A method for collecting environmental information for autonomous vehicles on public roads

Through drone companion technology, real and complete information about the surrounding environment of autonomous vehicles is collected on public roads, solving the problem of limitations in the perception range of traditional perception systems, and improving the authenticity and repeatability of the test.

CN114155511BActive Publication Date: 2024-11-26JILIN UNIVERSITY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111513620.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-12-13
Publication Date
2024-11-26
Estimated Expiration
2041-12-13

AI Technical Summary

Technical Problem

The existing environmental information collection methods of autonomous vehicles on public roads cannot effectively obtain the dynamic changes of traffic participants, resulting in limited perception range and low accuracy, making it difficult to obtain real and complete environmental information in public road tests.

Method used

Use drone to test vehicles for a long time, keep synchronized through the drone positioning system and control system, use the onboard environment information collection equipment to collect video frames, and use image pixel coordinate transformation, lane line detection, pedestrian and vehicle targets. Detection and tracking, obtain three-dimensional world coordinates and trajectory information, and realize more comprehensive environmental information collection.

Benefits of technology

It improves the authenticity and completeness of environmental information, avoids the limitations of sensor perception range, enhances the repeatability of public road tests and the accuracy of performance evaluation, and can be used for input in virtual test scenarios.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114155511B_ABST
    Figure CN114155511B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for collecting environmental information for self-driving cars on public roads. The method comprises: using a drone to accompany the test vehicle for a long time and steadily, using the drone's onboard environmental information collection equipment to collect road information ahead, preprocessing the collected video frames, and performing lane line inspection, pedestrian target detection and tracking, and vehicle target detection and tracking on the preprocessed video frames. The present invention obtains road information around the vehicle at a higher viewing angle and a larger range, effectively avoiding the situation where the vehicle's own sensor's perception range is not far enough and is blocked by road obstacles and other traffic participants. While ensuring the authenticity of the test, more complete public road scene information can also be obtained; the obtained data can be analyzed offline to obtain more accurate performance evaluation; the obtained public road scene information can be used as a scene input for virtual testing, and the relevant performance of the self-driving car can be repeatedly tested, thereby enhancing the repeatability of public road testing.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for collecting environmental information of a road, and in particular to a method for collecting environmental information of an autonomous driving vehicle on a public road. Background Art

[0002] With the development of intelligent assisted driving technology, the intelligence of automobiles is constantly improving, and the testing needs are also more diverse. The testing methods of autonomous driving vehicles mainly include virtual tests conducted in virtual environments such as model-in-the-loop testing, driving simulator testing, hardware-in-the-loop testing, and vehicle-in-the-loop testing, as well as real car tests conducted in closed test sites and public roads. Virtual testing has low testing costs and strong repeatability, but its test authenticity depends on the authenticity of the model, which is relatively low in authenticity; although closed site testing is more realistic, it lacks the dynamic change elements of other traffic participants; public road testing is consistent with the actual road environment of autonomous driving vehicles, and is the most realistic testing method, but its repeatability is poor. At the same time, the testing and evaluation links of public road testing occur in real time, and it is difficult to discover some evaluation details that are easy to overlook.

[0003] At present, there are two technical routes for intelligent vehicle environment perception technology. One is a pure visual perception solution dominated by cameras, with Tesla as a typical representative; the other is a multi-sensor fusion perception technology solution dominated by lidar and supplemented by other sensors, with typical representatives from companies such as Baidu and Google. However, both existing methods perceive the surrounding environment from the perspective of the vehicle. Due to the limitations of perspective, sensor perception principle and perception range, they can only perceive unobstructed objects within a certain range, and are powerless for obstructed objects or distant objects. Therefore, the scene obtained by reproducing the perception information of the perception system is not complete and has a low accuracy rate.

