Robust Pedestrian Tracking Using Interactive Multiple Model Particle Filter and Feature Matching

Abstract

Pedestrian tracking with a single motion model and a single cue cannot be robust against combination of changes of target motion and appearance, caused by either static or dynamic changes in surrounding environments. Different from previous work, this paper proposes to refer to both multiple motion models and cues to improve accuracy of single pedestrian tracking. By integrating the compensated motion model into model filtering of interacting multiple model particle filter (IMMPF) and updating mode probabilities using the state with the maximum color based likelihood weight, the interacting stage of IMMPF generates more accurate approximation of the mixed a priori probability distribution of the target state. Within the search window centering at the predicted state of the constant velocity (CV) model of IMMPF, feature matching of scale- and rotation-invariant feature descriptors helps locate the target. Experimental results show that the proposed scheme is robust against the change of target motion, occlusion, rotation, scaling, and illumination variations. The average RMSE and overlap ratio of the proposed scheme significantly outperforms color based IMMPF using compensated motion model and fast L1-tracker.