People Tracking Algorithms Used in Our Experiments. During the XXXXXXX project, we have conducted experiments with different person tracking algorithms of varying complexity. Next to a computationally cheap nearest-neighbor algorithm which we will describe in the following section, we converted an existing hypothesis-oriented multi-hypothesis tracker from our past work [1, 17] into a ROS package4 to exploit the visualization capabilities and interchangeability of detectors of our framework. Further experiments were conducted using a vision-based track-oriented multi-hypothesis tracker5 [12]. A detailed study of which person tracking algorithm to choose under which circumstances is subject of our ongoing research. We believe that in general, the choice of models (e. g. for motion prediction, occlusion handling, inclusion of visual appearance) has a greater influence on overall tracking performance than the tracking algorithm itself. Here, various methods have been proposed in the past, e. g. the use of a motion model influenced by social forces [19], adapt- ing occlusion probabilities and motion prediction within groups [17], improved occlusion models [16], or an online-boosted RGB-D classifier that learns to tell apart tracks from background or other tracked persons [18, 22].
