Vision Based Multi-pedestrian Tracking Using Adaptive Detection and Clustering

This paper proposes a novel vision based multi-pedestrian tracking scheme in crowded scenes, which are very common in real-world applications. The major challenge of the multi-pedestrian tracking problem comes from complicated occlusions, cluttered or even changing background. We address these issues by creatively combining state-of-the-art pedestrian detectors and clustering algorithms. The core idea of our method lies in the integration of local information provided by pedestrian detector and global evidence produced by cluster analysis. A prediction algorithm is proposed to return the possible locations of missed target in offline detection, which will be re-detected by online detectors. The pedestrian detector in use is an online adaptive detector mainly based on texture features, which can be replaced by more advanced ones if necessary. The effectiveness of the proposed tracking scheme is validated on a real-world scenario and shows satisfactory performance.