A Keypoint-based Fast Object Tracking Algorithm

This paper proposes a novel keypoint-based algorithm to track multiple objects in real time from a fixed camera. Instead of searching keypoints from the whole video image, we only detect keypoints from the region of interest, which can be obtained by adaptive background mixture models. Our approach is robust against object appearance changes because of its learning module. Furthermore, our algorithm makes a good use of fast keypoint detectors and binary descriptors to run in real-time, which qualifies it for online applications. It requires no camera or ground plane calibration and can efficiently work at a frame rate of 28-32 fps under the image resolution of 768 x 576.