The Research Of Tracking Based On Online Boosting

This paper focuses on tracking motion targets in video image sequences,and mainly researches and analyses online boosting algorithm based on statistical learning tracking method.The algorithm regards the motion target tracking problem as a kind of two classification problem of target and background,which is mainly to find a decision boundary which can distinguish target and background effectively,then use this decision boundary to get a specific location to track the target.The key idea is integrated learning,that is,to integrate several weak classifiers and classify the new samples together,so as to obtain better classification effect than single weak classifiers.Based on the robustness of the tracking algorithm,the paper analyzes the tracking drift problem of the traditional online boosting algorithm when the target is short period partial occluded and the target is moving quickly,Then,we propose a solution that makes full use of the background information around the target to assist in tracking,and adds a detector to the tracking framework to ensure that the target is recaptured when it is lost.Specific innovations include:Firstly,an online boosting motion target tracking algorithm based on spatial and temporal context information is proposed to solve the tracking drift problem when the target is short period partial occluded.Firstly,for each frame,a confidence function is obtained by using the spatial and temporal context model studied in the previous frame and the focus characteristics of the biometrics in the current frame,the space-time context model is a model built by using Bayesian principles for information about time context and space context.Then,Using the confidence function and the strong classifier in the boosting algorithm to evaluate each rectangle block in the current frame search area and return the rectangle block is the confidence value of the target rectangle block.Finally,the two confidence matrices are recombined linearly into a new confidence matrix according to the voting rights,looking for the maximum confidence value in the new confidence matrix,the corresponding rectangle block is the target rectangle block to track.After finding the target to track,the two classifiers are updated online with the current frame image,and the voting weights of the two classifiers are updated with the calculated occlusion factor.Experimental results show that the algorithm can track the robustness of the target when it is short period partial occluded.Secondly,an online boosting algorithm based on offline learning is proposed to solve the problem of tracking drift when the target is moving fast.Firstly,based on the combination of tracking and detection of TLD algorithm,this paper introduces the detector based on HOG feature of offline learning and the nearest neighbor detector of online learning into the tracing process to realize the combination of online learning and offline learning.When the tracking device fails or the confidence value of the target rectangle block is too low,the detector is used to research the target rectangle block in the whole image frame,and the target rectangle block tracked by the tracker is used to locate the current target rectangle block.Secondly,in order to improve the training update speed of classifier,the Haar feature used in the traditional online boosting algorithm is modified to HOG feature.Experimental results show that the algorithm can track the robustness of the target when it is moving fast.