Weapon Object Detection Using Quantized YOLOv8

Video surveillance is essential for creating a secure and hassle-free environment in all areas of life. It helps identify theft, detect unusual events in crowded locations, and monitor the suspicious behavior of individuals. However, monitoring surveillance cameras manually is quite challenging, and thus, fully automated surveillance with smart video-capturing capabilities is gaining popularity. This approach uses deep learning methodology to remotely monitor unusual actions with accurate information about the location, time of occurrence, and identification of criminals. Detecting criminal conduct in public settings is difficult due to the complexity of real-world scenarios. CCTV cameras can record suspicious incidents in public areas, such as carrying weapons, which helps authorities to take preventive measures to protect citizens. The proposed system employs the state-of-the-art YOLOv8 model for real-time weapon detection, which is faster, more accurate, and better than YOLOv5. To ensure fast performance, the weights of YOLOv8 were quantized. In our experiments, we evaluated the performance of the YOLOv8 and YOLOv5 models for weapon detection. The mean Average Precision (mAP) value achieved using YOLOv8 was 90.1%, which outperformed the mAP value of 89.1% obtained with YOLOv5. Furthermore, by applying weight quantization to the YOLOv8 model, we reduced the inference time by 15% compared to the original YOLOv8 configuration.