Violence Detection Using One-Dimensional Convolutional Networks

Abstract

Violence detection in surveillance video processing is a useful capability helping discover abnormal events in a variety of places. Utilizing methods considering the accuracy and complexity simultaneously can provide systems suitable for real-time applications. In this paper, the traditional approach of extracting temporal features has been investigated, while by exploiting one-dimensional convolutional networks, a new approach is proposed, which extracts these features across consecutive frames properly. This low-complexity convolutional-based approach represents a series of frames with a robust feature vector, which can be applied for real-time applications. The experimental results on Hockey, ViolentFlow reveal the efficiency of this proposed method.