Deep Learning for Violence Detection in Surveillance: The Role of Transfer Learning and Pre-Trained Models

In recent years, the use of CCTV footage for proactive crime prevention has surged, particularly in public places like airports, train stations, and malls. However, the efficacy of these surveillance systems becomes questionable for their substantial reliance on human resources which may lead to erroneous or delayed responses. This research takes a comprehensive approach for violence detection which was previously limited to binary classification, by categorizing violence into four distinct classes: abuse, arson, assault, and fight. It employs transfer learning combined with computer vision techniques for violence detection and classification in video footages. The study compares the performance of four pre-trained neural networks, namely DenseNet121, VGG16, MobileNet, and Xception. The dataset created for this research is compiled from three datasets available on Kaggle. The results reveal that the Xception model performed comparatively better achieving the highest AUC score of 98.39%, while the VGG16 model attained the lowest AUC score of 96%. In addition to the AUC score, precision, recall, and fl-score are employed as performance metrics. Transfer learning with convolutional neural networks (CNN) significantly reduced computational requirements and time. Automating the detection and categorization of violent behavior through the employed approach has the potential to reduce the risk of fatalities and injuries in public areas. It also improves the speed and accuracy of threat detection, enabling swift preventive actions by authorities.