A Robust Particle Filter for People Tracking

Among various tracking algorithms, particle filtering (PF) is a robust and accurate one for different applications. It also allows data fusion from different sources due to its inherent property without increasing the dimension of the state vector. In this paper, we propose three strategies to improve the performance of particle filters. First, our approach combines the foreground region with the particle initialization and similarity measure step to lower the background distraction. Second, we form the proposal distribution for particle filters from the dynamic model predicted from the previous time step. The combination of the two approach leads to fewer failure than traditional particle filters. Fusion of multiple cues including the spatial-color cues and edge cues is also used to improve the estimation performance. It is shown that with the improved proposal distribution above, the particle filter can provide greatly improved estimation accuracy and robustness for complicated tracking problems.