is input to the neural network, the calculation accuracy of the risk level can be improved.

  In addition, since the second image to the fifth image having different sizes are included in the neural network while including the detected edge portion, the degree of risk is considered in consideration of the local situation and the overall situation in the vicinity of the edge portion. Can be calculated. In addition, since the risk level is calculated in consideration of the local situation and the overall situation in the vicinity of the edge portion, it is possible to improve the calculation accuracy of the risk level. In addition, since the risk level can be calculated, the risk determination accuracy can be improved. Moreover, since a camera is included, a 1st image can be acquired. Moreover, since a speaker is included, danger can be notified.

  One embodiment of the present invention is as follows. A driving support apparatus according to an aspect of the present invention is a driving support apparatus that includes an input circuit and an output circuit and can be installed in a vehicle. In the first image including the first number of pixels input to the input circuit, At least a road region and a region other than the road region are divided, and in the first image, an edge portion that enters the road region from the region other than the road region is detected, and includes a second portion that includes the edge portion and is smaller than the first pixel Based on the second image having the number of pixels, the degree of risk is calculated, and when the calculated degree of risk is equal to or greater than a predetermined value, the output circuit outputs a predetermined signal.

  According to this aspect, since the edge part entering from the area other than the road area is detected in the road area divided in the first image, the risk is calculated based on the second image including the detected edge part. The degree of danger can be determined even when a moving object is not detected.

  In the driving support device, the input circuit may include at least a camera. In this case, since the camera is included, the first image can be acquired.

  In the driving support device, the output circuit may include at least a speaker that emits sound. In this case, since a speaker is included, danger can be notified.

  In the driving support device, the degree of risk may be an index indicating the probability of collision. In this case, since the probability of a collision is shown, it is possible to notify the danger that a danger will occur even if a moving body is not detected.

  In the driving support device, the first image may be divided into a road area and an area other than the road area using a neural network. In this case, since the neural network is used, the accuracy of the division into the road area and the area other than the road area can be improved.

  The driving support device may calculate the degree of risk using a neural network based on the second image including the edge portion. In this case, since the neural network is used, the calculation accuracy of the degree of risk can be improved.

  In addition to the second image, the driving support device further uses a third image that includes an edge portion and has a third pixel number that is less than the first pixel number and different from the second pixel number. When the degree of risk calculated based on the second image and the third image is greater than or equal to a predetermined value, the output circuit may output a predetermined signal. In this case, since the third image is used, the determination accuracy can be improved.

  A driving support device that uses a neural network based on a second image including an edge portion and a learning result and uses a neural network based on a third image including an edge portion and a learning result By doing so, the degree of risk may be calculated. In this case, since the neural network is also used for the third image, the accuracy of calculating the degree of risk can be improved.

  Another aspect of the present invention is a driving support method. This method is a driving support method that can be used in a driving support device that includes an input circuit and an output circuit, and that can be installed in a vehicle. In the first image that includes the first number of pixels that are input to the input circuit, A second pixel that is divided into a road region and a region other than the road region, detects an edge portion that has entered the road region from a region other than the road region in the first image, includes the edge portion, and is smaller than the first pixel number. The risk level is calculated based on the second image having a number, and when the calculated risk level is equal to or greater than a predetermined value, the output circuit outputs a predetermined signal.

  According to this aspect, since the edge part entering from the area other than the road area is detected in the road area divided in the first image, the risk is calculated based on the second image including the detected edge part. The degree of danger can be determined even when a moving object is not detected.

  In the driving support method, the input circuit may include at least a camera. In this case, since the camera is included, the first image can be acquired.

  In the driving support method, the output circuit may include at least a speaker that emits sound. In this case, since a speaker is included, danger can be notified.

  In the driving support method, the degree of risk may be an index indicating the probability of collision. In this case, since the probability of a collision is shown, it is possible to notify the danger that a danger will occur even if a moving body is not detected.

  In the driving support method, the first image may be classified into a road area and an area other than the road area using a neural network. In this case, since the neural network is used, the accuracy of the division into the road area and the area other than the road area can be improved.

  In the driving support method, the degree of risk may be calculated using a neural network based on the second image including the edge portion. In this case, since the neural network is used, the calculation accuracy of the degree of risk can be improved.

