Research On Detection Algorithm Of Lane Marking For Urban Road

At present,self-driving car has become one of the research hotspots in the field of artificial intelligence.And the research of advanced driving assistance system(ADAS)is of great significance for the realization of driverless driving.Traditional lane detection algorithms usually only recognize a single target scene,and need to manually set the corresponding parameters for feature extraction.However,in the complex environment of urban roads,lane detection is often affected by multiple factors such as light and shadow occlusion,making traditional algorithms less robust.Therefore,this thesis studies the lane detection in urban road scenes against this defect.The main contributions of the thesis are as follows:(1)Dual lane detection.A gray-scale transformation based on linear discriminant analysis(LDA)was proposed for feature extraction,and a chaotic perturbation particle swarm optimization(PSO)algorithm is used for lane line fitting.First,the high-dimensional RGB(red green and blue)color image is projected into the low-dimensional subspace through the optimal discriminative projection vector,and then the image is grayed out;then the chaos particle swarm optimization(CPSO)algorithm is used to traverse the space within the range of particle values according to the characteristics of the lane markings to find adapt the solution with the largest degree function;finally,the parameters of the straight line can be obtained according to the optimal solution.The experimental results show that the proposed algorithm can realize the lane detection function under various road conditions,and verifies that the algorithm has good robustness.(2)Multi-lane detection.First,an improved FCN(Fully Convolutional Networks)-based model is proposed for lane markings feature extraction.This neural network can implement pixel-level lane markings image classification.The parameters of the model were trained on two public datasets of Tusimple and Caltech Lanes;then,lane fitting is performed based on a hybrid model,and the fitting interval was mainly determined by the Hough transform.The least square method is used in the fitting interval to perform lane markings fitting.Experimental results show that the average accuracy of the algorithm on the Tusimple dataset is 98.74%,and the accuracy on the Caltech Lanes dataset is 96.29%.(3)In order to improve the real-time performance of multi-lane markings recognition,an improved model based on U-Segnet is proposed.This network uses different upsampling layers than FCN and has a smaller parameter volume.In addition,using the correlation between lane markings image frames,a long-short-term memory network(LSTM)is added in the middle of the feature extraction network.Experimental results show that the processing speed of the improved algorithm is 64% faster than that of the algorithm using FCN for feature extraction,which significantly improves the real-time performance of the algorithm for multi-lane detection.And the accuracy of the algorithm on the Caltech Lanes dataset is improved by 0.7 percentage point.