Object tracking is a challenging problem in computer vision. The task of object tracking is to determine the evolving states of objects according to measurement sequences. The key to a successful object tracking depends on an effective extraction of the useful information about the object state from measurement sequences. In this thesis we study nonlinear and non-Gaussian recursive Bayesian estimation problems in discrete time. Our research interest in these problems stems from the problem of tracking non-rigid visual multi-objects with a monocular camera regardless of background clutter, camera motion and frequent mutual occlusion between targets. The main contributions of the thesis can be concluded as follows. (1) The t-distribution based particle filter (SPF) is developed in this thesis. The SPF approximates the predictive and posterior distributions by t-distribution, respectively, and the mixture of the predictive distribution and prior distribution is chosen as the importance sampling distribution. In the case of the limited samples, MLEs of t distributions are computed by the ECME algorithm based on the regularized Mahalanobis distance. In the non-rigid object tracking and maneuvering object tracking, the performances between UKF, SISR and SPF are compared. (2) In the problem of tracking multiple non-rigid objects, especially in the case that the number of visual objects varies, how to update and maintain the multi-modality target distributions is crucial. To surmount the difficulty, a multi-object mixture tracking approach is developed based on modeling target distribution as finite t-distribution mixture models, which are termed as mixtures of t-distributions particle filters (MTPF). Due to a variable number of objects, two object detection algorithms are adopted to adjust mixture distributions, and add/remove the corresponding particle subsets. The Importance Sampling distributions in the MTPF consist of the probability transition model and prediction distribution of individual objects. The similarity between the sample and the reference is evaluated by the sample-based similarity measure. Comparisons between MTPF, MPF and BPF show the advantages and limitations of the new method over the football match sequence.
修改评论