Automated Smoking Scene Detection in Videos

Abstract

Automated human action detection is crucial in application areas such as video indexing, video surveillance, security applications and many more. Therefore many solutions have been proposed for the problem in the past. Significant observation of these solutions is that almost all the solutions address detection of basic human actions such as shaking hands , sitting down , raising arm , based on structure of the action pattern of the action. No considerable attention has been paid for more meaningful activities like smoking , fighting , riding bike etc. Nevertheless, ultimate expectation of human action detection is to identify more meaningful activities. It is very unlikely that a person is interested in searching for human actions such as waving arms or sitting down , however he/she would do for an activity like smoking , fighting . Therefore it is essential to address human activity detection problem properly for the completeness in human action detection domain. In this thesis, we discuss the necessity of paying attention on meaningful human activity detection and we propose a prototype solution using activity smoking . Since our solution focus on a set of context-specific information rather than focus on structure of action patterns, it is more like an evidence collecting approach. Therefore, our solution is robust to dynamic and cluttered backgrounds, occlusion and variations in action performance.