Multi-cue Based Discriminative Visual Object Contour Tracking

This paper proposes a discriminative visual object contour tracking algorithm using multi-cue fusion particle filter. A novel contour evolution energy is designed by integrating an incremental learning discriminative model into the parametric snake model, and such energy function is combined with a mixed cascade particle filter tracking algorithm fusing multiple observation models for accurate object contour tracking. In the proposed multi-cue fusion particle filter method, the incremental learning discriminative model is used to create observation model on appearance of the object, while the bending energy, calculated by the thin plate spline (TPS) model with multiple order graph matching between contours in two consecutive frames, together with the energy achieved from the contour evolution process, are both taken as observation models on contour deformation. Dealing with these multiple observation models, a mixed cascade important sampling process is adopted to fuse these observations efficiently. Besides, the dynamic model used in the tracking method is also improved by using the optical flow. Experiments on real videos show that our approach highly improves the performance of the object contour tracking.