Early Recognition of Suspicious Activity for Crime Prevention

1. Introduction

Video surveillance is attracting much of the researchers’ attention since it is a crucial tool for protecting people and public property and finds most promising application in computer vision. The recent acts of terrorism have necessitated the vital need for well-organized surveillance of suspicious human behaviors at important public places. Suspicious behavior in general incorporates an event that builds a disbelief or mistrust. These events have to be reported for further critical examinations. Currently the surveillance cameras are used extensively and the purpose of these cameras are primarily for security monitoring and identifying illegal actions and events. It is unfeasible for a human supervisor to inspect a huge volume of video recordings, since security personnel are also required to manage other tasks, such as access control, handling emergency calls, following up on fire alarms, radio communications control, etc. However, this is still a novel technology with many practical restrictions. How can one differentiate among an individual running because he/she is delayed or because he/she has just committed a crime? Under such scenario, activity detection alone is not sufficed; prevention of such crime becomes necessary. Current researches on video analysis for surveillance system are post investigation methods and necessitate real-time processing. For example, recognizing missing objects after they have been stolen is of no use. Hence, it is enviable to predict the intension of a person, before the complete happening of the suspicious event. This prediction will possibly prevent a criminal activity, thus providing enough time to the security person to react upon the critical condition to ensure public safety. Hence an automatic activity recognition methodology is necessary.

A greater part of the present surveillance systems only record the events and act as a post event investigation tool which is not preferred. Other systems still submissively observe the surroundings and raise the alarm once a suspicious action is detected. Such an alarm will not help to prevent an unwanted event. Predicting a person’s action before it is executed has an ample choice of applications in autonomous robots, surveillance and health care. While comparing the action recognition using full length video, action prediction with unfinished video observations requires identifying the well-grained factors intrinsic to the existing observations that would direct to the future action. For example, a person with stretched open arms indicates that he/she is going to hug. Although, traditional models like hidden markov models are used to approximate the prediction problem, they are found to be inappropriate for the sparse discontinuous features of the video. Thus it is indispensable to develop a novel prediction methodology for recognizing/predicting an incomplete activity from a video.

Activity recognition can be sensor-based or vision-based and this chapter addresses the problem of activity recognition using the later approach. In vision-based activity recognition, the computational procedure is done at four steps viz. human detection, tracking, activity recognition and then a complex activity evaluation.