Robust object tracking based on multiple cues.

摘要

Many real-world applications, e.g., video surveillance and video conferencing, require robust visual object tracking. However, due to the richness and large variations in the visual inputs, robust and efficient visual tracking in complex environments is still an open problem even after years of research. This dissertation is dedicated to achieving efficient and robust tracking of nonrigid objects based on multiple visual cues, constraints, and global prior knowledge of the objects. Data-driven algorithms and sampling-based methods are designed and integrated to fuse various sensors and prior knowledge for real-time probabilistic object tracking.; Two efficient data-driven tracking algorithms are proposed in Chapters 2 and 3, which focus on region-based and contour-based object tracking, respectively. The tracking is treated not only as a matching problem, but also as a grouping problem. The presented methods can incorporate both the foreground and background models to detect the new object position based on multiple visual cues (e.g., motion, color, and edges), spatial-temporal constraints, and prior knowledge of object dynamics and shape.; Chapter 4 addresses the issue of robustness by fusing multiple sensors and maintaining multiple hypotheses. An efficient hybrid Bayesian sensor fusion framework is proposed to combine the proposed individual data-driven trackers and high-level knowledge of the object likelihood model. Unlike the traditional sensor fusion algorithms, a closed-loop is formed to adapt, evaluate, and calibrate different data-driven trackers based on generative and discriminant models of the objects.; A real-time multiple-speaker tracking system based on sound source localization, color-blob, and object contour validate the efficacy of the proposed fusion framework.