Vision-Based Real-time Human Malicious Behavior Detection

Human detection and behavior analysis from surveillance videos is an active area of research in computer vision. Authorities and security administrators need a system that can detect human malicious behavior to take immediate necessary actions. In this paper, we propose an approach to detect the anomalous/malicious behavior of humans in the surveillance videos. The proposed approach models the human behavior using human joint motion information from skeleton sequence. We have divided the proposed approach into four sub-modules i.e. human detection and skeleton estimation, human ID assignment, feature extraction and classification. The proposed approach is evaluated on publicly available CASIA dataset in offline mode and accuracy of 90.81% has been achieved. The experimental results indicates that it can be exploited in real-time applications with low computational cost of 18 frames per second.