Identifying social groups in pedestrian crowd videos
Arun Kumar Chandran, Loh Ai Poh, Prahlad Vadakkepat · 2015
A Non-recursive Motion Similarity Clustering (NMSC) algorithm is proposed to identify pedestrians traveling together in social groups. The clustering algorithm is unsupervised and can automatically identify social groups within a region of interest in a video. Social groups are identified using only pedestrian motion information by imposing motion parameter thresholds defined by social psychological principles. Social groups are identified without any prior training. In addition to detecting small social groups, NMSC also detects short-term groups (occurring for a few seconds) and social groups with sparsely distributed pedestrians. The real-time performance and group identification accuracy reveal that the proposed clustering algorithm performs better compared to existing algorithms even for scenes with a large number of pedestrians.