Visual Tracking Based On Non-negative Coding Boosting

The object tracking is one of the essential problems in the field of computer vision. It involves a variety of applications including human-computer interaction, medical imaging, security, surveillance, etc. Due to the development of computer science, object tracking becomes more complicated. Meanwhile, owing to pose variation, illumination change, occlusion, and motion blur, developing an effective and efficient model for robust object tracking is a challenging task. Now, there is no an algorithm that can handle all problems in process of tracking. Robustness, timeliness and accuracy are the main difficulties of target tracking, so the tracking algorithm should be further researched.In this paper, a series of research work is done on the abnormal condition of the target in the video sequences. The main research results are as follows:1.Proposed a target tracking approach based on non-negative encoding boosting. In contrast to the existing boosting tracking approach, the proposed method focuses on the global optimal subset of classifiers for boosting, in which the classifier selection is achieved in term of the nonnegative coding. The generated codes reflect the importance of corresponding weak classifiers and are used as weights. In object tracking, by the predicted labels based on the training samples using the selected weak classifiers, we are able to further analyze the correct classification rate to identify the occurrence of outlier, which is beneficial to correctly updating classifier parameters and avoiding tracking drifting and achieving accurate and reliable tracking.2. Proposed a target tracking approach based on spatial constraint non-negative encoding boosting. The method is based on the original method, considering the similarity between adjacent frames in a complex scene, the spatial consistency of the video sequences is maintained by the addition of spatial constraints, which makes the tracking algorithm more stable. Meanwhile, a new target representation is adopted, and all the samples are integrated into a unified size to make the original tracking algorithm can deal with the problem of the target scale.