Multi-object Tracking Using Hidden Markov Model with SIFT Feature

Solving the problems of occlusion and illumination in the multi-object tracking task is the focus in current research. In the case of occlusion, improving the performance of the tracker can improve the tracking effect. This paper proposes an improved Hidden Markov Model for multi-object tracking. Firstly, we extract the SIFT feature points of the previous frame of image as the observation status and input it into the improved model. Secondly, we perform iterative training on each object to estimate the parameters of each model. Finally, the hidden status with the maximum probability of occurrence of the model is obtained according to the observed status and convergent parameters. The hidden status can reflect the SIFT points in the current frame to accomplish the tracking of object. In addition, we setup comparison experiment with other advanced trackers on MOT17 dataset. Experiment results show that our method achieves the best performance in MOTA and IDS, and have good robustness to occlusion problems.