VisLoiter: a system to visualize loiterers discovered from surveillance videos

Abstract

This paper presents a system for visualizing the results of loitering discovery in surveillance videos. Since loitering is a suspicious behaviour that often leads to abnormal situations, such as pickpocketing, its analysis attracts attention from researchers [Bird et al. 2005; Ke et al. 2013; A. et al. 2015]. Most of them mainly focus on how to detect or identify loitering individuals by human tracking techniques. A robust approach in [Nam 2015] is one of the state-of-theart methods for detecting loitering persons in crowded scenes using pedestrian tracking based on spatio-temporal changes. However, such tracking-based methods are quite time-consuming. Therefore, it is hard to apply loitering detection across multiple cameras for a long time, or take into account the visualization of loiterers at a glance. To solve this problem, we propose a system, named VisLoiter (Figure 1), which enables efficient loitering discovery based on face features extracted from longtime videos across multiple cameras, instead of the tracking-based manner. By taking the advantage of efficiency, the VisLoiter realizes the visualization of loiterers at a glance. The visualization consists of three display components for (1) the appearance patterns of loitering individuals, (2) the frequency ranking of faces of loiterers, and (3) the lightweight playback of video clips where the discovered loiterer frequently appeared (see Figure 1 (b) and (c)).