A Comparative Analysis of Violence Detection in Surveillance Video using TLD, GOTURN, YOLOv5, MDNet, KCF, DeepSort, and FairMOT

Abstract

The increasing deployment of surveillance systems has necessitated the development of robust algorithms for automatic violence detection. This paper presents a comparative study of various object detection and tracking algorithms—TLD, GOTURN, YOLOv5, MDNet, KCF, DeepSort, and FairMOT—for detecting violent activities in surveillance video footage. The evaluation focuses on accuracy, speed, and robustness under different environmental conditions. Our experiments demonstrate that while YOLOv5 combined with FairMOT provides superior accuracy, other algorithms offer competitive results in terms of processing speed and robustness, suggesting that the choice of algorithm should be based on specific application requirements.