Robust object tracking using Bi-model

Occlusion is one of the major problems that object tracking faces in a clustered environment. In this paper, a tracking method which can deal with partial occlusion is proposed. There are two novelties in this paper: (1) using SURF keypoints to represent the object, key-points are evaluated and online learned by Random Ferns. (2) Bi-model is proposed to store key-points from object and surrounding background. In each frame, key-points inside or around the object bounding box will be assigned labels by matching with points stored in the Bi-model. These labeled points will be further used for improving the tracking accuracy and learning of Random Ferns. Long-term tracking is achieved by combining detection and tracking together. Experiments on videos with occlusion conditions show that the proposed method has good performance on tracking partial occluded objects, compared to some of the state-of-art methods.