[0004] Based on this, it is necessary for autonomous driving vehicles today to use a method for collecting environmental information under public road testing to obtain real dynamic changing factors of traffic participants in order to evaluate the relevant performance of autonomous driving vehicles in the most realistic and complete manner. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method for collecting environmental information of an autonomous vehicle on a public road. The vehicle is tested by a drone flying with it for a long time. The collection steps are as follows:

[0006] (1) The UAV positioning system uses the real-time dynamic carrier phase differential positioning (RTK) method to accurately locate the UAV. The UAV positioning system outputs the UAV position information as the input information of the UAV control system;

[0007] (2) The UAV control system receives the position information of the test vehicle and the UAV from the test vehicle positioning system and the UAV positioning system, determines the control strategy of the UAV, and then issues control instructions to keep the UAV and the test vehicle in a synchronous flight state over a long period of time;

[0008] (3) During the flight of the UAV and the vehicle, the UAV's onboard environmental information collection device is used to collect road surface information ahead, and a video frame containing road surface environmental information is obtained;

[0009] (4) calibrating and dedistorting the camera, preprocessing the video frames acquired in step (3), and transforming the image pixel coordinates into three-dimensional world coordinates;

[0010] (5) Perform lane line detection on the preprocessed video frames to obtain road surface structure information;

[0011] (6) Detect pedestrians: Based on the preprocessed video frames, the algorithm based on the histogram of oriented gradients (HOG) and the linear support vector machine (SVM) is used to detect whether there are pedestrians. When a pedestrian is detected, the global position coordinates Global_X, Global_Y and the timestamp Global_Time of the pedestrian are obtained, and a pedestrian identification number Pedestrian_ID is assigned when the pedestrian is first identified.

[0012] (7) Pedestrians are tracked. The pedestrian identification number of the same pedestrian will not change, thereby generating a time series of pedestrian positions and obtaining pedestrian trajectory information. The pedestrian speed v_Pedestrian and acceleration a_Pedestrian are obtained based on the pedestrian trajectory information;

[0013] (8) Detect vehicle targets. According to the preprocessed video frames, the YOLOv3 algorithm is used to detect vehicle targets. When a vehicle target is detected, the global coordinate position Global_X, Global_Y, vehicle length Vehicle_Length, vehicle width Vehicle_Width, vehicle type Vehicle_Type, timestamp Global_Time and current lane position Line_ID of the vehicle target are obtained, and a vehicle identification number Vehicle_ID is assigned when the vehicle target is first recognized;

[0014] (9) Track the vehicle to generate vehicle trajectory information, and then obtain the vehicle speed v_Pedestrian and acceleration a_Pedestrian.

[0015] Furthermore, in step (2), the method for the UAV and the test vehicle to maintain a synchronous flight state over a longer time span is as follows: the tangent direction of the lane is taken as the longitudinal direction, the normal direction of the lane is taken as the lateral direction, the UAV maintains a constant horizontal height, no or little lateral displacement is generated (lateral displacement < 0.1m), and the average longitudinal velocity of the vehicle in the previous period T (T> 30s) is obtained as the longitudinal velocity of the UAV in the current period T.

[0016] Furthermore, in step (3), the environmental information acquisition device is a camera image sensor carried by a drone.

[0017] Furthermore, in step (4), the relationship between the 3D world coordinates and the image pixel coordinates is as follows:

[0018]

[0019] The image pixel coordinate system is (u, v) and the world coordinate system is (X W , Y W , Z W ), the pixel unit length size is d x and d y , the pixel coordinates of the principal point are (u0, ν0), f is the normalized focal length, R is the rotation matrix, T is the translation vector, the intrinsic and extrinsic matrices on the right side of the equation are obtained by Zhang Zhengyou's chessboard calibration method, and the height H of the drone from the ground can be known from the drone positioning module in step (1), and Z is calculated C , determine the coordinate transformation relationship between pixel coordinates and world coordinates.

