Reseach On Lane Detection And Tracking Based On Deep Learning

The "Intelligent Vehicle Innovation and Development Strategy" released by the National Development and Reform Commission in 2020 pointed out that,with the driven by a new round of technological revolution and industrial transformation,the intelligent vehicles have become the strategic direction of the global automobile industry.As the main part of intelligent vehicles,the auto pilot system(APS)is a comprehensive system integrating many high-techs such as environmental perception,logical reasoning and decision-making,and motion control.In APS,environmental perception is a key link to ensure safety of intelligence vehicle.As one of the core technologies for vehicle environmental information acquisition,lane detection is an important part of guiding vehicles and route decision.Currently,the greatest challenge of lane detection is that higher vehicle speed means faster detection speed,meanwhile,ensuring detection accuracy.Therefore,how to improve the precision and efficiency of lane detection is the main purpose of this paper,and then the model is deployed on the embedded vehicle platform to verify practicability.The main contents of the paper are as follows:(1)The UFSD-T-S-Net model,based on the UFSD-Net(Ultra Fast Structure-aware Deep Network)model,is refined to solve the problem of high-efficiency and highprecision detection of lane in this paper.the UFSD-T-S-Net model can be divided into TNet and S-Net,which are used in different stages.By refining the local structure of ENet as the T-Net backbone module,and using the SAD(self-attention distillation)module to enhance the feature expression ability of the T-Net,and specially designing Classifier module,the detection accuracy of T-Net model in Tusimple reached 96.75 %,while the F1-measure in CULane reached 73.1.By using knowledge distillation to train S-Net model,the feature expression ability of S-Net is greatly improved.The experiment results show that,the detection accuracy of S-Net is 0.5% higher than UFSD-Net on the Tusimple dataset,and the F1-measure of S-Net is increased by 3.7 compared with UFSD-Net on the CULane dataset.At the same time,the detection effiency of S-Net is similar with UFSD-Net.(2)In order to overcome the problem of insufficient features information of singleframe image and difficult to deal with environmental mutation,this paper refines the SNet structure by merging the Res Net block features and the optical flow features which was extracted by Fast Flow Net’s CDDC module,and then the refined model is named as Fast Flow S-Net in this paper.In order to experimentally verify the Fast Flow S-Net model,the Tusimple dataset was reconstructed,and the experimental results show that the detection accuracy of Fast Flow S-Net model is 0.11% higher than S-Net.The S-Net-KF algorithm based on Lane key-points and KF(Kalman Filter)is refined for the purpose of comparing the Fast Flow S-Net with the traditional tracking algorithm.Then the experimental results show that the detection accuracy of Fast Flow S-Net is 0.08% higher than S-Net-KF.(3)In order to solve practical application of the S-Net model,the S-Net is deploed on the embedded vehicle platform based on the Tensor RT deep learning framework and multi-threaded framework in this paper.Then the S-Net’s performance is improved by optimizing the structure and quantifing the precesion of the S-Net model.The experimental results show that compared with the FP32 S-Net,the detection efficiency of the FP16 model is more than 130 FPS,which is about 3 times of the FP32.The detection accuracy is only reduced by 0.17 % on Tusimple dataset and the F1-measure is reduced by 0.2 on CULane dataset.