Multiple Likelihoods and State Noises Based Particle Filter for Long-Lived Full Occlusion Handling

Reliable object tracking in complex visual environment is a challenging problem in the field of computer vision. One of the common problems in object tracking is partial and full object occlusions. And especially in the condition of long- lived full occlusion during which the full occlusion lasts for tens of frames, the tracking is more difficult. This paper proposes an occlusion handling scheme based on particle filter. Compared with the standard particle filter, multiple likelihood models - HSV color likelihood and gradient orientation likelihood, are employed in the observation model for occlusion handling. Also, multiple state noises are introduced under occlusion. Experiment results demonstrate the robust and accurate tracking performance in the condition of long-lived full occlusion.