Group Detection and Tracking. To detect groups, we perform graph-cutting in the social network graph by re- moving all edges below a fixed edge weight threshold. The remaining connected components of the graph then become the resulting groups. 14 People Detection, Tracking and Visualization on a Mobile Robot A simple group tracking algorithm can now be implemented by treating groups as separate entities, and tracking their centroids. This data association might fail if the group splits apart very suddenly, causing a large shift in the resulting new centroids. Alternatively, it is possible to track just the composition of a group, in terms of its individual group members, and then keep a memory of previously seen group configurations. This works as long as at least one person in the group is always visible. A simple Python-based implementation of this approach can be found in the xxxxxxx group tracking package. The output is in the form of TrackedGroups messages, as outlined in Fig. 8. In ongoing research shown in Fig. 6 (right), we use a more complex approach which interleaves a group-level data association layer with regular person-level data association in a multi-model multi-hypothesis tracker [17, 15]. By explicitly modelling group formation in terms of merge and split events, groups can be tracked more robustly. Internally, a prototype of this system has already been integrated with our visualization and detection framework and yielded good experimental results. While currently not publicly available, this system uses exactly the same message interfaces as the provided Python code. TrackedGroup TrackedGroups TrackedPerson
