Video-based affinity group detection using trajectories of multiple subjects

Affinity detection has been largely motivated by the increasing interest in modelling the social behavior of humans. This paper presents a supervised learning method for affinity detection which is based on an inference obtained from tracking trajectories of the human subjects captured in video sequences. In particular, the proxemic cues of group detection such as the pair-wise similarity of the positional and translational measurements of the tracked people are used in the well-known principal component analysis-based feature extraction process. The existence or non-existence of pair-wise affinities is recognized using the nearest neighbor detector applied on the proposed features and the majority voting-based fusion of decisions. Experiments conducted on surveillance video captured in diverse-type of movements of the subjects show favorable results in terms of accuracy of detecting affinities when compared with the ground truth.