Reserch Of Vehicle Tracking And Road Scene Understanding Based On Machine Learning

Intelligent transportation system integrates traffic engineering,communication technology,computer technology and other technologies to provide convenience for personal mobility.In future,intelligent transportation system can reduce traffic congestion and the rate of traffic accident.The demand for intelligent transportation makes vehicle tracking and road scene understanding become very meaningful.We research on the two parts respectively.In the aspect of vehicle tracking,aiming to deal with the unstable characteristics of current vehicle tracking algorithms in dealing with illumination changes,a vehicle tracking algorithm based on color features and local binary pattern features is proposed by using traditional machine learning method.The color features and local binary pattern(LBP)features of samples are extracted.The positive and negative samples are classified by structured output support vector machine(SVM).The predicted samples are imported into the classifier.Candidates corresponding to the predicted samples with the highest predicted score are the locations of the target vehicle.The simulation experiment are carried out in several video sequences involving illumination changes,motion blur,fast motion and other challenges.The results show that the proposed algorithm is robust to illumination challenges.In the aspect of 2D road scenes,in order to solve the problem of incomplete boundary information restoration in current semantic segmentation algorithms,Skip connection network(SCNet)and multi-scale dilation network(MDNet)for road scene segmentation are proposed by using deep learning method.the former can make full use of the combination of VGG16 feature extraction and deconvolution to obtain the target features,and achieve better segmentation results fortargets that pay more attention to traffic scenes,such as traffic lanes,cars andbicycles;the latter can make full use of the multi-scale features of the target and fuse them.In combination,pixel-level annotation can be achieved by introducing the predictor to small objects with fewer clues,such as trees and poles.In the aspect of 3D road scene,In order to solve the problem of information scarcity of RGB image in 2D road scene,double input network(DINet)is proposed,which adds disparity image to the input.Compared with similar 2D network,it has higher precision and efficiency of semantic segmentation.