Visual Detection of Events of Interest from Urban Activity

Abstract

Learning patterns of human-related activities in outdoor urban spaces, and utilising them to detect activity outliers that represent events of interest, can have important applications in automatic news generation and security. This paper addresses the problem of detecting both expected and unexpected activities in the visual domain. We use a foreground extraction method to mark people and vehicles in videos from city surveillance cameras as foreground blobs. The extracted foreground blobs are then converted to an activity measure to indicate how crowded the scene is at any given video frame. The activity measure, collected over the period of a day, is used to build an activity feature vector describing that day. Day activity vectors are then clustered into different patterns of activities. Common patterns in the data are not considered important as they represent the everyday norm of urban life in that location. Outliers, on the other hand, are detected and reported as events of interest.