Lane Keeping System Based On Deep Learning In Complex Working Conditions

As the core function of the ADAS system,the research on lane-keeping assistant system is especially important with the development strategy of intelligent vehicles.The research mainly includes two parts;one is the recognition of road conditions based on lane detection,the other is the vehicle lateral control strategy.However,the existing lane detection technology relies on traditional machine vision,which can not solve the problem that lane information is easily affected by illumination and occlusion;the impact of camera body motion on information recognition is not considered by most Lane detection techniques.Furthermore,lane-keeping system research has not integrated vision system and lane-keeping system.Most of the study is at the verification stage based on the software simulation.Therefore,this paper focuses on "lane identification under complex conditions" and "lane-keeping lateral control technology" to achieve human vehicle collaboration control based on machine vision learning.The specific content of the study is as follows:Firstly,accurate lane feature extraction is the key to lane detection.Based on the endto-end lane detection framework of deep learning,this paper presents a feature-enhanced point instance segmentation method for lane detection.The feature layer formed by feature enhancement retains more lane semantic information.Furthermore,the clustering problem of lane prediction feature points is transformed into the point instance segmentation problem to make the detection network is not limited by the number of fixed lane,which reduces the estimation of unnecessary prediction points and increases the accuracy of lane identification.The lane detection network is tested on the Tusimple dataset and compared with the PINet network.The results show that the accuracy of the lane detection network is improved by 2.7% and the loss function converges faster,the rate of miss detection and false detection is reduced to a certain extent,and the detection speed is 25 fps.Secondly,lane identification errors caused by the camera moving with the body should be considered.The principle of camera imaging is analyzed,and the primary camera calibration is completed;then,the additional Euler transformation correction matrix is established by using pitch angle,roll angle,and yaw angle generated during vehicle movement.In this paper,the correction matrix is incorporated into the transformation matrix between the pixel coordinate system and world coordinate system,a dynamic correction model is established and the lanes are then fitted using the RANSAC algorithm.Then,based on LQR and feedforward control,the lane keeping lateral control strategy is designed.Firstly,the dynamic model of tracking error is established based on the twodegree-of-freedom dynamic model in the Frenét coordinate system with the desired path as the reference line,and a reasonable tire model is adopted.LQR problem is constructed,the weight matrix is designed,the feedback control rate is solved,the steady-state error is analyzed,the feedforward control rate is designed to eliminate steady-state error,and the path tracking the performance of the control algorithm is verified based on Car Sim and Simulink.Finally,a hardware-in-the-loop dynamic simulation system is built to verify the accuracy of the lane-keeping control system.In the case of 40km/h and 80km/h,the verification results show that with the combination of depth learning lane detection,the lane-keeping control algorithm proposed by LQR and feedforward control can track the Navigation Lane well and meet the functional requirements of LKAS.