[0020] Furthermore, in step (5), the lane line detection step is specifically as follows:

[0021] (5.1) Convert the preprocessed video frame into a grayscale image: This step can directly take the mean or weighted average of the three color channel values ​​of the image;

[0022] (5.2) Perform Gaussian smoothing: Eliminate the influence of the inevitable noise and some unimportant small lines on the lane line detection effect when collecting video frames; for the size of the window template is (2k+1)×(2k+1), the calculation formula for the corresponding weights of each position in the template is as follows:

[0023]

[0024] H i,j represents the weight of the i-th row and j-th column based on the template center, σ is the variance, and k is obtained from the window template size. The obtained position corresponding weight is convolved with the grayscale image to obtain the image after Gaussian smoothing.

[0025] (5.3) Edge detection: Obtain the gradient g in different directions by dot multiplication with a Sobel or other operator x (m, n), g y (m, n), the comprehensive gradient value and gradient direction are calculated by the following formula:

[0026]

[0027]

[0028] G(m,n) is the comprehensive gradient value, θ is the gradient direction;

[0029] Then filter the edge points whose gradient values ​​are not the largest in the gradient direction, and finally use the upper and lower threshold method to detect the edge;

[0030] (5.4) Hough transformation: transform the points in the Cartesian coordinate system into the Hough space, select the points formed by as many straight lines as possible in the Hough space as the points transformed from the lane lines in the Cartesian coordinate system, and obtain the lane lines;

[0031] (5.5) Line extraction: The lane lines on the actual road will have many discontinuities due to dotted lines, wear and tear, and occlusion of obstacles. Therefore, the lane lines obtained by Hough transformation are subjected to line extraction to obtain complete lane lines.

[0032] Furthermore, in step (7), the process of tracking pedestrians uses the SORT algorithm, which is implemented by the target detection result of step (6) and the Kalman filter and Hungarian algorithm:

[0033] The Kalman filter algorithm uses a mathematical model to optimally estimate the state of the system and fuses the predicted value with the measured value. The algorithm is divided into two parts: state vector prediction and measurement update. The state vector prediction is used to predict the value of the current state vector and the state vector covariance matrix. The purpose of the measurement update is to update the state vector. The state vector prediction equation is as follows:

[0034] X=FX t-1

[0035] P=FP t-1 F T +Q

[0036] The measurement update equation is as follows:

[0037] K t =PH T [HPH T +R] -1

[0038] X t =X+K t (Zt -HX)

[0039] P t =(IK t H)P

[0040] Where X is the predicted state vector of the current state; F represents the transition matrix; X t-1 represents the state vector at the previous moment, P is the covariance corresponding to X; Q is the system noise covariance; K t is the current Kalman gain; H represents the conversion matrix from state variables to measurements; R is the observation noise; X t is the state vector at the current moment; Z t is the current observation value;

[0041] After the previous frame is predicted by the Kalman filter algorithm, multiple rectangular boxes will be predicted in the current frame. The task of the Hungarian algorithm is to match the rectangular box detected in the current frame with the rectangular box predicted by the Kalman filter to determine whether they are the same target, so as to achieve the purpose of tracking.

[0042] Furthermore, in step (9), the CamShift algorithm is used to track the vehicle. The algorithm tracks the existing detected vehicle results based on the color features of the moving vehicle in front. The color histogram of the vehicle area in the detection result is used as the matching basis. The color histogram of the image is calculated to obtain an area with a probability distribution similar to the tracked vehicle area, and the offset of the tracking area is calculated. After multiple iterative operations, when the offset is less than a threshold or the number of iterations exceeds a threshold, the tracking of the vehicle target area is completed.

