Social Relation. Estimation Before detecting groups, we first estimate the pairwise social relations of all tracked persons by feeding the output of the people tracker into a social rela- tion estimation module. This module, based upon the work described in [15], builds a social network graph as shown in Fig. 6 (left), where the edge weights encode the likelihood of a positive social relation between a pair of persons. To estimate these likelihoods, we rely on coherent motion indicators, which are People Detection, Tracking and Visualization on a Mobile Robot 13 motion-related features that were found to indicate group affiliation between people in large-scale crowd behavior analysis experiments [20]. They consist of relative spatial distance, difference in velocity and difference in orientation of a given pair of tracks. Using a probabilistic SVM classifier trained on a large dataset annotated with group affiliations, these features are mapped to a social relation probability that indicates the strength of the relation. This method al- lows for an easy integration of additional social cues (such as persons’ eye gaze, body pose) in the future. An implementation and a trained SVM model are provided in the xxxxxxx social relations package. Figure 7 shows the corre- sponding SocialRelation and SocialRelations ROS messages that constitute the resulting social network graph.
