Research On Video Target Tracking Based On Particle Filter

Video target tracking is an important research topic in the field of computer vision with numerous applications such as intelligent traffic, statistics of flow people in public places, video surveillance. The task of video target tracking is to track interested target in video sequence, and obtain the parameters of moving target, for further processing and analysis.The particle filter can be applied to any non-linearity and non-normality state space model. For video target tracking, researchers have turned to particle filter-based tracking methods due to the argument that the target movements are not linear. However, now most of the particle filter tracking algorithm only uses colour feature to calculate the weights of particles, thus in complex situations such as low contrast, the target is occluded, etc., tend to approach the wrong location. In order to improve the stability of target tracking algorithm under the condition of target tracked undergoing interfere with targets near the tracked target, and low contrast, we propose a multi-feature cascade-sample particle filter target tracking method.In the condition of target undergoes severely occlusion, ensure the stability of target tracking system is difficulty. Aim to this problem this paper propose a new occlusion judgement and procession strategy: calculate the Bhattacharyya distance between candidate region and target template region, if the Bhattacharyya distance is larger than a threshold, the target is occluded severely, split the target template into several sub-regions, for each sub-region use mean-shift algorithm to iterate, select the sub-regions with lower Bhattacharyya distance as"reliable sub-region", according to which infer the color information of occluded region, and revise the target position according to topology between the sub-regions. Select the discriminative features for the next time step.Based only on 2-D information from single camera tracking-by-detection approach, which does not require any camera or ground plane calibration, can track targets in the condition of camera moving, becomes increasingly popular. The main difficulties in tracking–by–detection approaches based on 2-D information are that the data association computation cost is large and the contradiction between high detection rate and low false alarm rate. Aim to these problems, feedback the targets tracking result to the target detection module, and provide position, size information of tracking targets for target detection module; feedback the response strength of target detect ion to particle filter target module, and guide sampling particles, thus improve the stability of multi-targets tracking. This multiple targets(human body) tracking method consists of three modules, that is: target detection module, targets tracking module, cascade matching data association module, these three modules coordinate each other to achieve the multiple targets tracing task.