Research And Application Of Lane Line Detection Algorithm Based On Vision

Lane line detection has always been a classic topic in automatic driving system,which can be applied to the key links of subsequent lane departure warning and lane keeping to improve vehicle safety.At present,with the development of deep learning,the mainstream lane detection method has been transformed from image processing method based on computer vision to deep learning method.Due to the complexity of the environment in the real world and the different representation methods of lane lines in different scenes,the existing lane line detection schemes still cannot meet the requirements of real-time,accuracy and robustness.The research object of this topic is the unmanned vehicles driving in the park and the surrounding city roads.This paper proposes a lane-line detection scheme based on deep learning for its specific scene,and realizes the function of lane departure warning on this basis.The first is the production of data sets.In this paper,videos collected by unmanned vehicles are used to produce data sets.For different scenes,frame extraction processing is carried out respectively to meet the differences between different samples.According to different lane-line detection models,different lane-line data sets are used respectively.Aiming at the slow speed of semantic segmentation method,the target detection method was selected to detect lane lines,and a lane line detection scheme based on YOLOv4 was proposed.K-means++clustering algorithm is adopted to optimize anchor parameters of YOLOv4,so that the training network has certain pertinence in lane line detection.According to the distribution characteristics of lane lines in the aerial view of the road,the detection density is increased in the longitudinal axis direction,and S*2S grid is used to divide the road image area,so that the detection effect of lane lines,which are dense objects in the longitudinal axis direction,is better.Since the lane line detection scheme based on YOLOv4 does not use the lane line structure information,the detection accuracy is low in the case of road wear or occlusion,so a lane line detection algorithm based on row classification is proposed.On the basis of UFLD lane line detection algorithm,three steps of large,medium and small are selected to sample the position of the image line on the vertical axis of the road image,so as to solve the problem that there are few points of lane line in the distant part of the road image.The conic curve is used to restrain the structure of the lane line,so that it can detect the curve better.At the same time,Res Net and Mobile Net V2 are selected as feature extraction networks,which are respectively applied to different devices to solve the problem of slow running speed on embedded devices.Tensor RT is also used to accelerate the model processing to get faster detection speed.Finally,the least square method was used to fit the lane lines,and the lane departure warning module was completed to judge whether the vehicle deviated from the center of the lane.