Detecting Violent Crowds using Temporal Analysis of GLCM Texture

The severity of sustained injury resulting from assault-related violence can be minimized by reducing detection time [10,28]. However, it has been shown that human operators perform poorly at detecting events found in video footage when presented with simultaneous feeds [30]. We utilize computer vision techniques to develop an automated method of violence detection that can aid a human operator. We observed that violence in city centre environments often occur in crowded areas, resulting in individual actions being occluded by other crowd members. Measures of visual texture have shown to be effective at encoding crowd appearance [8, 23, 31]. Therefore, we propose modelling crowd dynamics using changes in crowd texture. We refer to this approach as Violent Crowd Texture (VCT). Real-world surveillance footage of night time environments and the violent flows [16] dataset were tested using a random forest classifier to evaluate the ability of the VCT method at discriminating between violent and non-violent behaviour. Our method achieves ROC values of 0.98 and 0.91 for the realworld and violent flows data respectively.