Research On Key Technologies Of Particle Filter Based Visual Tracking And Application

Visual tracking is a hot issue in computer vision. And it is widely used in the fields of visual surveillance, sports video tactical analysis, intelligent human-computer interaction, and so on. By simulating human vision's motion sensing function, the objective of visual tracking is to give computer the ability of recognizing moving targets, and thereby to provide data basis for video analysis and understanding. Many interference factors challenge the visual tracking problem, such as, complex environment, variable illuminance, frequent occlusions, and object's changable postures. Though much effort has been made in this field, and many methods have been proposed, there are still many unresolved problems, and more stable key technologies and methods are still in urgent need to solve the above problems.This dissertation focuses on resolving the problems of complex tracking conditions and object's changeable observation, and mainly researches on Particle Filter based tracking algorithm in dealing with occlusion, improving tracking accuracy, and modeling changeable observation. The major research achievements include:(1) Multi-Regions'Joint Particle FilterIn order to deal with the occlusion problem, this paper proposed to represent target as multiple regions, and construct Multi-Regions'Joint Particle Filter (MR-JPF) to track multiple regions jointly. By this way, when one region is occluded, its status can be calculated by the other regions. In the proposed method, multiple region's joint Particle Filter is constructed based on a joint motion model specified by an undirected graph, and on a regions'relation based observe-and-estimate scheme. Specially, Markov Chain Monte Carlo method is employed to replace inefficient importance sampling. The experimental results tested on videos with various occlusion situations demonstrate that the proposed algorithm is more effective in solving long-time partial or total occlusion problem than the tracking method based on single region Particle Filter.(2) State prediction and classification combined visual trackingTo improve the accuracy of Particle Filter based visual tracking, this paper proposed to combine it with classification based target location method. In the proposed tracking framework, Particle Filter is employed to approximate target's states, according to which, a Support Vector Machine is used to locate target's state accurately. Specially, an initial SVM is constructed using a method called foreground-background joint histogram. And at each time step, SVM is real-time updated to adapt to the change of tracking condition. In the tracking procedure, at each frame of the video sequence, Particle Filter provides a matched region by approximately estimate target's state, and SVM formed in pre-frame outputs target's accurate state by classifying the pixels in that region. The proposed method shows improvement to bayesian state prediction or classification based methods, and makes promotion in tracking accuracy.(3) Adaptive multi-cue integration based observation feature representationTo adapt target's observation to the change of tracking environment, this paper proposed an adaptive weights updating scheme for multi-cue integration. And three weights updating algorithms are introduced. Firstly, by analyzing particles'distribution, this paper proposed a frame-by-frame weight adapting approach. Secondly, a principal and subordinate weight updating method is proposed, according to human vision characteristic. Thirdly, taking weight's time-series consistency into account, this paper proposed a weight tracking method. The proposed three weight updating algorithms provide a stable and reliable basis for accurate and robust visual tracking. Experimental results on videos with various tracking conditions show the significant improvements of the proposed methods, when comparing with the existing integration algorithms.(4) Short-track speed skating sliding data measurement systemThe proposed multi-region based joint Particle Filter is applied in a short-track speed skating sliding data measurement system, to track high-speed skater in the video captured by a moving camera. In the system, a probabilistic density gradient based corner detector and a global parallax constraint matching points filter are proposed, to realize effective and accurate registration of long video sequences. To track high-speed skater with frequent occlusions, the proposed multi-regions'joint Particle Filter is employed with careful design of object's observation model. And the system can output skater's accurate trajectory and motion data. The application of the proposed method in this system demonstrates it to be effective and practical in real scene, and can be extended in a wider range of applications.