A state vector augmentation technique for incorporating indirect velocity information into the likelihood function of the sir video target tracking filter

We consider the problem of tracking moving targets in heavily cluttered video sequences and introduce a new state vector augmentation technique to incorporate indirect velocity information into the likelihood function of the SIR particle filter. While the importance of motion information in video tracking has been well recognized, the standard SIR filter typically weights particles using a likelihood function that considers the appearance model only. Our goal is to prevent particles with poor velocity estimates from receiving large weights. The key modifications involve saving the previous values of the state variables in the state update equation and reformulating the measurement model to deliver both the current and previous observations. This leads to a straightforward calculation of likelihood across pairs of temporally adjacent frames. Our preliminary experimental results show that the proposed method is effective for avoiding track losses due to the filter locking onto structured clutter.