Research On Visual Target Tracking Based On Machine Learning

In the intelligent video detection system,the correct identification and tracking of the target vehicle is the main function of the intelligent transportation system,in the machine learning framework,as the most basic part of video surveillance,target detection and tracking.The problem of target tracking is regarded as the two classification problem in machine learning,and the suitable feature and the appropriate classification pattern are found by machine learning,and then the target tracking is achieved.There are several main problems in target tracking: One: moving objects in the scene interference.Two: occlusion.Three: shadows.Four: target’s style change.To solve the above problems,the introduction of machine learning methods,through the improvement of the traditional algorithm,improve the real-time target tracking,accuracy and robustness,this paper is based on the advantages and disadvantages of existing algorithms,to improve the comprehensive performance of the algorithm for target tracking.The main research work includes two aspects: the design and implementation of the two algorithms,one is the use of graph model,triangulation method of input image grid,target and background separation,To extract the target corner energy feature,feature vectorization as improved SVM classifier of positive and negative samples,the optimal parameters of the learning classifier according to the results of classification,target tracking.The two is based on the theory of deep learning,build a deep self code learning network,the target vehicle pixel level feature aggregation into structured features,training and classification of network weights,parameter tuning and the weight updating according to the follow-up test sample,the complex conditions and subject to tracking target occlusion accuracy significantly improved.