Multimodal IR 3-D target tracking based on non-uniform sampling

Aiming at the tracker failure problem induced by target pose changes during IR 3-D target tracking, a multimodal method based on non-uniform sampling was presented. Firstly, several prototype views were used to represent 3-D target, the corresponding prototype shapes were regarded as the multimodal shape representation, and the transition probability matrix between these prototype shapes was established. Then, under the frame of particle filtering, the reference target models were defined as the log-polar transformed prototype views' gray distribution features. By sampling the shape transition probability, the transfer and propagation of the sample's shape state were achieved. The method improves the robustness of the tracker to target pose's change. Meanwhile, log-polar transform possessing the non-uniform sampling characteristics can suppress target's scale and rotate deformation, and decrease the computation burdens. Simulation results show that the proposed method can track 3-D targets effectively.