Vehicle automatic driving visual ranging learning method and device

Abstract

The invention discloses a vehicle automatic driving visual ranging learning method and device, and the method comprises the steps: obtaining a driving image, generating a preset number of driving points in the driving image in the driving direction, and obtaining a road curve image according to all the driving points; recording the time, the distance and the vehicle speed of driving to each driving point; obtaining a road mapping function according to the road curve image and the time, the distance and the vehicle speed corresponding to all the current driving points; selecting a point location after the current target driving point location in the road curve image as a detection point location; obtaining the target time of the detection point according to the road mapping function; and calculating the actual distance between the current target driving point position and the detection point position according to the time of the current target driving point position, the vehicle speed and the target time. The target time of the detection point is obtained based on the road mapping function, and the actual distance is calculated by combining the time of the previous target driving point and the vehicle speed, so that the distance measurement precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power distribution network inspection, and in particular to a method and device for learning visual ranging of an automatic driving vehicle. Background Art

[0002] With the continuous development of science and technology, the technology of artificial intelligence is constantly improving, and the era of autonomous driving is getting closer to people's lives. In the autonomous driving system, ranging technology is an important part of realizing vehicle autonomous navigation and following. However, traditional ranging methods mainly rely on lidar. Although these devices can provide high-precision distance information, they are expensive. In contrast, visual ranging technology relies on cameras and image processing technology for ranging. Although the cost of cameras is low, they need to rely on algorithms to achieve ranging. However, the current algorithms still cannot provide accurate distance information. Summary of the invention

[0003] Generating a preset number of driving points in the driving direction, and obtaining a road curve image according to all the driving points; Record the time, distance and speed of each driving point; Obtaining a road mapping function according to the road curve image and the time, distance and vehicle speed corresponding to all the current driving points; Selecting a point after the current target driving point in the road curve image as a detection point; Obtaining the target time of the detection point according to the road mapping function; The actual distance between the current target driving point and the detection point is calculated according to the time and vehicle speed of the current target driving point and the target time.

[0004] In order to solve the above technical problems, another technical solution adopted by the present invention is: A route identification planning device in patrol route planning includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned vehicle autonomous driving visual ranging learning method is implemented.

[0005] The beneficial effects of the present invention are as follows: a preset number of driving points are generated in a driving image, and a road curve image is obtained according to all the driving points; at the same time, the time, distance and speed corresponding to each driving point are recorded during the driving process of the vehicle, and then a road mapping function corresponding to the road curve image is obtained according to the data corresponding to each driving point; then, after selecting a detection point in the road curve image, the target time of the detection point is obtained based on the road mapping function, and finally the actual distance is calculated according to the time and speed of the current target driving point and the target time, and the distance measurement is realized by relying on the visual distance measurement technology, thereby improving the distance measurement accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Figure 1 A flowchart of the steps of a method for learning visual ranging of an autonomous driving vehicle in an embodiment of the present invention; Figure 2 A schematic diagram of driving point generation in a visual ranging learning method for autonomous driving of a vehicle in an embodiment of the present invention; Figure 3 A schematic diagram of mapping function generation in a vehicle autonomous driving visual ranging learning method in an embodiment of the present invention; Figure 4 This is another step flow chart of a method for learning visual ranging of an autonomous driving vehicle in an embodiment of the present invention; Figure 5 Schematic diagram of the structure of a vehicle autonomous driving visual ranging learning device in an embodiment of the present invention. DETAILED DESCRIPTION

[0007] It can be seen from the above description that the road curve image is converted horizontally and vertically and stretched at a 45° angle, so that the road curve image can be better mapped to the road mapping function.

[0008] Furthermore, the initial mapping function includes: at 3 +bt 2 +ct+d; Among them, when the road curve image is a straight line, a and b are 0, when it is a single curve, a is 0, and when it is an S-curve, both a and b are not 0.

[0009] It can be seen from the above description that after identifying the road curve image, the parameters corresponding to the initial mapping function are adjusted so that the initial mapping function can fit different road curves, thereby improving the mapping accuracy of the road curve.

[0010] Furthermore, after obtaining the road mapping function according to the road curve image and the time, distance and vehicle speed corresponding to all the current driving points, the method further includes: Get historical road mapping function; The road mapping function and the historical road mapping function are learned through a first deep learning model to optimize the road mapping function.

[0011] From the above description, it can be seen that by obtaining the historical road mapping function generated during vehicle driving, learning the historical road mapping function based on the first deep learning model, and calculating the corresponding loss function, the mapping accuracy of subsequent road mapping functions can be improved.

[0012] Furthermore, generating a preset number of driving points along the driving direction in the driving image includes: Acquiring road pattern information and environmental information in the driving image; Converting the road pattern information and the environment information into a feature vector; The feature vector is processed by a second deep learning model to obtain the driving point.

[0013] From the above description, it can be seen that by collecting road pattern information and environmental information and converting them into feature vectors, the corresponding driving points are generated based on the second deep learning model, which can improve the accuracy of road recognition.