[0043] Beneficial effects of the present invention:

[0044] The present invention aims to provide a method for collecting environmental information for use in public road tests of autonomous vehicles, which can obtain road surface information around the vehicle at a higher viewing angle and a larger range, effectively avoiding the situation where the vehicle's own sensor's perception range is not far enough and is blocked by road obstacles and other traffic participants. While ensuring the authenticity of the test, more complete public road scene information can be obtained; the obtained data can be analyzed offline to obtain more accurate performance evaluation; the obtained public road scene information can be used as a scene input for virtual testing, and the relevant performance of the autonomous vehicle can be repeatedly tested, thereby enhancing the repeatability of public road testing. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a scene schematic diagram of the present invention;

[0046] Figure 2 is a system architecture diagram of an embodiment of the present invention;

[0047] Figure 3 It is a coordinate transformation relationship diagram of the present invention;

[0048] Figure 4 This is a flow chart of lane line detection of the present invention. DETAILED DESCRIPTION

[0049] See also Figure 1-4 As shown:

[0050] The present invention provides a method for collecting environmental information of an autonomous vehicle on a public road, wherein a drone is used to accompany the vehicle for a long time and stably to test the vehicle, and the drone airborne system includes: a drone positioning module, a drone motion control module, an environmental information collection module, an environmental information preprocessing module, a lane line detection module, a pedestrian detection module, a pedestrian tracking module, a vehicle detection module and a vehicle tracking module;

[0051] The environmental information collection method includes the following steps:

[0052] (1) The UAV positioning module uses the real-time dynamic carrier phase differential positioning (RTK) method to process the difference of the carrier phase observations of the two measurement stations in real time. The positioning accuracy can reach the centimeter level to achieve accurate positioning of the UAV. The UAV positioning module outputs the UAV position information as the input information of the UAV control module;

[0053] (2) The UAV control module receives the position information of the test vehicle and the UAV from the test vehicle positioning module and the UAV positioning module, determines the control strategy of the UAV, and then issues control instructions to keep the UAV and the test vehicle in a synchronous flight state over a long period of time;

[0054] When the UAV flies in complete synchronization with the vehicle, its flight trajectory will be too complicated, resulting in too many changes in the scene boundary. The dynamic traffic participants at the scene boundary will appear and disappear in the collected video frames, causing unnecessary false detection and missed detection in the subsequent target detection and tracking process, and making subsequent scene reproduction more difficult. Therefore, in the UAV control module, a method of "making the UAV and the vehicle fly synchronously over a longer time span" is proposed: the tangent direction of the lane is the longitudinal direction, and the normal direction of the lane is the lateral direction. The UAV maintains a constant horizontal height, and generates no or little lateral displacement (lateral displacement < 0.1m). The average longitudinal velocity of the vehicle in the previous period T (T> 30s) is obtained as the longitudinal velocity of the UAV in the current period T.

[0055] (3) During the flight of the UAV and the vehicle, the airborne environmental information acquisition module of the UAV is used to collect the road surface information ahead, and obtain a video frame containing the road surface environmental information; the environmental information acquisition module is a camera image sensor carried by the UAV;

[0056] (4) using the environment information preprocessing module to calibrate and dedistort the camera, and preprocess the video frames collected in step (3) to transform the image pixel coordinates into three-dimensional world coordinates;

[0057] The relationship between 3D world coordinates and image pixel coordinates is as follows Figure 3 As shown, the coordinate transformation relationship is as follows:

[0058]

[0059] The image pixel coordinate system is (u, v) and the world coordinate system is (X W , Y W , Z W ), the pixel unit length size is d x and d y , the pixel coordinates of the principal point are (u0, ν0), f is the normalized focal length, R is the rotation matrix, T is the translation vector, the intrinsic and extrinsic matrices on the right side of the equation are obtained by Zhang Zhengyou's chessboard calibration method, and the height H of the drone from the ground can be known from the drone positioning module in step (1), and Z is calculated C , determine the coordinate transformation relationship between pixel coordinates and world coordinates.

[0060] (5) Use the lane detection module to perform lane detection on the preprocessed video frames to obtain road surface structure information, such as Figure 4 As shown in the figure, the lane line detection steps are as follows:

[0061] (5.1) Convert the preprocessed video frame into a grayscale image: This step can directly take the mean or weighted average of the three color channel values ​​of the image;

[0062] (5.2) Perform Gaussian smoothing: Eliminate the influence of the inevitable noise and some unimportant small lines on the lane line detection effect when collecting video frames; for the size of the window template is (2k+1)×(2k+1), the calculation formula for the corresponding weights of each position in the template is as follows:

[0063]

[0064] H i,j represents the weight of the i-th row and j-th column based on the template center, σ is the variance, and k is obtained from the window template size. The obtained position corresponding weight is convolved with the grayscale image to obtain the image after Gaussian smoothing.

