Studies Of Object Tracking Algorithms

This thesis has conducted comprehensive discussions and researches on object tracking algorithms without losing deep insights. The first chapter expounded the research back-ground of object tracking as well as its significance, discussed its prospect for applications on the civil and military affairs, displayed its rich and attractive value in practice; In the first chapter, the status quo of object tracking research at home and abroad was also dis-cussed, and various existing tracking algorithms were categorized by their latent ideas into two classes:Top-Down and Down-Top, and were expounded detailedly in simple language; The last part of first chapter exposed two problems which had prevailed in object tracking technical world:1) How to represent the object effectively; 2) How to learn the changes of object appearance correctly and timely, brought out the focus of research of this thesis:1) introduce new technical tools into tracking system; 2) understand background by the way of we understanding object; 3) Embed priori knowledge. The second chapter focused on the feature point based tracking algorithms; Firstly, It expounded concisely the characteristic of local feature points, such as richness, distinctiveness and robustness; Then, It gave a brief sketch of SURF tracking algorithm, and discussed relative algorithms and their differences with SURF tracking; In the next paragraphs, it displayed algorithm details of SURF track-ing, explained the feature motion generative model which plays a core role in the framework, demonstrated the online EM algorithm for learning model parameters, and also talked over the problem of occlusion. Substantial experiment results showed in this chapter verified the outstanding performance of SURF tracking algorithm, specially in dealing with various chal-lenging cases. The correctness of SURF tracking's mechanism for active occlusion detection was also proved in the experiment part. The third chapter aimed to discuss level set based tracking algorithms; In the beginning it introduced theory of level set and its numerical com-puting methods, and talked about the existing level set based tracking algorithms and their capabilities; In the next level set embedded ensemble tracking algorithm, as the core of this chapter, was described detailedly. Experiments for this chapter highlighted the capability of level set method to precisely represent object contour.Sum up, the main study achievement and innovation of this thesis could be concluded as the following points: 1. By their latent ideas, we categorized various existing tracking algorithms into two classes:Top-Down and Down-Top; Under the proposed classification criteria, we dis-cussed a great number of prevailing tracking algorithms, analyzed their common points as well as differences, and also demonstrated their failure situations and reasons.2. Innovatively, we proposed SURF tracking algorithm; Its latent idea is that:estimat-ing object global motion parameters by motion observations of local feature points of object; In the SURF tracking framework, the relationship between object local feature points'motions and object global motion was depicted by generative model, and an online EM algorithm was employed to learn model parameters. Meanwhile, an active occlusion detection mechanism was proposed and verified in the experiments.3. We improved the classical ensemble tracking, proposed level set embedded ensemble tracking algorithm; As a high dimension representation of contour, level set function had great power in describing evolution process of contour; Level set embedded en-semble tracking algorithm seamlessly integrated level set function and ensemble track-ing together, made them promoted by each other; It was proved in the experiments that tracking accuracy and robustness was greatly improved by the new algorithm.