Multi-body tracking and behavior analysis

Abstract

This thesis presents research work in the area of application of visual surveillance to traffic videos. Viable solutions to some problems of visual surveillance viz; multi-target tracking and behavior analysis have been presented in this thesis. The focus here has been on developing a real time multi-target tracking algorithm for moving objects in videos from a stationary camera used for traffic monitoring and developing a behavior recognition system which does not require training to detect behaviors. This is an important consideration as it is difficult to get large data sets of adverse behaviors. The key tasks of a visual surveillance system considered in this thesis are (1) detecting moving objects, (2) tracking their motion and shape in spite of splits and merges, and (3) classifying them and recognizing their behaviors. Conceptually the system has been divided into above three main modules, which interact with one another to achieve the final goal of behavior recognition.