TOP-ALCM: A novel video analysis method for violence detection in crowded scenes

Despite the Video Violence Detection (VVD) plays a critical role in video surveillance, it is not trivial in crowded scenes due to the complexity and diversity of violence. Generally, the most typical features of violence are its drastic, disordered, and chaotic motion in contrast to non-violence. To capture these features for violence analysis in a video clip, we propose a novel Angle-level Co-occurrence Matrix (ALCM). Given a video volume, we treat it as a tensor of rank 3, which consists of a bound of fibers in one plane. ALCM records the co-occurrence of two specific quantized angle levels between fibers with their neighbors, which is the distribution of the co-occurrence of fiber pairs with specific similarities in one plane of the tensor of rank 3. To completely characterize the violence in the volume, we compute three ALCMs for three orthogonal planes to form a TOP-ALCM, respectively. We also propose both conventional and deep-learning-based VVD frameworks, in which the former one leverages the features such as entropy, homogeneity, and energy computed from TOP-ALCM for classification, while the latter one directly uses CNN to classify the TOP-ALCMs. Experimental results analysis demonstrates that the proposed TOP-ALCM outperforms the state-of-the-art methods for VVD.