A novel particle filter with implicit dynamic model for irregular motion tracking

Abstract

The particle filter is an effective approach for virtual object tracking. However, it suffers from the inaccuracy of tracking performance and track drifts which are caused by the inaccurate dynamic model. In irregular motion tracking, because of the large motion uncertainty and the poor prediction of the dynamic model, these two problems will definitely occur. We propose to model the object motion by an implicit dynamic model which is optimized by an iterative optimization method. We observe that the state with the biggest value of the sum of all particles' likelihoods will reach or be close to the ground truth which is testified by many experiments. Based on this, the dynamic model is formulated by maximizing an objective function. By evolving particles to new positions to obtain the maxima of summed particle likelihood, this particle shift strategy considerably improves the sampling efficiency. Extensive experiments show that the proposed algorithm outperforms other six trackers in dealing with irregular motions.