Efficient particle filter and its application in contour tracking

Abstract

In order to improve the real-time performance of particle filter, this paper proposes an efficient particle filter algorithm and evaluates its usage in object contour tracking application. This new filter uses only one particle to predict next state in certain situations. As particle set size is one, there is no need to resample the particles before prediction. Therefore the real-time performance of particle filter is improved. To maintain the performance of particle filter, we use one particle sample only in prediction step and only in situations where particles are close enough with each other. After prediction step, the original number of particle samples is recovered. Experiment shows that the new particle filter uses less time to process the particles while retain tracking performance comparable with the regular particle filter.