[0065] (5.3) Edge detection: Obtain the gradient g in different directions by dot multiplication with a Sobel or other operator x (m, n), g y (m, n), the comprehensive gradient value and gradient direction are calculated by the following formula:

[0066]

[0067]

[0068] G(m,n) is the comprehensive gradient value, θ is the gradient direction;

[0069] Then filter the edge points whose gradient values ​​are not the largest in the gradient direction, and finally use the upper and lower threshold method to detect the edge;

[0070] (5.4) Hough transformation: transform the points in the Cartesian coordinate system into the Hough space, select the points formed by as many straight lines as possible in the Hough space as the points transformed from the lane lines in the Cartesian coordinate system, and obtain the lane lines;

[0071] (5.5) Line extraction: The lane lines on the actual road will have many discontinuities due to dotted lines, wear and tear, and occlusion of obstacles. Therefore, the lane lines obtained by Hough transformation are subjected to line extraction to obtain complete lane lines.

[0072] (6) Pedestrian detection module is used to detect pedestrian targets: Based on the preprocessed video frames, the algorithm based on the histogram of oriented gradients (HOG) and the linear support vector machine (SVM) is applied to detect whether there is a pedestrian target. When a pedestrian target is detected, the global position coordinates Global_X, Global_Y and the timestamp Global_Time of the pedestrian target are obtained, and a pedestrian identification number Pedestrian_ID is assigned when the pedestrian target is first recognized;

[0073] (7) Pedestrians are tracked by a pedestrian tracking module. The pedestrian identification number of the same pedestrian will not change, thereby generating a time series of pedestrian positions, obtaining pedestrian trajectory information, and obtaining pedestrian speed v_Pedestrian and acceleration a_Pedestrian according to the pedestrian trajectory information; the pedestrian tracking module uses a SORT algorithm, which is implemented by the target detection result of step (6) and the Kalman filter and the Hungarian algorithm;

[0074] The Kalman filter algorithm uses a mathematical model to optimally estimate the state of the system and fuses the predicted value with the measured value. The algorithm is divided into two parts: state vector prediction and measurement update. The state vector prediction is used to predict the value of the current state vector and the state vector covariance matrix. The purpose of the measurement update is to update the state vector. The state vector prediction equation is as follows:

[0075] X=FX t-1

[0076] P=FP t-1 F T +Q

[0077] The measurement update equation is as follows:

[0078] K t =PH T [HPH T +R] -1

[0079] X t =X+K t (Z t -HX)

[0080] P t =(IK t H)P

[0081] Where X is the predicted state vector of the current state; F represents the transition matrix; X t-1 represents the state vector at the previous moment, P is the covariance corresponding to X; Q is the system noise covariance; K t is the current Kalman gain; H represents the conversion matrix from state variables to measurements; R is the observation noise; X t is the state vector at the current moment; Z t is the current observation value;

[0082] After the previous frame is predicted by the Kalman filter algorithm, multiple rectangular boxes will be predicted in the current frame. The task of the Hungarian algorithm is to match the rectangular box detected in the current frame with the rectangular box predicted by the Kalman filter to determine whether they are the same target, so as to achieve the purpose of tracking.

