Human Group Clustering in a Crowded Public Place Using Multiple Object Detection and Tracking

Donggoo Kang, Yeongheon Mok, Yeong-Joon Kim, Sunkyu Kwon, Joonki Paik

Abstract: Most people have their own social group that connects with each other. Therefore, the group is the basic element that composes the crowd. It is key to analyze the social behavior of the crowd. However, since the complexity of interaction, capturing the behavior of a group is hard to define. In this paper, we present a novel algorithm that detects pedestrian groups in view of the trajectory of their tracklet. The algorithm is composed of two main parts, detection-tracking and group clustering. First, we use a real-time detector to densely detect pedestrians and a multi-object tracker to keep their individual ID. Second, we compute the relative distance of each ID and assign group ID based on their distance. The proposed algorithm keeps the personal ID and also the group ID. Experimental results show that the proposed algorithm capture group successfully on a complex real-world scene.