Violent Behavior Detection in Surveillance Videos Using MoSIFT and SVM

Most of the recent research in action recognition is mainly focused on recognizing activities of human like clapping, jogging, running, etc. Detection of the violent behavior has been relatively less studied. The applications of violent behavior detection are huge and it can be used in surveillance videos, prisons, university campuses, international borders and anywhere where the cameras provide sufficient line of sight for a sensitive region. In this paper for finding violent behavior in surveillance videos, we used the popular Bag-of-visual-words framework. The feature descriptor that we used is MoSIFT and STIP that provide the feature vector for Bag-of-visual-words to detect violence using Support Vector Machine classifier. The linear classifier provides labels of both violent and non-violent scenes along with description of the video. The experimental evaluation is done with two different datasets while the model is implemented on Raspberry Pi. Experiment results suggest that we can achieve more than 90% accuracy with the MoSIFT descriptor.