Particle Filter-based Video Object Tracking Method

Tracking moving object in video sequences is one of the key subjects in computer vision, which has a wide range of applications involving in video surveillance, human-computer interaction, robot vison navigation and intelligent traffic control. Recently, particle filter is often employed in tracking algorithms that has been proven to be a very effective framework for solving the non-linear/non-gaussian tracking problems. Particle filter is a kind of statistical simulation method based on the idea of recursive Bayesian estimation,which uses a set of weighed particles sampled randomly to approximate the posterior possibility density function. In this approach, the problem of object tracking can be regarded as the estimation of a state vector which characterizes the object in a state space.In this dissertation, the research is focused on a single target tracking problems in the complex environments. The target can be defined as rapidly changing deformable non-rigid object in gray or color video sequences. For gray case, the work is mainly focused on the design of a robust observation model with effective measurement. For color case, the work is focused on the tracking problems of target's fast moving, temporary occlusions, rotation and scale invariant.The main contributions of this dissertation are summarized as follows:1. A novel approach in combination with H_∞filter and particle filter is proposed. An H_∞filter has superior performance in prediction when the system parameter has uncertain perturbation or unknown noise sources. It ensures that the system satisfies some performance criteria if the perturbation is the most serious, yet cannot ensure that the system has better performance than a Kalman filter with an appropriately designed process noise covariance matrix when the perturbation is small. In the proposed method, the H_∞filter is used to estimate a new proposal distribution in particle fiter, and draws the sample particles from the new proposal distribution instead of the prior transitional possibility density function. This approach effectively improved the accuracy of estimates compared with standard particle filter.2. For gray target tracking problem, a new template-match based randomized template tracker is developled in this dissertation. This approach combines the classical intensity-correlation-matching method with particle filter, using the robust maximum matching pel count distance as the similarity criterion to design the observation model. This proposed tracking algorithm can track target robustly in temporary occlusions situation.3. For color target tracking in complex environments, a new reliable and stable tracking algorithm using not only color information but also texture information and motion edge information is proposed, which is called multi-cue fusion based adaptive mean-shift particle filter. This method incorporates those multiple cues into particle filter, and selects proper fusion weight for each cue, then uses resample scheme to deal with degeneracy problem. When occlusion occurs, the particle filter stops updating and the particles are propagated by system model only. When the target moves fast, the mean-shift important sample step replaces the nomal important sample step in the filter. Considering the target's deformation, two scale variables which represent x direction and y direction respectively are added into the state vector. Experimental results show that the proposed methotd has a better robust than single-cue tracking algorithms.