Detection of Violence Behavior using Deep Learning Technique

Abstract

Deep learning-based violence detection approaches from video streams are a rapidly expanding subject of study, this is due to the necessity to develop appropriate and automated violence detection techniques based on visual data obtained from security cameras mounted in various locations. In this research, a modified pre-trained deep learning technique named Convolutional Neural Network-Visual Geometry Group16 (CNN-VGG16) are employed to implement a low complex and uncomplicated model for the detection of violence. The transfer learning technique is applied to take advantage of the pre-knowledge VGG16 in detecting shapes and edges. The final layers of the default VGG16 structure are replaced to accommodate the purpose of the research. The efficiency of this approach is evaluated using two datasets (Automatic Violence Detection Dataset (AvdDS) and Surveillance fight dataset (SfDS)). The experimental outcomes prove the efficiency of the proposed model against alternative methods. In experiment accuracy results for Dataset 1 94% and 91% after applying Canny filter, as for dataset 2, the accuracy is 99% and 92% using canny filter. Also In this paper, the effect of applying edges detector on classification accuracy results and processing time for training the model is observed.