[0014] Furthermore, all previous processing results of the second deep learning model are obtained, and the loss function corresponding to the second deep learning model is optimized according to all the processing results.

[0015] From the above description, it can be seen that calculating the loss function corresponding to the second deep learning model based on the historical processing results can improve the accuracy of the second deep learning model for subsequent driving points.

[0016] Another embodiment of the present invention provides a route identification planning device in patrol path planning, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, each step in a vehicle autonomous driving visual ranging learning method as described above is implemented.

[0017] Another embodiment of the present invention provides a route identification planning device in patrol path planning, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, each step in a vehicle autonomous driving visual ranging learning method as described above is implemented.

[0018] The line identification planning method and device in the inspection path planning provided by the present invention can be applied to the power grid inspection path planning scenario to realize automatic inspection path planning, which is described below through specific implementation methods: Embodiment 1 Please refer to Figure 1, a vehicle automatic driving visual ranging learning method, comprising: S1. Acquire a driving image, generate a preset number of driving points in the driving image along the driving direction, and obtain a road curve image according to all the driving points; wherein the specific generation method of the driving points is as follows: S11. Obtain the point spacing according to the current vehicle speed; S=(V*t0) / N; wherein S represents the point spacing; N represents the number of driving points; V represents the current vehicle speed; t0 represents time, t0≤0.1s; for example, t0=0.1s, which means that after acquiring an image every 0.1s, a corresponding number of driving points are generated on the image. If the vehicle speed is fast, the driving points appear dense, and if the vehicle speed is slow, the driving points appear sparse.

[0019] S12, all driving points are sequentially arranged along the driving direction according to the point spacing, that is, the driving points are generated according to the obtained point spacing S. Figure 2 As shown in the figure, points A, B, and C are respectively the current vehicle driving point, the front short-distance sight point, and the front long-distance sight point. The spacing between the nodes can be adjusted according to the current vehicle speed. Point T represents the point to be detected; that is, the first N points are the points required to build the function model. Generally, the points split out in a short section of road before the driving sight distance are mainly used to build a complete function model. Point T is the point that requires sight distance detection.

[0020] At the same time, in the process of generating driving points, after the road pattern information and surrounding environment information are obtained by zooming in and scanning with the camera, the road pattern information and environmental information are converted into feature vectors, and then the feature vectors are processed by the second deep learning model to obtain the driving points. Among them, the road pattern information and the surrounding environment are mainly to find the sight point during vehicle driving. For example, there are many sight points marked on the road ahead within the camera's sight. The road pattern is detected at these sight points, and when the subsequent vehicle drives to the road pattern position, it is detected as the vehicle driving to the sight point. In this way, the time can be obtained and can be used for the subsequent combination of function parameters and subsequent T points. Furthermore, by obtaining the previous processing results of the second deep learning model, the loss function corresponding to the second deep learning model is optimized according to all the processing results, that is, the feature is trained and learned by the deep learning algorithm to obtain low loss, so that high-precision recognition can be achieved in the subsequent road pattern detection. In actual applications, the road pattern will be combined with the actual detection, and the loss will be reduced again by learning again.

[0021] S2. Record the time, distance and speed of each driving point; for example, when the vehicle drives to point A and point B, record the corresponding time, distance, speed and other vehicle-related data.

[0022] S3. Obtain a road mapping function based on the road curve image and the time, distance and vehicle speed corresponding to all current driving points; that is, drive multiple points in a short time interval and obtain the function amplitude and offset improvement function based on the recorded time, the distance traveled at the speed maintained in the time interval and the left and right offset of the vehicle; at the same time, the road function is continuously drawn and improved during the driving of the vehicle, and the improvement process needs to be combined with the above-mentioned road pattern detection to drive to the sight distance point, and the time obtained after the vehicle drives to the sight distance point is calculated, so as to map it to the vertical axis and horizontal axis in the sight distance function of the function, where the vertical axis mapped to the function is the sight distance point, and the horizontal axis is the time point when the vehicle drives to the sight distance point, please refer to Figure 3 , is a schematic diagram of road image mapping: S31, obtaining an initial mapping function according to the road curve image; converting the road curve image horizontally and vertically and extending it by 45° to obtain the initial mapping function; the specific expression of the mapping function is as follows: at 3 +bt 2 +ct+d; Among them, a and b are 0 when the road curve image is a straight line, a is 0 when there is a single curve, and both a and b are not 0 when there is an S-curve; since the current road conditions generally only have straight lines, ordinary curves and S-curves, straight lines can be directly represented by first powers, ordinary curves can be represented by second powers, and S-curves can be represented by third powers; so that the points on the mapping function correspond to the points on the road image one by one; this step is recorded by a computer, and when all subsequent points correspond to the mapping function, the same behavior will be used to find the point on the mapping function.

