Research On Video Object Tracking Based On Feature Information

With the rapid development of social economy and popularity of information technology,object detection and tracking based on vision have become a hotspot.No matter military missile tracking technology or civil video surveillance system,they are all closely linked with the computer vision technology.In this thesis,we do some theoretical and practical research on target feature matching and tracking.The main work of the thesis includes the following aspects:(1)The thesis does some research on target tracking based on motion analysis and target tracking based on image matching.For the former,firstly,the thesis analyses the advantages,disadvantages and applications of three basic target detection algorithms: background difference method,inter-frame difference method and optical flow method.The thesis comes to a conclusion: the inter-frame difference method is the most appropriate for the scene of real-time tracking.Then,the thesis compares the Gaussian Mixture Model and the CodeBook model.The experimental results of the comparison show that both algorithms can detect targets,but the speed of the CodeBook model is faster than the other.For the latter,firstly,the thesis clarifies the technical requirements and difficulties of target tracking.Secondly,the thesis divides the target tracking problem into four parts: feature extraction,motion model,observation model and model update.Finally,the thesis analyses some suitable features and motion models for this topic.(2)The thesis proposes an improved tracking algorithm based on the Compressive Tracking(CT)algorithm.Firstly,the thesis dissects the theory and practice flow of the CT algorithm.The CT algorithm can track accurately and fast,but cannot adapt to the change of target scale.In order to solve the problem,the thesis uses the method of block tracking to improve the original CT algorithm.The method can evaluate the situation of target's scale change and adapt to bounding box's scale to track the target more accurately.Experimental results about tracking scale changing targets show that the proposed tracking algorithm is better than the CT and the Multiple Instance Learning(MIL)algorithms in terms of both the success rate and the center error rate.And in most cases,our algorithm is also better than the Fast Compressive Tracking(FCT)algorithm.(3)The thesis proposes a fusion algorithm of Major Color Spectrum Histogram and component information.The algorithm is used in a PTZ camera tracking system.The system needs to choose the main target every once in a while.When the system detects two targets,it extracts the Major Color Spectrum Histogram features of the two targets,comparing to the previous record of the main target.Then it chooses the current main target.If the above steps can not choose the main target,the algorithm adopts LBP features of the component information and the SVM classifier to choose the main target in this frame.Experimental results show that the matching accuracy of our algorithm is higher than that of the MCSH method.