Approximating Inference on Complex Motion Models Using Multi-model Particle Filter

Abstract

Due to its great ability of conquering clutters, which is especially useful for high-dimensional tracking problems, particle filter becomes popular in the visual tracking community. One remained difficulty of applying the particle filter to high-dimensional tracking problems is how to propagate particles efficiently considering complex motions of the target. In this paper, we propose the idea of approximating the complex motion model using a set of simple motion models to deal with the tracking problems cumbered by complex motions. Then, we provide a practical way to do inference on the set of simple motion models instead of original complex motion model in the particle filter. This new variation of particle filter is termed as Multi-Model Particle Filter (MMPF). We apply our proposed MMPF to the problem of head motion tracking. Note that the defined head motions include both rigid motions and non-rigid motions. Experiments show that, when compared with the standard particle filter, the MMPF works well for this high-dimensional tracking problem with reasonable computational cost. In addition, the MMPF may provide a possible solution to other high-dimensional sequential state estimation problems such as human body pose estimation and sign language estimation and recognition from video.