摘要
Part-based tracking is a promising approach to visual tracking, which has been demonstrated more flexible than whole-object-based methods and more robust than local-feature-based methods. However, most of traditional part-based tracking methods are model-based, in which the priori knowledge, such as a set of part detectors and a structure model, becomes essential and critical. This constraint limits the extension of part-based trackers in many real-world tracking tasks. This paper presents a generic part-based approach to visual tracking. Different from existing part-based tracking which typically involves a structure model to handle object deformation, the proposed approach employs a matrix model which is non-deformable and generic. Each element of the proposed matrix corresponds to a part of the object, while the spatial relation between parts is encoded. Each part is associated with two attributes: a detector and a weight. To capture the variations of objects'; appearance, these attributes are updated online. The proposed approach can work generically in real-time and handle kinds of appearance variations. We compared the proposed approach to state-of-the-art tracking methods and demonstrated its robustness and effectiveness.
- 出版日期2012
- 单位中国科学技术大学; Microsoft