Tracking periodic motion using Bayesian estimation

This work presents a Bayesian approach to achieve efficient and accurate motion tracking in monocular image sequences. We first extract a deterministic motion model with six degrees of freedom in an on-line learning phase. This is followed by predicting the image points in successive frames, and achieving correspondence in the context of Monte Carlo estimation. Meanwhile, the motion parameters of the camera are simultaneously estimated. The experimental results show that the stable and accurate ego-motion parameters can be obtained.