A two-lane line detection method based on vision system

Abstract

The invention discloses a double-lane line detection method based on a vision system. For a training structure, the marked key points of the lane line are processed by multi-point interpolation to obtain the key points with appropriate density; the position channel is added to the picture on the basis of three channels; Compression and feature extraction are carried out on images using a 7-layer 3x 3 or 1 x 1 convolution kernel and / or pooling operation; Up-sampling and multi-scale prediction are carried out on features; tensor characteristics of nx1x1x2 are output by convolution operation; The point coordinates are fitted into two curves by cubic spline interpolation algorithm, and two lane lines are obtained. The invention has the advantages of high semantic parsing ability, sufficient feature extraction and high accuracy of the result.

Description

technical field [0001] The invention relates to a double-lane line detection method, in particular to a vision system-based double-lane line detection method. Background technique [0002] At present, the existing lane line detection methods adopt methods based on feature extraction (Roberts operator, Sobel operator, Prewitt operator, Krisch edge operator, Gauss-Laplacian operator) and methods based on neural network models. (Baseline, ReNet, DenseCRF, MRFNet, ResNet-50, ResNet-101, SCNN). Among them, the method based on feature extraction has natural disadvantages in terms of edge continuity, edge smoothness, edge refinement, edge positioning, noise resistance, etc., especially the huge disadvantage in noise resistance restricts the reliability of the automatic driving vision system. sex. Although the methods based on the neural network model alleviate the above problems to a certain extent, most of the methods are GPU-level, which cannot meet the real-time performance of...

AI Technical Summary

Problems solved by technology

[0003] (1) The detection accuracy needs to be improved: the existing methods are accurate in detecting lane lines, especially in complex scenarios including traffic congestion, night, large turns of vehicles, discontinuous lane lines, many shadows, and missing data. will be greatly reduced

[0004] (2) Low detection speed: The detection speed of existing methods is mostly at the GPU level. In the field of automatic driving, the detection method at the GPU level obviously cannot meet the requirements of real-time detection. The engineering requirements for lane line detection must be at least at the CPU level. Detect multiple images per second so that the vehicle can react quickly when special situations arise

[0005] (3) The ability to analyze the semantic information of pictures is poor: the method based on feature extraction has natural disadvantages in terms of noise resistance

Due to the poor ability to analyze the semantic information of the picture, the existing methods more use the external feature processing of the picture to obtain other useless lane lines except the left and right lane lines.

Images