Nobody likes Mondays: foreground detection and behavioral patterns analysis in complex urban scenes

Abstract

Streams of images from large numbers of surveillance webcams are available via the web. The continuous monitoring of activities at different locations provides a great opportunity for research on the use of vision systems for detecting actors, objects, and events, and for understanding patterns of activity and anomaly in real-world settings. In this work we show how images available on the web from surveillance webcams can be used as sensors in urban scenarios for monitoring and interpreting states of interest such as traffic intensity. We highlight the power of the cyclical aspect of the lives of people and of cities. We extract from long-term streams of images typical patterns of behavior and anomalous events and situations, based on considerations of day of the week and time of day. The analysis of typia and atypia required a robust method for background subtraction. For this purpose, we present a method based on sparse coding which outperforms state-of-the-art works on complex and crowded scenes.