Human Abnormal Activity Recognition from Video Using Motion Tracking

The detection of violent behavior in the public environment using video content has become increasingly important in recent years due to the rise of violent incidents and the ease of sharing and disseminating video content through social media platforms. Efficient and effective techniques for detecting violent behavior in video content can assist authorities with identifying potential hazards, preventing crimes, and promoting public safety. Violence detection can also help to mitigate the psychological damage caused by viewing violent content, particularly in vulnerable populations such as infants and victims of violence. We have proposed an algorithm to calculate new descriptors using the magnitude and orientation of optical flow (MOOF) in the video. Descriptors are extracted from MOOF based on four binary histograms each by applying various weighted thresholds. These descriptors are used to train Support Vector Machine (SVM) and classify the video as violent or nonviolent. The proposed algorithm has been tested on the publicly available Hockey Fight Dataset and Violent Flow dataset. The results demonstrate that the proposed descriptors outperform the state-of-the-art algorithms with an accuracy of 91.5% and 78.5% on the Hockey Fight and Violent Flow datasets, respectively.