Research On Data-driven Based Visual Object Tracking

Visual object tracking is a fundamental task in computer vision field. However, suffered from lots of interference caused by natural scenes variation, model-driven which is most widly used in current tracking algorithms, failed to satisfy accuracy and robustness for actual applications due to its inherent defects.We analyze the framework of model-driven and find that data from different tracking stages could reflect variations intuitively. On this basis, we descript the feasible of data-driven and its implementation methods. Different from passive mode of data in model-driven, data-driven defines a series of QoI(Quality of Information) attributes based on real-time data stream, offers reliable information for model construction and updating and configures system well abilities to handle with variations.Based on mid-level description, we define the tracking algorithms combined with model-driven and data-driven.From the perspective of feature selection, we propose the adaptive selected sub-blocks tracking algorithom(ASST) based on current frame information, weakening degradation caused by historical information which is focused by traditional feature selection algorithms. Different sub-blocks are selected according to discrimination and uniqueness and participate in localization during tracking process. Experiments proved ASST with adaptive selection mechanism could well handle with occlusion and partial variations.From the perspective of template updating, an adaptive-weighted appearance model tracking(AAMT) algorithm is proposed. AAMT transforms the hard threshold of voting or not in ASST into soft one with variable weight, which ensures accuracy and avoids improper update of other sub-blocks. After that, we develop a novel template update strategy with changing rate and updating confidence. Lots of experiments validated the robustness and accuracy of AAMT with adaptive appearance variation mechanism.