Research On Key Technologies Of Lane Detection And Tracking Based On Machine Vision

In recent years,in order to reduce road traffic accidents and ensure the safety of driving,the technology of Advanced driver assistance system(ADAS)in automobiles has become a research hotspot.As one of the key technologies of ADAS,the lane detecting and tracking provide the service of lane departure warning for drivers by detecting the lanes in real time.Lane detection is to extract the lanes from real-time road images collected by vehicle camera.Lane tracking predicts the position of lane in the next frame based on the lane parameters detected by the current frame.Aiming at the lane of the expressway,the paper uses the machine vision sensor to collect the road images and proposes the methods that combine machine learning and traditional image processing algorithms to detect and track lane in multiple actual road datasets.The main work and innovations of the dissertation are as follows:(1)A lane detection based on clustering and geometric constraint is proposed.Firstly,based on traditional image analysis techniques,the paper uses the Hough transform to detect all the lines in each lane area.Then,the k-means algorithm is used to classify the lines belonging to the same lane area.Finally,the lanes are fitted by integrating the parameters from the same type of lines.The proposed algorithm is simple and clear with low computational cost.(2)A lane detection based on convolutional neural network and connected components constraint is proposed.In order to solve the problem that traditional algorithms are difficult to overcome the disturbance of complex road environment,a convolutional neural network called Tiny-LaneNet is proposed for the detection of the lane's feature.It combines the connected components constraint and least square method to classify the lane feature points and to fit the lane,which can effectively improve the anti-interference ability and average detection accuracy of the lane detection.The results of the experiments show that the average detection accuracy is improved by 2.23% compared with the traditional algorithms.(3)Research on lane track based on kalman filter and LSTM.Based on the convolutional neural network of lane detection,this paper constructs a lane tracking model on Kalman filter or LSTM by using the spatial continuity of the lanes from forward and backward road images.It can effectively solves the problem that the lane detection fails due to the blurring of the lane or the occlusion of the surrounding vehicles.The experimental results show that the average detection accuracy of the algorithm using tracking is improved by 1.32% compared with the algorithm without using tracking.A large number of experiments has been carried on Xi'an Jiaotong University TSDmax datasets,Tusimple Benchmatk datasets,Baidu ApolloScape datasets and selfcollected datasets.The experimental results show that the proposed algorithm can effectively detect and track the lanes with good generalization ability and detection accuracy.