A method for evaluating lane line detection performance

Abstract

The application discloses a kind of lane line detection performance evaluation methods, in this evaluation method, the true value mark of lane line is established, the position of lane line in physical world is sampled and true value mark in a certain length range of lane, and the lane line of benchmark mark is obtained.In real time collection is carried out to lane line on intelligent car installation camera, obtains lane line video, and extracts key frame in lane line video, and carries out lane line detection.The lane line detected is based on world coordinate system Data conversion, least square method is used to obtain the fitting parameter of lane line, and the fitting of lane line is carried out.The center coordinate position of fitting lane line and the center coordinate of lane line obtained by benchmark mark are calculated deviation value, and the stability of fitting parameter is calculated.The lane line detection performance evaluation dimension of the application is richer, which breaks through the evaluation from the angle of segmentation and classification, so that the lane line detection performance evaluation method is more comprehensive and more scientific.

Description

TECHNICAL FIELD

[0001] The present application relates to a lane line detection performance evaluation technology, belonging to the field of automatic driving or intelligent driving. BACKGROUND

[0002] At present, the unmanned driving technology is developing rapidly, and the requirements for lane line detection are also increasing, requiring higher accuracy and higher real-time performance. In actual roads, the lane conditions are complex and changeable, including the diversity of lane line shapes, visual disturbances caused by obstacles, shadows, fog, light intensity changes, image blurring caused by vehicle speed, image quality affected by light changes, object occlusion, stains, or serious deformation of the acquired marking line due to changes in the camera shooting angle, which will all affect the detection of lane lines and thus the accuracy of lane line recognition. In actual situations, road conditions present different states, so a set of lane line detection system can only robustly detect lane lines in a certain road condition, and may not be applicable in other road environments, that is, the same system may have a large difference in processing results under different environments. Therefore, in scientific research or company product development, the evaluation of lane line detection performance is very important, and an effective lane line detection performance evaluation method needs to be developed.

[0003] Chinese patent 201911239409.8 discloses a lane line detection method and system, which includes the following steps: acquiring a road image in front of a vehicle; preprocessing the road image in front of the vehicle to obtain a preprocessed image; processing the preprocessed image using a pre-trained deep learning CNN network to obtain lane line position point information, and fitting each lane line according to the lane line position point information; binarizing the preprocessed image to obtain a binary image, and determining lane line color information according to the binary image and the lane line position point information; performing image threshold segmentation on the preprocessed image after grayscale to obtain a threshold segmentation image, and determining lane line solid information according to the threshold segmentation image and the lane line position point information.

[0004] Chinese patent 202180000475.9 provides a lane line detection method and device, which can be applied in the field of automatic driving or intelligent driving. The method includes: acquiring echo data of the ground including lane lines using a radar; preprocessing the echo data to obtain first data; filtering the first data to obtain second data, which is data with an amplitude less than a first threshold; and performing imaging processing on the second data to obtain a synthetic aperture radar (SAR) image of the lane lines. The lane line detection method and device can obtain a high-resolution SAR image of the lane lines, effectively improving the accuracy of lane line detection.

[0005] Current lane detection technologies primarily rely on segmentation or classification-based evaluation methods, which only assess whether detected pixels are on the lane lines, neglecting the degree of deviation of the detected lane line positions and the stability of the fitted parameters. Summary of the Invention

[0006] Based on the above background technical analysis, most current lane line evaluation methods are based on segmentation or classification methods, which only evaluate whether the detected pixels are pixels on the lane line, but ignore the degree of deviation of the detected lane line position and the evaluation of parameter stability.

[0007] The technical method designed in this invention focuses on the degree of lane line position deviation and the stability of parameter fitting during the evaluation process, including: calculating the lane line center deviation, average deviation, number of lane lines within a certain deviation range, and variance of parameters.