[0023] S32. The time, distance and speed data corresponding to all driving points are brought into the initial mapping function to generate a road mapping function. Specifically: In combination with deep learning road detection, multiple points are driven in a short time interval and points a, b, c, and d of the mapping function are obtained based on the recorded time, the distance traveled at the speed maintained in the time interval, and the left and right offsets of the vehicle, thereby obtaining a complete mapping function.

[0024] S4, selecting a point after the current target driving point in the road curve image as a detection point; Figure 2 As shown, the point that needs to be measured is selected as the detection point T.

[0025] S5. Obtain the target time of the detection point according to the road mapping function; obtain the target time corresponding to the detection point T by the position of the detection point T in the road curve image in the road mapping function. Figure 4When obtaining the target time corresponding to the detection point T, the matching degree of the detection point T is also tested. Specifically: combined with the road pattern detection, driving until all N points are detected, that is, the function is completed and then the point T is detected and matched with the point C within the line of sight.

[0026] S6. Calculate the actual distance between the current target driving point and the detection point according to the time and speed of the current target driving point and the target time.

[0027] Embodiment 2 Please refer to Figure 5 A route identification and planning device in patrol path planning includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, each step in the above-mentioned vehicle autonomous driving visual ranging learning method is implemented.

[0028] In summary, the present invention provides a vehicle automatic driving visual ranging learning method, which generates a preset number of driving points in a driving image and obtains a road curve image based on all the driving points; at the same time, the time, distance and speed corresponding to each driving point are recorded during the vehicle driving process, and then the road mapping function corresponding to the road curve image is obtained according to the data corresponding to each driving point; then, after selecting a detection point in the road curve image, the target time of the detection point is obtained based on the road mapping function, and finally the actual distance is calculated through the time and speed of the current target driving point and the target time, and the ranging is realized by relying on the visual ranging technology, thereby improving the ranging accuracy.

[0029] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent transformations made using the contents of the present invention's specification and drawings, or directly or indirectly applied in related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A method for learning visual ranging for vehicle autonomous driving, characterized in that: include: Acquire a driving image, generate a preset number of driving points along the driving direction in the driving image, and obtain a road curve image according to all the driving points; Record the time, distance and speed of each driving point; Obtaining a road mapping function according to the road curve image and the time, distance and vehicle speed corresponding to all the current driving points; Selecting a point after the current target driving point in the road curve image as a detection point; Obtaining the target time of the detection point according to the road mapping function; The actual distance between the current target driving point and the detection point is calculated according to the time and vehicle speed of the current target driving point and the target time.

2. The method for learning visual distance measurement for automatic driving of a vehicle according to claim 1, characterized in that: Generating a preset number of driving points along the driving direction in the driving image comprises: Obtaining the point spacing according to the current vehicle speed; All the driving points are arranged in sequence along the driving direction according to the point spacing.

3. The method for learning automatic vehicle vision distance measurement according to claim 2, characterized in that: The step of obtaining the point spacing according to the current vehicle speed includes: S=(V*t0) / N; Among them, S represents the point spacing; N represents the number of driving points; V represents the current vehicle speed; and t0 represents time.

4. The method for learning visual distance measurement for automatic driving of a vehicle according to claim 1, characterized in that: The road mapping function is obtained according to the road curve image and the time, distance and vehicle speed corresponding to all the current driving points, including: Obtaining an initial mapping function according to the road curve image; The time, distance and vehicle speed data corresponding to all the driving points are brought into the initial mapping function to generate the road mapping function.

5. The method for learning visual distance measurement for automatic vehicle driving according to claim 4, characterized in that: The obtaining of the initial mapping function according to the road curve image comprises: The initial mapping function is obtained by converting the road curve image horizontally and vertically and extending it by 45 degrees.

6. A vehicle automatic driving visual ranging learning method according to claim 4 or 5, characterized in that: The initial mapping function comprises: at 3 +bt 2 +ct+d; Among them, when the road curve image is a straight line, a and b are 0, when it is a single curve, a is 0, and when it is an S-curve, both a and b are not 0.

7. The method for learning visual distance measurement for automatic vehicle driving according to claim 1, characterized in that: After obtaining the road mapping function according to the road curve image and the time, distance and vehicle speed corresponding to all the current driving points, the following further comprises: Get historical road mapping function; The road mapping function and the historical road mapping function are learned through a first deep learning model to optimize the road mapping function.

8. The method for learning visual distance measurement for vehicle automatic driving according to claim 1, characterized in that: Generating a preset number of driving points along the driving direction in the driving image comprises: Acquiring road pattern information and environmental information in the driving image; Converting the road pattern information and the environment information into a feature vector; The feature vector is processed by a second deep learning model to obtain the driving point.

9. A vehicle automatic driving visual ranging learning method according to claim 8, characterized in that: Obtain all previous processing results of the second deep learning model, and optimize the loss function corresponding to the second deep learning model according to all the processing results.

10. A route identification and planning device in inspection route planning, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, each step in the vehicle autonomous driving visual ranging learning method as described in any one of claims 1-9 is implemented.

Images