Group activity recognition using belief propagation for wearable devices
Dawud Gordon, Markus Scholz, Michael Beigl · 2014
Humans are social beings and spend most of their time in groups. Group behavior is emergent, generated by members' personal characteristics and their interactions. It is therefore difficult to recognize in peer-to-peer (P2P) systems where the emergent behavior itself cannot be directly observed. We introduce 2 novel algorithms for distributed probabilistic inference (DPI) of group activities using loopy belief propagation (LBP). We evaluate their performance using an experiment in which 10 individuals play 6 team sports and show that these activities are emergent in nature through natural processes. Centralized recognition performs very well, upwards of an F-score of 0.95 for large window sizes. The distributed methods iteratively converge to solutions which are comparable to centralized methods. DPI-LBP also reduces energy consumption by a factor of 7 to 40, where a centralized unit or infrastructure is not required.