This paper details the development of a framework for fast and active video-based surveillance system (FAVSS) for remote monitoring of events. The key idea is to analyze video streams, detect the presence of human and recognize predefined subset of behaviors/activities (i.e., violence, suspicious activities and shop lifting, etc.) in the scene. The proposed system records only the meaningful events thus saving storage requirements and analysis time. It also notifies the concerned person via email with attachment of the captured images or picture message to mobile phone or stream the video to an web application if a human/face is detected or predefined activities/behavior is recognized. Development of such intelligent and event driven recording and communication of data is of extreme interest in building inexpensive, semi-autonomous surveillance system for the Department of Homeland Security as well as private use. The detection of human and body parts were performed by adopting the routines from the OpenCV and other freely available software resources. To model the behavior/activities a new video representation called visual elements was introduced. A host of individual classifier, ensemble classier from the RapidMiner tool box was used to model human behaviors/activities. To further improve the accuracies and robustness of models obtained using various machine learning methods were fused based on the measures of diversity. Initial tests using Human Behavior Video Database (HBViD) consisting of 221 video sequences shows up to 71.5% (5 fold cross validation) accuracy. In depth performance evaluation of the models were performed to illustrate the utility of each classifier in modeling behaviors/activities from video data and finding classifier that can be used for the deployment of the system.
展开▼
