Social group discovery in protest activities via two-level social network

Abstract

Social group is defined as two or more humans interacting with each other and sharing a common identity. This paper proposes a method to detect and discover social groups in protest activities in Hong Kong, through social network (SN) on visual video data. For example, three social groups, namely as police, protestors and onlookers, are usually identified in a protest activity. Every individual person in the protest scene is detected by a well-established three-dimensional (3D) model-based human tracking method. There are three contributions in this paper. First, to the best of our knowledge, this is the first attempt in the world of social group discovery (SGD) in a real-world protest activity. Second, a new two-level SN structure based on an appearance model utilizing the torso colors in people of different groups is constructed. Third, the capability of the two-level SN structure is verified. A preliminary evaluation is carried out in five real-world video frames in the database of protest activities with an average 94.34% accuracy for a discovery of three social groups.