Research On Visual Multi-Target Tracking In Complex Scenes Based On Discriminative Classifier

In recent years,video surveillance is used widely in the field of security monitoring,especially in some occasions requiring high security,such as school,airport and subway,etc.A video surveillance system is made up several modules including moving object detection,target tracking,behavior detection and abnormal information filing.Where multi-target tracking is applied widely and is a research hotspot in the field of computer vision,human-computer interaction system and pattern recognition,etc.However,there are still many to be studied in multi target tracking.Multi target tracking algorithm are summarized both at home and abroad in the paper,and they advantages and disadvantages are analyzed,mainly including target vision presentation model and the research of multi-target tracking algorithm.Moreover,the difficulties of multi target tracking algorithm are summarized,finally,the future development trend of multi target tracking algorithm is discussed.The main innovative work and research results are as follows:1)Firstly,due to the problem of poor reliability resulted from target vision presentation model based on single feature describing accurate the target image hardly under the complex scenes,such as illumination change,direction change and scale change,a multi feature fusion algorithm based on Fuzzy Theory(FT)and Principal Component Analysis(PCA)is proposed in the paper.Firstly,the color feature,texture feature and gradient feature of the target are extracted.Moreover,features are chosen crudely by FT.Finally,features after chosen crudely are fused by PCA to reduce algorithm complexity effectively.Compared to signal feature,the fusion feature is better on accommodating scene change,and more robust tracking results are obtained.2)The establishment of the discriminative classification model theory based on temporal and spatial constraints.Aiming at the problem of discriminative classification model not using temporal and spatial relations among multi target in the multi target tracking algorithm based on discriminative classification model at present,a discriminative classification model based on temporal and spatial relations and structured SVM is established according to the spatial relations among multi targets in the same frame and the motion relationship of target.Finally,the parameters of discriminative classification model are updated by Passive Aggressive(PA).3)The multi target tracking algorithm based on discriminative classification model with temporal and spatial constraints.Multi targets are tracked by discriminative classification model with temporal and spatial constraints.Firstly,the HOG feature of every target region is extracted,and combines with structured SVM to conduct sample training and to obtain appearance model of every target.Moreover,spatial relations among targets are established by using minimal spanning tree(MST)model in the same frame,and the time constraint is built based on motion relationship of target.In the process of track,parameters are studied online by using structured SVM frame,and appearance models of every target as well as constraint of space and time are updated in real time.Compared to the multi target tracking algorithm only based on space constrain,the method proposed in the paper fully uses relation of space and time of targets,experimental results on a common data set show that the method possess a certain improvement.