[0008] The technical solution adopted in this invention is a method for evaluating lane line detection performance, which includes the following steps:

[0009] Step 1: Mark the baseline true value of the lane lines;

[0010] Before evaluating the detection performance of lane lines, establish the ground truth labeling of lane lines; within a certain length of the lane, collect points and mark the ground truth of the lane lines in the physical world to obtain the benchmark lane lines, which are used to compare the accuracy with the detected lane lines during the evaluation process.

[0011] Step 2: Evaluation of lane line detection performance;

[0012] Step 2.1: Real-time data collection of lane markings;

[0013] Cameras are installed on smart cars to collect lane line data in real time, acquire lane line videos, and extract key frames from the lane line videos for lane line detection.

[0014] Step 2.2 Lane line fitting;

[0015] The detected lane lines are transformed based on the world coordinate system, and the least squares method is used to obtain the fitting parameters of the lane lines, and then the lane lines are fitted.

[0016] Step 3: Calculation of lane line detection deviation;

[0017] Step 3.1 The fitted lane line in Step 2.2 has a certain width. Calculate the center coordinates of the fitted lane line.

[0018] Step 3.2 Calculate the deviation value of the lane line center coordinate fitted in step 3.1 and the lane line center coordinate obtained by the reference marking in step 1.

[0019] Step 3.3 Sum all lane line center deviation values obtained in step 3.2.

[0020] Step 3.4 Average the lane line center deviation sum obtained in step 3.3.

[0021] Step 3.5 Count the number of lane lines whose deviation values are within a certain range.

[0022] Step 4 Stability calculation of fitting parameters;

[0023] Step 2.2 The stability of the fitting parameters of the lane line is evaluated.

[0024] Further, the reference true value marking process of the lane line is as follows: the lane line is drawn and dotted by RTK integrated inertial navigation, and the true value marking of the lane line is formed based on the drawing and dotting.

[0025] Further, the fitting process of the lane line in the evaluation process of the lane line detection performance is as follows:

[0026] The video of the lane line real-time collection contains N frames, the current frame is the qth frame, N is a natural number, 1≤N, 1≤q≤N. The relevant point coordinates of each lane line in the current frame and the previous a frame images are obtained from the video, and the relevant point coordinates are converted to the world coordinate system, wherein a is a natural number, 1≤a≤q. In the world coordinate system, for each lane line, the least squares method is used to perform curve fitting on the relevant point coordinates of the current frame and the previous n frames, and the fitted curve includes t parameters; the fitted parameter set of the qth frame is recorded as is an element of the fitted parameter set. The lane line fitting is also performed on the reference marking points of the Nth frame.

[0027] Further, the calculation process of the lane line detection deviation in the evaluation process of the lane line detection performance is as follows:

[0028] The entire lane line real-time collection video contains N frames, the current frame is the qth frame, N is a natural number, 1≤N, 1≤q≤N. Each lane line equation of the current frame is vertically sampled at an interval of 1m, and there are M times of sampling within the lane line output range (the starting point to the ending point), and each sampling is performed horizontally offset calculation. For the i-th vertical sampling point, the calculation is as follows:

[0029] (1) Calculation of lane line center coordinate;

[0030] Get left line coordinate and right line coordinates The center coordinates of this lane line are calculated as

[0031] Read the left line coordinates of the reference marking of the lane line of the i-th longitudinal sampling point and right line coordinates The center coordinates of the lane line reference marking are 1≤i≤M.

[0032] (2) Calculation of lane line center deviation;

[0033] Calculate the Euclidean distance between two points p i and

[0034]

[0035] (3) Calculation of lane line center deviation and

[0036] The lane line center deviation value is

[0037] (4) Calculation of the average value of the lane line center deviation of the video;

[0038] The average value of the lane line center deviation of the video is

[0039] (5) Statistics of the number of lane lines in the range (c, d) of the lane center deviation;

[0040] If d i ∈(c,d) and the current lane line has not been counted, then s=s+1. s represents the number of lane lines in the range (c, d) of the lane center deviation, and the initial value is 0.

