Detection of abnormal human activities in video using combining classifiers

Abstract

Control of public places such as stadiums, banks etc. by CCTV cameras require so many manpower. These systems are vulnerable to error because of fatigue or human error. So, we must develop a system that be able to detect normal and abnormal activities besides aware the security forces about the situation to provide better protection. In this research we introduce an integrated descriptor where after cutting videos, we perform Fourier transform to map the movement information to frequency spectrum/domain and then extract information. We apply Gabor filter for produced frequency spectrum and extract features. Then, we have dimension reduction with JMI feature selection and training with SVM and finally, behavior detection. Non-isolated background from foreground causes that our system works like a referee in sports games for activity detection when we need to foreground and background information simultaneously. With feature selection step in dimension reduction and hierarchical SVM, we show the superiority of this method according to accuracy and speed criterion compared to other methods.