Multi-target Tracking in Crowded Scenes

Abstract

In this paper, we propose a two-phase tracking algorithm for multi-target tracking in crowded scenes. The first phase extracts an overcomplete set of tracklets as potential fragments of true object tracks by considering the local temporal context of dense detection-scores. The second phase employs a Bayesian formulation to find the most probable set of tracks in a range of frames. A major difference to previous algorithms is that tracklet confidences are not directly used during track generation in the second phase. This decreases the influence of those effects, which are difficult to model during detection (e.g. occlusions, bad illumination), in the track generation. Instead, the algorithm starts with a detection-confidence model derived from a trained detector. Then, tracking-by-detection (TBD) is applied on the confidence volume over several frames to generate tracklets which are considered as enhanced detections. As our experiments show, detection performance of the tracklet detections significantly outperforms the raw detections. The second phase of the algorithm employs a new multi-frame Bayesian formulation that estimates the number of tracks as well as their location with an MCMC process. Experimental results indicate that our approach outperforms the state-of-the-art in crowded scenes.