[0041] Further, the stability calculation process of the fitting parameters in the evaluation process of the lane line detection performance is as follows:

[0042] The stability of the fitting parameters is measured by calculating the standard deviation of the fitting parameters.

[0043] The standard deviation of the parameters is {V1, V2, …, V j , …, V t}

[0044] Where, 1≤j≤t, The number of lane lines in the range (c, d) of the lane line center deviation of the video is s.

[0045] ​Compared with existing technologies, the lane detection performance evaluation method of this invention has richer dimensions, breaking through the limitations of evaluating solely from the perspectives of segmentation and classification. This makes the lane detection performance evaluation method more comprehensive and scientific, with the following advantages:

[0046] (1) An evaluation method based on distance deviation was invented. This method breaks through the limitations of simply evaluating whether a pixel is a pixel on the lane line, and incorporates the degree of deviation of the lane line position into the evaluation method, including: calculating the lane line center deviation, average deviation, and the number of lane lines within a certain deviation range.

[0047] (2) A method for evaluating the stability of fitting parameters was invented. This method designs a method for evaluating the stability of lane line fitting parameters. Stable fitting parameters are of great reference significance for ensuring the stable and safe operation of vehicles. Attached Figure Description

[0048] Figure 1 The video used in Example 1 is shown in the figure.

[0049] Figure 2 A schematic diagram of tracing and marking lane lines.

[0050] Figure 3 This is a flowchart illustrating the implementation of this method. Detailed Implementation

[0051] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0052] Example 1

[0053] like Figure 1 As shown, the video uses lane line video data from a 2MP 60-degree forward-looking camera on the Donghai Bridge. The video was edited into segments, each 2 minutes long, and keyframes were extracted. Due to the high similarity between adjacent frames, one video image was extracted every 5 frames, with a resolution of 2304×1296. The lane line dataset contains a total of 3000 images.

[0054] Example 2

[0055] like Figure 2 As shown, the true value of the lane line is generated by using the RT-3002 trolley device to trace and mark the lane lines.

[0056] Example 3

[0057] The coordinates of the relevant points on the lane line are fitted with a cubic curve using the least squares method, and the equation is:

[0058] y = r0 + r1x + r2x 2 +r3x 3

[0059] r0, r1, r2, r3. r0 represents the lateral distance from the vehicle head to the target lane line, r1 represents the heading angle of the lane line, r2 represents the second order coefficient of the lane line world coordinate system, and r3 represents the third order coefficient of the lane line world coordinate system.

[0060] Example 4

[0061] Assuming that a lane line detection image contains only one lane line, in the image pixel coordinate system, the vertical direction contains 5 sampling points, the lane line coordinates of the reference mark and the lane line coordinates detected, the lane line center coordinates and the center deviation values are shown in Table 1:

[0062]

[0063] From the data in Table 1, the minimum value of the lateral deviation is 1, the maximum value of the lateral deviation is 10, the sum of the lateral deviation is 26, the average value of the lateral deviation is 5.2, the standard deviation of the lateral deviation of the lane line is 3.12, and the number of lane lines with lateral deviation in the range of (15, 20) is 0.

[0064] Example 5: The parameters in Table 2 are fitted for 5 times, and the mean, variance and standard deviation of the parameters are calculated.

[0065] C0 C1 C2 C3 First -1.483 0.001 0.046 0.231 Second -1.497 0.011 0.116 0.346 Third -1.488 0.015 0.089 0.194 Fourth -1.511 0.006 0.101 0.218 Fifth -1.489 0.009 0.134 0.257 Mean -1.4986 0.0084 0.0972 0.2492 Variance 0.3698 0.000093 0.0028 0.0648 Standard Deviation 0.608 0.0096 0.0529 0.2545

