Research For Dynamic Crowd Target Tracking Method Based On Deep-learning

The research and application of target tracking method,which as an important branch in the field of computer vision,now is widely used in science and technology,national defense construction,aerospace,and the various fields of national economy.So the target tracking technology has significant practical value and broad space for development.At present,based on video technology under the large crowds of pedestrians in target tracking is a hotspot issues for researching.The main conventional methods have frame differential method,optical flow method,template matching,kalman filter etc.But facing up to crowd and complextracking problem,the effect is not ideal as in a simple environment significantly.The main reason is that the number of people is more,in addition strong correlation of the person and the person or people with complex environment,together with coupling and dynamicity are difficulties,which makes the classified problems challenging and having more research value.To solve above problems,this thesis proposes a dynamic crowd target tracking method based on deep-learning.Specifically,the core of this study is using deep learning model to do training for image samples,and get the characteristic description.But before the learning,we need to do pixel division first,which makes the current image being segmented into smaller pieces,then generate no labels of samples.After completing above processing,the foundation is made for follow-up tracking,then the tracking targets are set as samples of identifying,which is different from the backgroud.This thesis use multiple Support Vector Machine(SVM)to design the classifier.Immediately following,sample collection will be updated base on the existing recognition results and tracking results,which is used for later loop trace.Here the sample size will be adopted according to the determination of dynamic adaptive algorithm.At the end of the thesis,we conducted a series of experimental analysis to evaluate the method.Experimental results shows that the proposed tracking method can meet the requirements of various application scenarios and guarantee its reliability,and the accuracy is significantly superior to the traditional tracking algorithm.