Part-based Data Association for Visual Tracking

We present a method that integrates a part-based sparse appearance model in a Bayesian inference framework for tracking targets in video sequences. We formulate the sparse appearance model as a set of smoothed colour histograms corresponding to the object windows detected by the Deformable Part Model (DPM) detector. The data association of each body part between frames is solved based on the position constraint, appearance coherence, and motion consistency. To deal with missing and noisy observations, the part detection window in the following frame is also predicted using an interacting multiple model (IMM) tracker. We have tested our tracking method on all the video sequences that involve people in upright poses from the TB-50 and TB-100 benchmark videos datasets. Our experimental results show that our tracking method outperforms six state-of-the-art tracking techniques.