MoWLD: a robust motion image descriptor for violence detection

Automatic violence detection from video is a hot topic for many video surveillance applications. However, there has been little success in designing an algorithm that can detect violence in surveillance videos with high performance. Existing methods typically apply the Bag-of-Words (BoW) model on local spatiotemporal descriptors. However, traditional spatiotemporal features are not discriminative enough, and also the BoW model roughly assigns each feature vector to only one visual word and therefore ignores the spatial relationships among the features. To tackle these problems, in this paper we propose a novel Motion Weber Local Descriptor (MoWLD) in the spirit of the well-known WLD and make it a powerful and robust descriptor for motion images. We extend the WLD spatial descriptions by adding a temporal component to the appearance descriptor, which implicitly captures local motion information as well as low-level image appear information. To eliminate redundant and irrelevant features, the non-parametric Kernel Density Estimation (KDE) is employed on the MoWLD descriptor. In order to obtain more discriminative features, we adopt the sparse coding and max pooling scheme to further process the selected MoWLDs. Experimental results on three benchmark datasets have demonstrated the superiority of the proposed approach over the state-of-the-arts.