Single Visual Object Tracking Under Complex Scenarios

Visual object tracking is a research hotspot in the computer vision field. It has been applied with great success to a variety of application areas including intelligent surveillance, medical applications, human-machine interaction and sports analysis. Although many promising visual tracking algorithms have been proposed which greatly enhanced the development of visual ob-ject tracking, the existing tracking methods cannot handle the practical application difficulties due to the increasing complexity and challenging property of the visual scenarios and more and more difficulties for the tracking algorithms to deal with. These challenging scenarios may in-clude:severe appearance changes induced by severe illumination change, abrupt motion induced by sudden dynamic change, and so on. Researchers have been working hard to design efficien-t and effective algorithms to deal with the tracking scenarios. Under this circumstance, this dissertation aim to solve visual object tracking difficulties in complex scenarios using Bayesian tracking framework combining the particle filters and Markov chain Monte Carlo(MCMC) sam-pling methods. The main contributions of this dissertation are summarized as follows:(1) We study the particle filtering algorithm and its application in visual object tracking. A robust particle tracker using MCMC posterior sampling and second-order Markovian assump-tion is proposed in order to solve the problems encountered in traditonal particle tracker. This algorithm is based on the particle filtering framework, which is based on second-order Markov assupmtion, that is, the current state relates to the previous two historical states. The posteri-or probability density is the joint density of the two states. We also substitute the traditional importance sampling method with a MCMC sampling method, which can avoid the sample im-poverishment problem encountered in traditional particle filter based tracking algorithms.(2) Abrupt motion tracking problem is intensively studied in this thesis. Stochastic sam-pling based tracking algorithms are widely used in solving abrupt motion tracking problem. We proposes an efficient visual tracking algorithm based on the Hamiltonian Markov chain Monte Carlo method (HMCMC) in order to sovle the abrupt motion tracking difficulties. The proposed algorithm uses the Hamilton dynamics to simulate a Markov Chain, which can to some extends suppress the random walk behavior in traditional MCMC based tracking methods. The HMCM-C tracker can avoid being trapped in local minimum during searching the state space, leading to successfully capturing the abrupt motion. The proposed algorithm can effectively track the object with different kinds of abrupt motions.(3) This thesis proposes a robust abrupt motion tracking algorithm using adaptive ordered over-relaxation MCMC sampling method, aiming at suppressing the random walk behavior in traditonal MCMC tracking algorithms. The ordered overrelaxtion method is introduced in the Gibbs sampling stage of the HMC based tracking framework which contains random walk be-havior. The iterations needed for searching the promising object area can be reduced. We also proposed to adaptively adjust the step length of leapfrog in order to reduce the simulation error of Hamilton dynamics. The proposed algorithm can track the object with abrupt motion effec-tively. It also shows superior performance in dealing with partial or full occlusion as well.