Video Object Tracking And Behavior Recognition Using Sequential Bayesian Methods

Visual object tracking and behavior analysis is a fundamental topic in computer vision, andis widely applied to surveillance, robotics, human machine interface, video retrieval, digital enter-tainment, object-based video coding, etc. However, it remains a great challenge to achieve robusttracking and intelligent inference of underlying target status. This thesis is aimed to solve theseproblems within the framework of sequential Bayesian theory, where the major innovations areenumerated as follows:1. Popular tracking methods, such as particle filter and mean shift, can not achieve the tradeoffbetween accuracy and efficiency, which limits their usefulness in online tracking tasks undercomplex environment. Therefore, a novel algorithm - CamShift guided particle filter (CAMS-GPF) - is proposed to track object in video sequence by combining the relative strengths ofparticle filter and CamShift in tracking robustness and efficiency. CamShift is incorporatedinto the probabilistic framework of particle filter as an optimization scheme for proposal dis-tribution. Meanwhile, in the context of particle filter, the scale adaptation of CamShift isimproved and the computation complexity is reduced.2. The techniques of multi-target tracking and active tracking are still under developed, althoughthey bare great practical significance in real applications. Motivated by such demand, wedesign and implement a real time surveillance system, which is capable of actively trackingand detecting multiple targets with multiple features. In order to simultaneously track multipletargets, we apply Markov random field to model their interactions. A new mechanism ofcamera pose adjustment is put forth so that a good viewpoint of target can be maintained asit moves. We extract the joint feature of color, Gabor coefficient and motion to model targetappearance and the tracking robustness is further improved.3. The goal of video tracking is to provide baseline information for the intelligent analysis of ob-ject behaviors. Most existing statistic models for temporal behavior rely on the stationary as- sumption, and may fail to describe those non-stationary activities in real world due to widenedsemantic gap. A novel time sequence model - time varying hidden Markov model (TVHMM)- is proposed by extending traditional HMM to time varying scenario. In TVHMM, the statetransition density is explicitly allowed to change as the time spent in a particular state passesby. The temporal correlation between these transition densities is exploited by applying ahierarchical Dirichlet prior, which leads to a more robust model formulation. Markov ChainMonte Carlo (MCMC) sampling is employed to get the MAP estimate of time varying param-eters.The above-mentioned works are tested thoroughly on various real video sequences, syntheticscenarios and numerical simulations. It is demonstrated that the proposed CAMSGPF outperformsstandard particle filter or mean shift-based trackers in both tracking robustness and computationalefficiency. Our surveillance system works reliably regardless of the difficulties such as fast motion,partial occlusion, and multiple objects. The proposed model of TVHMM is shown to achieve higherrecognition rate in comparison to other stationary HMM-based methods, especially on date set ofbasic human actions such as walking, running and jumping.