Complex Context Of The Video Moving Target Tracking

Video moving object tracking is an important component of artificial intelligence and a key of solving many computer vision problems. A mass of research achievements for object tracking have been proposed in the past. However, when appearance of the object change on the condition of illumination variation, rotation, occlusion and scale variation, furthermore, the object is disturbed by other neighboring objects with smiliar appearance, how to track the object is not to be solved until today. In the thesis, we attempt to solve the issue by improving the process of object detection, object modeling and object tracking.The main contributions of thesis as follows:(1) Object detection method: In pixel domain, a novel background model method based on sparse presention theory is proposed. When fontground has been detected by background subtraction, an adaptive fusing color and texture information method is adapted to remove moving shadow with moving object. The refinement result as foundation is used in object tracking. Experiment results show the effective of the proposed method in complex scene. In H.264 compressed domain, due to motion vector noise, moving object segmentation only depending on motion vector may be often ineffective. In the thesis, a method combining intra prediction with motion vector is proposed to segment moving, on the other hand, a new spatial-temporal filter is proposed to remove motion vector noise. Experimental results for several H.264 compressed video sequences demonstrate the good segmentation quality of the proposed approach.(2) Object modeling method: In the complex background, the appearance changing of object due to illumination variation and rotation is major reason that those present object tracking methods were failed. In the thesis, scale invariant feature which is stabilize to illumination variation and rotation is used to construct object model and combined with particle filter framework. As a result, object can be tracked correctly with illumination variation, rotation and occlusion. In complex scene, tracking using only one feature is possible to lose. However, using too many features in tracking, computing complexity would increase very much and'dimension curse'which dissipations tracking performance could be induced. In the thesis, Fisher linear discrimination theory is used to select a few discriminative features which can distinguish forntground from background in feature set. Then, object is tracked by fusing these discriminative features.(3) Object tracking method: Particle filter has been used widely owing to universalness and robustness, however, particle filter has a high computation complexity, on the other hand, it is possible to occur particle impoverishment and fail in tracking. In the thesis, a hierarchic particle swarm optimization based target tracking is proposed. The proposed method maintains high performance in multi-model scene.