Claims

1. A method of evaluating lane line detection performance, characterized by: The evaluation method comprises the following steps, Step 1: lane line reference true value labeling; Before evaluating the detection performance of the lane line, the true value labeling of the lane line is established; within a certain length range of the lane, the position of the lane line in the physical world is sampled and true value labeled to obtain the reference labeled lane line, which is used for accuracy comparison with the detected lane line in the evaluation process; Step 2: evaluation of lane line detection performance; Step 2.1: real-time collection of lane lines; Install a camera on the intelligent vehicle to collect lane lines in real time, obtain lane line videos, and extract key frames in the lane line videos for lane line detection; Step 2.2: lane line fitting; Convert the detected lane line into data based on the world coordinate system, obtain the fitting parameters of the lane line using the least squares method, and perform lane line fitting; Step 3: calculation of lane line detection deviation; Step 3.1: the fitted lane line in step 2.2 has a certain width, and the center coordinate position of the fitted lane line is calculated; Step 3.2: calculate the deviation value of the center coordinate position of the fitted lane line in step 3.1 and the center coordinate of the lane line obtained in step 1; Step 3.3: sum all the lane line center deviation values obtained in step 3.2; Step 3.4: average the lane line center deviation sum obtained in step 3.3; Step 3.5: count the number of lane lines within a certain range of deviation value from the lane line center deviation value obtained in step 3.2; Step 4: calculation of the stability of the fitting parameters; Step 2.2: stability evaluation of the fitting parameters of the lane line; The calculation process of the lane line detection deviation in the evaluation process of the lane line detection performance is as follows: The entire lane line real-time collection video contains N frames, the current frame is the qth frame, N is a natural number, 1≤N, 1≤q≤N; each lane line equation of the current frame is sampled vertically at an interval of 1m, a total of M times within the lane line output range, and a horizontal offset calculation is performed each time; for the i-th vertical sampling point, the calculation is as follows: (1) calculation of the lane line center coordinate; acquire left lane coordinate and right lane coordinate calculate the center coordinate of this lane line as a left edge line coordinate of a reference marking of a lane line of the ith longitudinal sampling point and a right edge line coordinate a center coordinate of the lane line reference marking is 1≤i≤M; (2) calculation of the lane line center deviation; Compute the Euclidean distance of two points p i and p (3) calculation of the lane line center deviation sum; The lane line center deviation value is (4) calculation of the average value of the lane line center deviation of the video; The lane line center deviation value average value of the video is (5) statistics of the number of lane lines within the range (c, d) of the lane center deviation; If d i If ∈(c,d) and the current lane line has not been counted, then s = s + 1; s represents the number of lane lines whose center deviation is in the range (c,d), and the initial value is 0.

2. The method of claim 1, wherein: The lane line reference true value labeling process is as follows: the lane line is drawn and dotted by RTK integrated inertial navigation, and the true value labeling of the lane line is formed based on the drawn and dotted lane line.

3. The method of claim 1, wherein: The lane line fitting process in the evaluation process of the lane line detection performance is as follows: The lane line real-time collection video contains N frames, the current frame is the qth frame, N is a natural number, 1≤N, 1≤q≤N; the relevant point coordinates of each lane line in the current frame and the previous a frames are obtained from the video, and the relevant point coordinates are converted to the world coordinate system, a is a natural number, 1≤a≤q; in the world coordinate system, for each lane line, the relevant point coordinates of the current frame and the previous n frames are used to perform curve fitting by the least squares method, and the fitted curve includes t parameters; The set of fitted parameters for the qth frame is recorded as is an element of the set of fitted parameters; The lane line reference labeling points of the Nth frame are also fitted.

4. The method of claim 1, wherein the stability of the fitting parameters is calculated as follows during the evaluation of the lane line detection performance: The stability of the fitting parameters is measured by calculating the standard deviation of the fitting parameters. The number of lane lines in the video whose lane line midline deviation is in the range (c, d) is s. The standard deviation of the parameters is {V1, V2,..., V j ,..., V t} where 1 < j < t, ​

Images