[0083] (8) Use the vehicle detection module to detect vehicle targets. According to the preprocessed video frames, the YOLOv3 algorithm is used to detect vehicle targets. When a vehicle target is detected, the global coordinate position Global_X, Global_Y, vehicle length Vehicle_Length, vehicle width Vehicle_Width, vehicle type Vehicle_Type, timestamp Global_Time and current lane position Line_ID (the lane the vehicle is in) of the vehicle target are obtained, and a vehicle identification number Vehicle_ID is assigned when the vehicle target is first recognized;

[0084] (9) Track the vehicle using a vehicle tracking module to generate vehicle trajectory information, and then obtain the vehicle speed v_Pedestrian and acceleration a_Pedestrian. The vehicle tracking module uses the CamShift algorithm, which is based on the color features of the moving vehicle in front to track the existing vehicle detection results. The color histogram of the vehicle area in the detection result is used as the matching basis. The color histogram of the image is calculated to obtain an area with a similar probability distribution to the tracking vehicle area, and the offset of the tracking area is calculated. After multiple iterative operations, when the offset is less than a threshold or the number of iterations exceeds a threshold, the tracking of the vehicle target area is completed.

Claims

1. A method for collecting environmental information of an autonomous driving vehicle on a public road, characterized in that: The drone is used to fly the vehicle for a long time and steadily. The collection steps are as follows: (1) The UAV positioning system uses a real-time dynamic carrier phase differential positioning method to accurately locate the UAV. The UAV positioning system outputs the UAV position information as input information for the UAV control system; (2) The UAV control system receives the position information of the ego vehicle and the UAV from the ego vehicle positioning system and the UAV positioning system of the test vehicle, determines the control strategy of the UAV, and then issues control instructions to keep the UAV and the test vehicle in a synchronous accompanying flight state over a long time span; the method for keeping the UAV and the test vehicle in a synchronous accompanying flight state over a long time span is as follows: the tangent direction of the lane is taken as the longitudinal direction, the normal direction of the lane is taken as the lateral direction, the UAV keeps the horizontal height unchanged, no or little lateral displacement is generated, the lateral displacement is <0.1m, and the average longitudinal velocity of the ego vehicle in the previous period T is obtained as the longitudinal velocity of the UAV in the current period T, T>30s; (3) During the flight of the UAV and the vehicle, the UAV's onboard environmental information collection device is used to collect road surface information ahead, and a video frame containing road surface environmental information is obtained; (4) calibrating and dedistorting the camera, preprocessing the video frames acquired in step (3), and transforming the image pixel coordinates into three-dimensional world coordinates; (5) Perform lane line detection on the preprocessed video frames to obtain road surface structure information; (6) Detect pedestrians: Based on the preprocessed video frames, the algorithm based on the histogram of oriented gradients (HOG) and the linear support vector machine (SVM) is used to detect whether there are pedestrians. When a pedestrian is detected, the global position coordinates Global_X, Global_Y, and timestamp Global_Time of the pedestrian are obtained, and a pedestrian identification number Pedestrian_ID is assigned when the pedestrian is first identified. (7) Pedestrians are tracked. The pedestrian identification number of the same pedestrian will not change. Thus, a time series of pedestrian positions is generated to obtain pedestrian trajectory information. The pedestrian speed v_Pedestrian and acceleration a_Pedestrian are obtained based on the pedestrian trajectory information. The process of tracking pedestrians uses the SORT algorithm, which is implemented by the target detection result of step (6) and the Kalman filter and Hungarian algorithm: The Kalman filter algorithm uses a mathematical model to optimally estimate the state of the system and fuses the predicted value with the measured value. The algorithm is divided into two parts: state vector prediction and measurement update. The state vector prediction is used to predict the value of the current state vector and the state vector covariance matrix. The purpose of the measurement update is to update the state vector. The state vector prediction equation is as follows: X=FX t-1 P=FP t-1 F T +Q The measurement update equation is as follows: K t =PH T [HPH T +R] -1 X t =X+K t (Z t -HX) P t =(I-K t H)P Where X is the predicted state vector of the current state; F represents the transition matrix; X t-1 represents the state vector at the previous moment, P is the covariance corresponding to X; Q is the system noise covariance; K t is the current Kalman gain; H represents the conversion matrix from state variables to measurements; R is the observation noise; X t is the state vector at the current moment; Z t is the current observation value; After the previous frame is predicted by the Kalman filter algorithm, multiple rectangular boxes will be predicted in the current frame. The task of the Hungarian algorithm is to match the rectangular box detected in the current frame with the rectangular box predicted by the Kalman filter to determine whether they are the same target, so as to achieve the purpose of tracking; (8) Detect vehicle targets. According to the preprocessed video frames, the YOLOv3 algorithm is used to detect vehicle targets. When a vehicle target is detected, the global coordinate position Global_X, Global_Y, vehicle length Vehicle_Length, vehicle width Vehicle_Width, vehicle type Vehicle_Type, timestamp Global_Time and current lane position Line_ID of the vehicle target are obtained, and a vehicle identification number Vehicle_ID is assigned when the vehicle target is first recognized; (9) Tracking the vehicle, thereby generating vehicle trajectory information, and then obtaining the vehicle speed v_Pedestrian and acceleration a_Pedestrian; The CamShift algorithm is used to track the vehicle. This algorithm is based on the color features of the moving vehicle in front to track the existing detected vehicle results. The color histogram of the vehicle area in the detection result is used as the matching basis. By calculating the color histogram of the image, the area with a probability distribution similar to the tracked vehicle area is obtained, and the offset of the tracking area is calculated. After multiple iterative operations, when the offset is less than a threshold or the number of iterations exceeds a threshold, the tracking of the vehicle target area is completed.

