Research Of Vehicle Detection Based On Fusion Of Multi-view Classifiers

Vehicle detection subsystem is an important part of the intelligent auxiliary driving system.The traditional vehicle detection method detection performance is poor in the complicated traffic environment.A robust detection algorithm is proposed in this paper.The algorithm has good performance in different weather,strong light and other environments.In the part of vehicle candidate region determination,superpixels are obtained by Simple Linear Iterative Clustering(SLIC).Each superpixels block in the image is taken as a node of the graph model,and then clustered according to the spatial position relation,color similarity and other information between node and node to get a large number of vehicle candidate regions.According to the characteristics of the vehicle sample,Vehicle Proposal Score(VPS)of the vehicle candidate region is calculated by the similarity between the nodes.Finally,Non-maximum suppression(NMS)is used to eliminate the cross-repeated candidate region.Through the analysis of experimental results,vehicle candidate region determination proposed is proved to be accurate.In the part of car detection,image features such as image color and gradient are extracted to construct the fast pyramid model of multi-feature fusion.The fast pyramid model can enrich the image information and improve the efficiency of feature extraction.The AdaBoost algorithm is chosen as the classifier of vehicle detection by comparing the classification algorithms.In the part of training classifier,the training samples were divided into multiple classes according to the running angle of the vehicle,and the multi-view classifier was obtained.In order to test the performance of the vehicle detection algorithm,the comparative experiment in this paper is carried out in two public vehicle databases.Compared with the existing vehicle detection method,the vehicle detection algorithm proposed in this paper has better robustness and precision.For the situation that there is no public vehicle database in China,this paper establishes a Chinese traffic database.The establishment of this database enriches the positive samples of vehicle training set,which has application value and reference meaning to the research and development of vehicle detection algorithm.