An Efficient Vision-Based Group Detection Framework in Crowded Scene

Abstract

Visual surveillance systems are now widely used for monitoring the events. The challenging task in these events is effectively analyzing the crowd and its behavior. For better understanding the behavior of crowd or analyzing it, the group is considered as the basic element. The exigent task in a crowded scene is to distinguish between the groups and individuals. In this paper, we have proposed a video-based framework that efficiently identifies the group of people from the crowd. The framework is composed on boundary extraction of a group called contours in the literature. The proposed approach makes use of background subtraction algorithm called ViBe, to extract the relevant features and incur contours in the video frames. Further we detect the group in a crowd on the basis of threshold frames obtained by calculating the area and distance between them. Analysis has been carried out on a self-gathered dataset from the university campus. The proposed framework is able to distinguish the group and no group with an average accuracy of 86.06%.