Research On Visual Object Tracking In Particle Filter Framework

Visual tracking is an important research issue in computer vision, and also is important branch in machine intelligent. It has been widely used in automatic driving, video surveillance, human-machine interaction, post producing of films, virtual reality, etc. Particle filter framework, which is based on Bayesian filter and Monte Carlo sampling, is the most widely used architecture in visual tracking. Most modern visual tracking methods are under the particle filter framework, since particle filter is robust in dynamic system with nonlinear and multimodality. In general, particle filters can be decomposed into two steps: the predicting and the updating phases. In the predicting phase, the most import thing is the state transition model, and it predicts next state based on current state in a Bayesian manner. In the updating phase, the focus is the likelihood between the predictions and the target, whose aim is amending the errors of the predictions and the true position and appearance of the target. These two phases run alternately. If they can accurately robust collaborate of each other, the tracking would get better performance.In addition to the advantages of particle filter, there are some shortcomings of it, such as more particles sampling from the distribution for better performance of the tracking, and the high computational needs of estimating the likelihood. Based on the discussion above and considering the state-of-the-art tracking methods, several models and algorithms are proposed, which are as follows.Firstly, an adaptive particle filter with a velocity-updating based transition model and an adaptive state estimation model is proposed to improve the results of visual tracking.In the adaptive particle filter, the motion velocity embedded into the state transition model is updated continuously by a recursive equation, and the state estimation is obtained adaptively according to the state posterior distribution. The experiment results show that the proposed model can increase the tracking accuracy and run efficiently in complex environments.Visual tracking based on sparse representation is one of the most popular visual tracking methods; its essence is estimating the observed likelihood in the sparse representation model. It has been demonstrated that these methods can reduce the effects of illumination and occlusion. However, on one hand, these methods need to solve a complex L1minimization optimal problem, and cannot meet the needs of real-time applications; on the other hand, they need the templates of the target in advance, which cannot capture the dynamic changes of the environments. As a result, two more models are proposed.Secondly, a new L1tracker by clustering particles via improved K-means algorithm is proposed to reduce the high computaional complexity. Sparse representation is one of the most popular visual tracking methods and has been demonstrated robust to illumination changes and occlusion. However, it needs to solve a complex L1minimization optimal problem, and cannot meet the needs of real-time tracking. After having been clustered, the particles that close to each other and similar to the latest target template are selected out as important particles for L1minimization to determine the target state of current frame. The importance particles selection keeps the reliability of particles and reduce the also largely promotes the tracking stability. The good performance shows that the proposed method promotes the tracking accuracy and efficiency simultaneously.Thirdly, a robust L1tracking based on online logistic discriminate learning is proposed after analyzing the distribution of the particles sampled by particle filtering and their difference with the true state of moving targets. The discrimination of the moving target and backgrounds is enhanced by the discriminating and updating procedures of logistic discriminate function, at the same time, the proposed L1tracking is more efficiently compared to the original L1tracking as particles with little relevancy are rejected in advance of the L1optimization. Experiments demonstrate that, the proposed tracking method obtains robust results with a very lower running time compared with original L1tracker.The proposed tracking methods focus on tracking single target, which can also be used for multiple targets tracking by combining with data association methods. In addition, the studies of this thesis are still having deep influence on other related applications of computer vision.