Research On Lane Detection Algorithm In Complex Environment Based On Machine Vision

With the development of the automobile industry,the automatic driving technology plays an increasingly important role in the safe driving of vehicles,and lane detection is an important part of the automatic driving system.Because the road environment is complex and changeable,the existing lane detection algorithm based on machine vision is vulnerable to illumination changes,pedestrian occlusion and road damage,resulting in weak robustness and reduced accuracy of the algorithm.Traditional lane detectionalgorithm based on Lane feature morphology can meet practical requirements under certain conditions because it does not need a lot of labeling data.Deep learning has the ability of automatic learning target,and its application in lane detection can improve the robustness and accuracy of detection.In this paper,lane detection algorithms based on feature morphology and depth learning are studied and optimized respectively,and the performance of the algorithm is verified by experiments.The main contents of this paper are as follows:1)Improve the detection algorithm based on Lane shape feature,and optimize the algorithm.Firstly,image preprocessing technology is analyzed,including image ROI division,image graying,median filtering to remove image noise,adaptive segmentation threshold is obtained by OTSU binarization,and then Sobel operator is used to extract image edges.For the traditional Hough transform and least square method,the improved algorithm is proposed,and the improved algorithm is combined to fit the lane curve.The Kalman filter is used to track the lane line to solve the problem of lane jitter and short-term missing.Finally,combined with the improved Hough and least squares algorithm,the algorithm flow of lane detection is designed.The test results show that the detection accuracy of the optimized lane detection algorithm is improved by 3.8% and the operation time is reduced by 40%.2)Collecting and labeling lane data sets.Real vehicle captured road images in different scenes,and proposed a new marking method to mark the lanes in the images.By extracting regions of interest and downsampling from images,and merging data sets of different annotation formats with script functions,the data sets needed for network training are obtained.3)Designing lane detection algorithm based on instance segmentation method,and verifying it with real vehicle.The algorithm consists of lane segmentation network L-Net and lane fitting network H-Net.Based on the improved VGG-16 framework,L-Net network extracts lane lines from images.The network includes two branches: lane line segmentation and lane line clustering.Finally,it realizes the instance segmentation of lane line pixels and solves the problem of multi-lane detection.The H-Net network further fits the lane line pixels to get the parameterized lane line curve.H-Net network acquires variable perspective transformation parameters through training of neural network to transform the image perspective,so as to solve the problem of inaccurate lane detection on uphill and downhill.Finally,the algorithm is tested.The detection speed is 55.6 ms/frame,and the average detection accuracy is 96.6% in different scenarios.The performance of the instance segmentation algorithm is better than that of the improved feature-based algorithm,and it meets the real-time and robust requirements of detection.