Statistical Features-Based Violence Detection in Surveillance Videos

Research over detecting anomalous human behavior in crowded scenes has created much attention due to its direct applicability over a large number of real-world security applications. In this work, we propose a novel statistical feature descriptor to detect violent human activities in real-world surveillance videos. Standard spatiotemporal feature descriptors are used to extract motion cues from videos. Finally, a discriminative SVM classifier is used to classify violent/non-violent scenes present in the videos with the help of feature representation formed out of the proposed statistical descriptor. Efficiency of the proposed approach is tested on crowd violence and hockey fight benchmark datasets.