A Detailed Study of the Association Task in Tracking-by-Detection-based Multi-Person Tracking

Many multi-person trackers follow the tracking-by-detection paradigm applying a person detector in each frame and linking detections of the same target to form tracks in the association task. While the basic concept is the same among these methods, various motion models, distance metrics to measure the similarity of targets, and matching strategies are used. This makes it difficult to compare different methods and also to assess the influence of single tracking components on the final performance. For these reasons, all parts of the association task are thoroughly investigated in this study. Starting with a simple baseline which is consequently improved with the help of experimental results, a strong tracking-by-detection-based framework is developed that achieves state-of-the-art performance on two multi-person tracking benchmarks.