Research On Lane Detection Technology Under Poor Vision

Lane detection technology is one of the key technologies in the development of intelligent transportation technology.Advanced assisted driving technology including lane detection has been pre-installed on many new vehicles,which can meet some basic needs.In normal road scenes,many technologies can achieve good detection results,but the actual road scenes include lane wear,rainy weather,glare and other factors.Many scholars in the industry have also turned their attention to the detection in such scenes.The deep learning method shows higher accuracy and robustness in the field of lane line detection,which makes it possible to detect lane lines in more scenes.Therefore,based on deep learning,this paper attempts to realize an accurate and real-time lane detection technology for the scene under bad visual conditions.For this kind of lane detection,this paper mainly starts from expanding the receptive field and introducing the Self-Attention mechanism.The lane detection algorithm based on spatial CNN uses the backbone network Large FOV and the improved SCNN＿DURL is used to learn image features,and a new lane detection model is proposed by introducing Self-Attention mechanism.Experiments show that the trained model has an improvement of up to 3.39 points in the five scenes of road congestion,road glare,road shadow,turning road and night road in the CULane data set.The main work of this paper is as follows:(1)Self built small-scale data set to supplement the bad scene of open-source data set,because it is a 3-lane data set with low resolution,which is convenient for small-scale training and verification in the process of model improvement.It is found that the effect of many models decreases obviously in the lane fuzzy scene,so it is necessary to study this kind of scene.(2)Proposes to use atrous convolution in SCNN＿DURL,which expands the receptive field without increasing parameters;The Encoder is embedded in the feature layer of convolutional neural network to establish long-distance dependence from different dimensions and capture the global key context information.Experiments show that this method can improve the detection of five difficult scenes in CULane dataset.(3)A bilateral parallel update is proposed to replace the sequential slice update of original SCNN＿DURL.Experiments show that the time-consuming of the improved model is slightly lower than that of the original spatial CNN.