Samenvatting

To improve security in crowded environments, such as airports, shopping malls and railway stations, the number of surveillance cameras (CCTV) is rapidly increasing. However, the number of human operators remains limited and only a selection of the video streams can be observed. This makes it hard for an operator to be proactive. This paper gives an overview of novel developments that may lead to more efficient camera surveillance and a more proactive role for camera operators. It focuses on three main steps in this process of video content analysis: pedestrian tracking, action recognition and behavior analysis. Tracking and re-identification (i.e. recognizing a person in another camera) was initially only evaluated on off-line benchmark datasets, though recently it has gained in maturity with live demonstrations in realistic crowded environments and measured improved operator efficiency. For action recognition and automatic behavior recognition, we observe that the simple patterns, such as loitering detection, are emerging in many applications. Human action recognition obtains very high performance values in controlled environments and it is progressing towards more realistic environments. More advanced approaches, such as pickpocket recognition in a shopping mall and the detection of threats to trucks on a parking lot have been developed and the first systems have been presented in live demonstrations. Our main contribution is that we structure the recent advances and the emerging applications of video analysis for security applications, explain and interpret the results, and identify opportunities for the near future.