Utilization of Occluded Detections and Target Information in Multi-Person Tracking

Multi-person tracking has many applications such as surveillance or automated driving. Existing approaches exploit available motion and appearance cues insufficiently. In contrast, this work introduces several methods to improve the usage of detections and target information, including novel association strategies, distance measures, and an occlusion-aware initialization. The proposed framework achieves state-of-the-art results on multiple benchmarks and tracks hundreds of persons in real time.