Research On Visual Tracking Methods Based On Joint Decision From Multiple Regions

The visual tracking is one of key issues in computer vision feld, which has been widelyused in many vision-based applications, such as media production, medical diagnose, intel-ligent surveillance, unmanned vehicle, interactive gaming and military guidance. Althoughvisual tracking methods have been been widely researched and many efective algorithmshave been proposed, there are still a lot of difculties in designing a robust tracking algo-rithm due to the challenging complex scenarios such as signifcant illumination changes inenvironment, pose variations of the object, non-linear deformations of shape, and noise anddense clutters in complex background, etc. To overcome above difcult problems, we fo-cus the research on two major aspects: target presentation and target searching, and proposeseveral novel tracking methods based on joint decision from multiple regions. The maincontributions of this dissertation are summarized as follows:1. A dual-kernel-based tracking approach for visual target is proposed. When evaluat-ing a target candidate, we use two criteria: the similarity between it and target model, and thecontrast between it and its neighboring background. The similarity is measured using Bhat-tacharyya coefcient, the contrast is calculated using Jensen-Shannon divergence, and theyare integrated into a novel objective function. Performing multi-variables Taylor expansionon the function, we obtain its linear approximation. By maximizing the linear approxima-tion, we induce a dual-kernel-based target location-shift relation from current location to anew location. Based on the location-shift relation, the optimal target location can be recur-sively captured in the mean shift procedure. The target state is jointly determined by target candidate and its background region, so we name the new method as tracking method basedon joint decision from dual-regions.2. A detection approach of discriminative stable regions is developed by imitating thespatial selective attention. Three inter-related aspects are discussed, including mathematicaldescription of stability of a region, mathematical defnition of discrimination of a region,and fast localization and scale selection for discriminative stable regions. The early selec-tion process extracts a pool of salient regions from target image, and these regions have highstability. The late selection process identifes a subset of salient regions with strong localdiscriminative power as discriminative stable regions by using a fast search method. Exper-iments on several image sets demonstrate that these discriminative stable regions performwell in terms of repeatability and matching score.3. A new tracking approach based on discriminative stable regions is proposed. Thediscriminative stable regions are obtained based on the criteria of maximal local entropy andspatial discrimination, which enables the tracker to handle well distracters and appearancevariations. By incorporating K-means clustering, the collaborative tracking of multiple dis-criminative stable regions can tolerate more motion noise and occlusions. In addition, as anefcient technology, the subspace analysis is used to discover the potential afne relationbetween the discriminative stable region and the target, which timely adapts to the shapedeformation of the target. The target state is jointly determined by multiple discriminativestable regions, so we name the method as tracking method based on joint decision frommulti-regions.Extensive experiments demonstrate that these proposed algorithms can gain accuratetarget location and strong identifcation power to false target, and they are also robust todeformation and partial occlusion.