There is a broad range of applications of visual object tracking in fields such as military guidance, intelligent video surveillance, vision-aided navigation robotics and human-computer interaction,which has become an important topic for computer vision. Although extensive efforts have been taken, due to intrinsic factors, e.g. pose change, shape deformation and extrinsic factors, e.g. occlusion, illumination change make distinction lower, which seriously affect the capability of target tracking. Consequently, it is urgently need to study the novel effective and efficient theories and methods. Cognitive psychology shows that visual attention saliency (shorthand for saliency) has the ability to filter region of interest and suppress background. The study on saliency for object tracking and algorithm has important academic significance and actual application value. The traditional method of saliency-based object tracking is to detect target region as a preprocessing mechanism, however, the ambiguous boundary between target and background will make tracking drift. From the aspect of feature, saliency enforces the requirement that high distinction degree, strength discriminant and adaptability. For the above key issue, from the view of feature model and location searching scheme, we have proposed four efficiency and effectiveness trackers respectively. The one is Mean Shift tracking algorithm based on local region feature contrast saliency. The next one is Mean Shift tracking algorithm based on Top-Down evaluation local spatio-temporal saliency. The other one is Particle Filter tracking algorithm based on global context combined saliency. The final one is Particle Filter tracking algorithm based on frequency and motion saliency with evaluation mechanism. The effectiveness of the proposed algorithms is verified by a variety of experiments and results show that they has better distinction degree and handle the drifting significantly. The main works and contributions of the research in this thesis are detailed below. (1) The pros and cons for the tracking algorithms and saliency are detailed. For the key problem of distinctions degree, discriminant and adaptability for object tracking, we provide insights for future about the saliency-based tracking algorithms. (2) To reduce the disturbance of intrinsic factors, based on feature integrated theory and contrast, we propose Mean Shift tracking algorithm based on local nearest region feature contrast saliency. W...
修改评论