Analyzing video produced by a stationary surveillance camera Paolo Buono Dipartimento di Informatica Universit` a degli Studi di Bari “Aldo Moro” via Orabona, 4, Bari, Italy buono@di.uniba.it Abstract — Today surveillance systems are every- where. Human observers watching live videos of specific areas are not efficient due to the likely loss of attention. On the other side, unattended surveillance systems re- quire that people analyze hours of recordings when they have to search for some specific events, e.g. identify peo- ple responsible of violence, theft or other offences. In many cases a specific search in the video has to be ac- complished in the shortest amount of time. This paper presents MotionFinder, a tool that performs video anal- ysis by computing an interactive summarization of the movements in a scene. Once the summarization process is complete, the tool responds in real time to inquires. For example, human investigators may search for spe- cific areas in the video that show high levels of activity or where they know that something occurred (e.g.: prop- erty damaged or stolen). The tool responds by showing only the scenes in which some activity occurred for that specific area of the video. Video summarization, video analysis, visual analytics, stationary surveillance cameras 1 Introduction One morning, John enters into his work building and he notices that the director, the responsible of security, the guardian and a technical employee are looking at the guardian’s monitor in the guardian’s room. After a couple of hours, John passes by the guardian’s room again and the four people are still there, so John asks what is going on. It took about four hours of the four people to get to the exact moment in the video in which a thief stole a laptop from an office. The four people moved back and forth in the video and also used the fast-forward feature without noticing any change in the office scene, since only a few frames within a video recording of about 24 hours referred to the very quick action of the thief getting the laptop from the office. The previous scenario describe a typical situation in the video surveillance context. Today video surveillance sta- tionary cameras have built-in software able to recognize movements in the scene, i.e. to capture scenes in which something moves. Motion detection doesn’t always helps, in particular for heavy traffic areas, where almost always something moves. Having a human guardian watching surveilled areas could be not optimal, particularly when there are many cam- eras. Human attention span can drop below acceptable lev- els after only 20 minutes, even in trained observers [4]. If we add that humans can handle a number of items of seven, plus or minus two, the use of observers is very ineffective when more than a tenth of video surveillance cameras are installed. Nevertheless, a surveilled area is a deterrent for people and reduces the number of causalities or damages. Entire cities have networks of surveillance cameras in or- der to cover specific locations for detecting and identifying potential threats or suspicious events. These systems often adopt real-time algorithms for detecting anomalies, identify objects and track them. This paper describes a novel technique we have devel- oped for video analysis and a sofware tool, called Motion- Finder, which implements this technique. A first proposal of the technique was presented in [2]. MotionFinder allows a human investigator to speed up its search for anomalies by quickly selecting excerpts of the video in which an event occurred. The tool is intended for post-processing activities, not for providing real-time alerts. Nevertheless, it is possi- ble to work in real-time, since the adopted summarization technique is very fast. Next section provides an overview of related work. In Section 3, the summarization technique is presented. Mo- tionFinder is illustrated in Section 4. Section 5 describes an interaction session in order to show how the tool works. Finally, conclusions and possible future research directions are reported.