Stable Target Tracking Algorithm In Complex Scene

Target tracking of complex scene is one of the most popular research in Computer Vision, the main task of tracking is to detect and track the moving target in videos or cameras. Target tracking algorithms are used in many different fields, such as Intelligent Surveillance, Robot Navigation, Video Indexing and Virtual Reality etc.. Target tracking algorithm based on video has developed for many years. However, there are many challenging problems in target tracking, such as illumination change, target deformation, scale change etc.. These problems often cause the drifting problem as the error accumulates over time and finally lead to tracking failure. In order to solve these problems, we develop our research in terms of feature extraction, classifier construction,tracking decision and Spatio-Temporal context etc..In order to handle the illumination change, scale change and the target occlusion that lead to tracking failure, we propose a long-term tracking algorithm based on tracking failure detection and Weighted Random Forest. Firstly, in order to handle the illumination change problem, we extract the target multi-scale feature based onr bYC C illumination invariant color space. In order to compressive the dimension of target feature, we use the random projection to reduce the high-dimensional target feature.And then the low-dimensional feature subspace is used to initialize the Weighted Random Forest classifier that we proposed in this paper. Secondly, we introduce a tracking failure detection strategy calculated by the forward-backward trajectory error to decide whether the tracking is a failure. Finally, if the tracking is a failure, in order to relocate the target in the follow-up frames, we also proposed a Weighted Random Forest(WRF) classifier to retrieve the target position after the tracking failure situation,and the classifier is updated online.In order to solve problems caused by occlusion, we firstly propose an occluded target tracking algorithm based on spatio-temporal context information. Our algorithm first use compressive illumination invariant color space to construct and initialize the spatio-temporal context model. We decide whether the target it the current frame is occluded. For the target that is not occluded, the spatio-temporal context model is used to track the target accurately. But when the occlusion occurs(static occlusion and dynamic occlusion), our proposed cascade classifier is used to detect the position of thetarget after the occlusion disappear. The spatio-temporal context model and the classifier are both updated online which improve the performance of our algorithm over tracking.