  In addition to the second image, the driving support method further uses a third image including an edge portion and having a third pixel number that is smaller than the first pixel number and different from the second pixel number. When the degree of risk calculated based on the second image and the third image is greater than or equal to a predetermined value, the output circuit may output a predetermined signal. In this case, since the third image is used, the determination accuracy can be improved.

  A driving support method using a neural network based on a second image including an edge portion and a learning result, and using a neural network based on a third image including an edge portion and a learning result By doing so, the degree of risk may be calculated. In this case, since the neural network is also used for the third image, the accuracy of calculating the degree of risk can be improved.

  In the above, this invention was demonstrated based on the Example. This embodiment is an exemplification, and it will be understood by those skilled in the art that various modifications can be made to each of those constituent elements or combinations of processing processes, and such modifications are also within the scope of the present invention. .

  10 dashboard, 12 windshield, 14 ceiling, 16 room mirror, 20 driving support device, 22 network, 24 server, 30 input circuit, 32 storage device, 34 CPU, 36 output circuit, 100 vehicle.

Claims (16)

A driving support device that includes an input circuit and an output circuit and can be installed in a vehicle,
In the first image having the first number of pixels input to the input circuit, at least a road area and an area other than the road area are classified,
In the first image, an edge portion that has entered the road area from the area other than the road area is detected,
Based on a second image including the edge portion and having a second number of pixels smaller than the first number of pixels, a risk level is calculated,
When the calculated risk is a predetermined value or more, the output circuit outputs a predetermined signal.
Driving assistance device. The driving support device according to claim 1,
The input circuit includes at least a camera.
Driving assistance device. The driving support device according to claim 1 or 2,
The output circuit includes at least a speaker that emits sound,
Driving assistance device. The driving support device according to any one of claims 1 to 3,
The degree of risk is an index indicating the probability of collision,
Driving assistance device. The driving support device according to any one of claims 1 to 4,
In the first image, the road area and the area other than the road area are classified using a neural network.
Driving assistance device. The driving support device according to any one of claims 1 to 5,
Based on the second image including the edge portion, the risk is calculated using a neural network.
Driving assistance device. The driving support device according to claim 6,
In addition to the second image, the degree of risk is calculated using a third image that includes the edge portion and has a third pixel number that is smaller than the first pixel number and different from the second pixel number. And
When the degree of risk calculated based on the second image and the third image is a predetermined value or more, the output circuit outputs the predetermined signal.
Driving assistance device. The driving support device according to claim 7,
Using a neural network based on the second image including the edge portion and the learning result, and using a neural network based on the third image including the edge portion and the learning result. By calculating the risk,
Driving assistance device. There is a driving support method that can be used in a driving support device that includes an input circuit and an output circuit and can be installed in a vehicle.
In the first image having the first number of pixels input to the input circuit, at least a road area and an area other than the road area are classified,
In the first image, an edge portion that has entered the road area from the area other than the road area is detected,
Based on a second image including the edge portion and having a second number of pixels smaller than the first number of pixels, a risk level is calculated,
When the calculated risk is a predetermined value or more, the output circuit outputs a predetermined signal.
Driving support method. The driving support method according to claim 9,
The input circuit includes at least a camera.
Driving support method. The driving support method according to claim 9 or 10,
The output circuit includes at least a speaker that emits sound,
Driving support method. The driving support method according to any one of claims 9 to 11,
The degree of risk is an index indicating the probability of collision,
Driving support method. The driving support method according to any one of claims 9 to 12,
In the first image, the road area and the area other than the road area are classified using a neural network.
Driving support method. The driving support method according to any one of claims 9 to 13,
Based on the second image including the edge portion, the risk is calculated using a neural network.
Driving support method. The driving support method according to claim 14,
In addition to the second image, the degree of risk is calculated using a third image that includes the edge portion and has a third pixel number that is smaller than the first pixel number and different from the second pixel number. And
When the degree of risk calculated based on the second image and the third image is a predetermined value or more, the output circuit outputs the predetermined signal.
Driving support method. The driving support method according to claim 15, comprising:
Using a neural network based on the second image including the edge portion and the learning result, and using a neural network based on the third image including the edge portion and the learning result. By calculating the risk,
Driving support method.

Images