Visual Tracking Of Multiple Targets In Complex Scenes

One of key area of computer vision researching is visual tracking technique, which is the foundation of higher-level of motion recognition, event monitoring and video analysis etc. It is widely used in many fields including civilian and military application domain. A comprehensive study of the key techniques of multi-target visual tracking is conducted in this dissertation.Firstly, fixed appearance descriptor is applied in many existing visual tracking algorithms which assume the objects' appearance and surrounding illumination do not change drastically over time. Based on HSV color space histogram and region covariance target representation, an incremental appearance descriptor updating algorithm is proposed to avoid failure in dynamic scene. The proposed strategy improves the robustness of tracker. The performance of scene recognition is improved using region covariance descriptor as well.Secondly, a mixed probability data association method used for multi-target and multi-observation is introduced. It includes label gating, observation likelihood and position distribution. Label gating labels the position of target within a certain scope, observation likelihood distinguishes the target by its appearance features and position distribution predicts the current probability of targets' candidate positions according to it's position in the last sampling, and is also the fundamental part for particles dynamic assignment. The excellent ability to response the tracking is achieved by establishing the restrictive conditions of data association for targets.Thirdly, a dynamic assignment method of particles is developed to distribute the limited particle source more effectively. By using the position distribution, a map of position relationship of multi-target is constructed. In the re-sampling process, the particles are assigned to each target according to the proportions which are decided by the map.Lastly, to validate the proposed work, a multi-target visual tracking system using target appearance updating strategy and particle filtering is implemented; and a scene recognition algorithm based on region covariance is accomplished as well. The region covariance based object detection and visual tracking are also implemented. The experimental results demonstrate the feasibility and validity of the proposed techniques.