2. The method for collecting environmental information of an autonomous driving vehicle on a public road according to claim 1, characterized in that: In step (3), the environmental information acquisition device is a camera image sensor carried by the drone.

3. The method for collecting environmental information of an autonomous driving vehicle on a public road according to claim 1, characterized in that: In step (4), the relationship between the 3D world coordinates and the image pixel coordinates is as follows: The image pixel coordinate system is (u, v) and the world coordinate system is (X W , Y W , Z W ), the pixel unit length size is d x and d y , the pixel coordinates of the principal point are (u0, v0), f is the normalized focal length, R is the rotation matrix, T is the translation vector, the intrinsic and extrinsic matrices on the right side of the equation are obtained by Zhang Zhengyou's chessboard calibration method. From the UAV positioning system in step (1), we can know the height H of the UAV from the ground, and calculate Z C , determine the coordinate transformation relationship between pixel coordinates and world coordinates.

4. The method for collecting environmental information of an autonomous driving vehicle on a public road according to claim 3, characterized in that: Step (5) Lane line detection steps are as follows: (5.1) Convert the preprocessed video frame into a grayscale image: This step can directly take the mean or weighted average of the three color channel values ​​of the image; (5.2) Perform Gaussian smoothing: Eliminate the influence of the inevitable noise and some unimportant small lines on the lane line detection effect when collecting video frames; for the size of the window template is (2k+1)×(2k+1), the calculation formula for the corresponding weights of each position in the template is as follows: H i,j represents the weight of the i-th row and j-th column based on the template center, σ is the variance, and k is obtained by the size of the window template; the obtained position corresponding weight is convolved with the grayscale image to obtain the image after Gaussian smoothing; (5.3) Edge detection: Gradients g in different directions are obtained by multiplying an operator x (m, n), g y (m, n), the comprehensive gradient value and gradient direction are calculated by the following formula: G(m,n) is the comprehensive gradient value, θ is the gradient direction; Then filter the edge points whose gradient values ​​are not the largest in the gradient direction, and finally use the upper and lower threshold method to detect the edge; (5.4) Hough transformation: transform the points in the Cartesian coordinate system into the Hough space, select the points formed by as many straight lines as possible in the Hough space as the points transformed from the lane lines in the Cartesian coordinate system, and obtain the lane lines; (5.5) Line extraction: The lane lines on the actual road will have many discontinuities due to dotted lines, wear and tear, and occlusion of obstacles. Therefore, the lane lines obtained by Hough transformation are subjected to line extraction to obtain complete lane lines.