Research On Lane Detection Algorithm In Complex Road Scene

Lane detection is a basic but necessary task in automatic driving.It is an important indicator to guide vehicles,which directly affects the direction control and path planning of vehicles.Therefore,in order to ensure the safety of vehicles,it is particularly important to accurately identify the lane line.Although the past few years have witnessed significant progress in lane marking extraction by the deep learning model,it is still challenging in extreme cases,such as strong light changes in complex road scenes,lane lines with low definition,severe occlusion,etc.,and hinders the commercial application of automatic driving technology in natural road scenes.Therefore,two algorithms are proposed in this paper to enable it to quickly and accurately detect lane lines in complex road scenes.The main specific contents are as follows:(1)A new lane feature extraction module,multidirectional slice feature aggregator(MSFA),is proposed.For the sake of illustration,the network is called MSFANet.In the vertical and horizontal directions,slice features with different steps are used to collect information to avoid information loss,and the computational efficiency is improved.And add diagonal direction,which is divided into two directions: main diagonal and anti diagonal.Extracting features from diagonal direction can better deal with occlusion than traditional convolution layer.In the decoder stage,a coarse-grained and detailed bilateral up sampling decoder is used.The method is tested on two popular lane detection benchmarks(CULane and Tusimple),and the results showed that the F1-measure value of lane line detection results has increased under various complex road scenarios in CULane data sets,with the total F1-Measure score reaching 74.8%.Its speed is 164 fps.On the Tusimple dataset,the F1 value reached 96.57%.(2)In order to further improve the detection accuracy,a new lane detection network MSFANet-GAN is proposed based on the proposed MSFANet algorithm and the generation of confrontation network.The two models of GAN network,i.e.generation model and discrimination model,can help to learn the advanced image features of the line.When some lane lines are only partially available or completely blocked,the lane lines will also continue along without fracture.The MSFANet-GAN network consists of two parts,M-S and M-F.Finally,the method is tested on the CULane data set of lane detection.The F1 value reaches76.8% and the speed reaches 131 fps.The F1 value reached 96.81% on the Tusimple dataset.