Enhancing Violence Detection in Surveillance Footage Using a Fine-Tuned YOLOv8n Model with Domain-Specific Optimization

Abstract

Early detection of violent behavior is essential for preventing criminal activities such as murders, rapes, and theft. It is a critical component of public safety, recognizing hostile behavior across diverse scenarios. Motivated by the significance of this domain, we aimed to contribute to this area by investigating the efficacy of using the YOLOv8n model for detecting violent activities. We trained and validated the model using two violence detection datasets, and its performance was evaluated using multiple metrics, including mean Average Precision (mAP), precision, and recall. The results demonstrate that the proposed fine-tuned YOLOv8n model outperforms the pre-trained version, achieving a mAP of 0.95 on Dataset-1 and 0.96 on Dataset-2, representing a significant improvement over the baseline. Additionally, the model accurately detected weapons, such as knives and guns, achieving high precision and recall. These findings have important implications for improving security features, assisting law enforcement, and advancing surveillance technology in real-world applications.