A part-based online modeling algorithm and its applications on robust visual tracking

This paper proposes a part-based online object modeling and tracking method. The desired object is first tracked by IVT tracker which returns the appearances of objects in several leading frames. On these appearance subimages of object the SIFT key points are collected to build the training set for part detecting and modeling. NCut clustering then partitions key points into groups and the Gaussian models are employed to build the structure of object. Groups with the minimum variances of Gaussian models are considered as the parts. Given the learned parts, their corresponding appearance-based detectors are set by the solutions to the out-of-sample problem of NCut. Once a new key point is found in testing frames, the key point is classified by the part detectors and only the member of the learned parts are the survivals. The detected new part then spreads its scores around the potential centers indicated by its corresponding Gaussian model. The assembly of parts thus is completed by sum all scores up for all new instances in one frame. The maximum of sum scores points out the object locations if the score is above a fixed threshold. Otherwise, the re-training stage is called to update the detectors and the models. The experimental results on several challenging video sequences demonstrates the successful use of the proposed method.