Crowd Violence Detection in Videos Using Deep Learning Architecture

Violence is considered as a serious and sensory issue and visualizing the violence is therefore subjected to ethical implications. Surveillance cameras may be placed in several locations to monitor and capture people’s activities. From the massive data collected through surveillance camera, detecting violence is a tedious task. Recently computer vision-based methods are used to identify violent activities automatically. Researchers were using key frame extraction technique to delete duplicate frames and detect violence from the input video. In the recent day’s, they use a pretrained network for selecting visually important frames and perform violence detection. In a constrained environment these networks perform well, but the accuracy drops for inputs acquired in unconstrained environment. In this work, a keyframe extraction technique is initially performed to remove duplicate consecutive frames followed by a pretrained VGG 16 deep learning architecture on Hockey fight dataset for detecting and recognizing violence. By minimizing duplicate frames, the cost of training data and thus the computational cost is reduced. Previously proposed sequential CNN architectures used a single kernel size for feature selection and classification tasks, whereas the proposed VGG16 architecture uses 3 kernels. The proposed VGG 16 network is found to be improving on following aspects: model size is small, speed is faster, faster convergence is done by residual learning, generalization is better and degradation problem is solved. Experimental methodology tested on Hockey fight dataset has shown that the suggested architecture is able to perform well even in an unconstrained environment.