Detecting Social Groups in Crowded Surveillance Videos Using Visual Attention

In this paper we demonstrate that the current state of the art social grouping methodology can be enhanced with the use of visual attention estimation. In a surveillance environment it is possible to extract the gazing direction of pedestrians, a feature which can be used to improve social grouping estimation. We implement a state of the art motion based social grouping technique to get a baseline success at social grouping, and implement the same grouping with the addition of the visual attention feature. By a comparison of the success at finding social groups for two techniques we evaluate the effectiveness of including the visual attention feature. We test both methods on two datasets containing busy surveillance scenes. We find that the inclusion of visual interest improves the motion social grouping capability. For the Oxford data, we see a 5.6% improvement in true positives and 28.5% reduction in false positives. We see up to a 50% reduction in false positives in other datasets. The strength of the visual feature is demonstrated by the association of social connections that are otherwise missed by the motion only social grouping technique.