Targets Tracking Based On Joint Algorithm Of Motion Compensation And RJ-MCMC In Video Sequence

Moving target tracking from videos is an important research topic of computer vision, which receives a wide attention from domestic and international scholars. It has broad application prospects in missile visual guidance, robot and autonomous navigation, military targets localization, tracking and identification, safety monitoring, intelligent driving, traffic, virtual reality, etc. In reality, the existences of various noises increase the difficulty of object detection and tracking from videos.Target detection is the foundation of the target tracking, and the detection effect will bring significant influence for tracking. Various noises brought by light changing, target shade and camera sensors make target detection very complex, especially when the camera moves, background and foreground subsequently move accordingly in the images, so it brings more difficulties to distinguish the foreground from the background. This thesis adopts motion compensation and adaptive background updates to detect the target motion in dynamic scenes. Firstly calculate the pyramidal optical flow field of the two consecutive frames, the pyramidal optical flow field data contains not only target's and background's optical flow value but also that of noise. Because we cannot predict the number of targets and the interference caused by noise, we cannot use fixed number clustering algorithm. This thesis adopts an online clustering algorithm named leader-follow clustering algorithm, which can adaptively classify the pyramidal optical flow field data according to the input data. We get the compensation value when the optical flow value of the background is separated, then we use this value to compensate the previous frame. The background pixels of the previous frame are the same as the correspondence pixels in current frame after compensation. Then the background update algorithms applied in static scenes can be used to generate background.In video target tracking field, usually target tracking can be treated as a Bayesian estimation, that is, estimate the states from the observation data. Because the erratic target motion is difficult to describe by mathematic formulae, and the particle filter do not need any prior assumptions to states transfer, so many scholars adopt it in target tracking field. The particle degradation problem in ordinary particle filter seriously limits its development. As a particle filter, MCMC can effectively solve this problem. But in multi-targets tracking, the number of targets in the scene is not fixed, moreover the MCMC can not deal with the dimensions of solution space that is alterable, so the MCMC can not deal with multi-target tracking. The RJ-MCMC is designed to solve this problem, so we adopt this method to deal with multi-target tracking. Aiming at the multi-target tracking in the dynamic scene, we proposed a new RJ-MCMC particle filtering method based on a twice-observation model, The first observation is to modify the motion model though motion compensation to approach the true motion equation of the targets, while the second observation is the RJ-MCMC particle filtering procedure. Time variant motion model can increase the efficiency of the RJ-MCMC algorithm, reduce the number of ineffective particles, and enable it convergence to the real value faster.The features of targets may be obscure or its foreground regions are very small, this will increase the difficulty of tracking. Aiming to this problem we proposed a joint color histogram with foreground appearance comparability model. This observation model makes use of foreground information to enhance the features of targets. Experiments show that this improved observation model can effectively handle little targets' tracking.