Deep Learning Framework for Firearms Detection YOLOv8 Optimized for Smart Environment Society

As cities become smarter, the need for real-time threat detection in surveillance systems increases. Firearm-related incidents in public areas are highly critical, where even a small delay can lead to serious consequences. In this study, we propose a deep- learning-based framework for firearm detection using the YOLOv7 model. To improve performance, we integrated a Convolutional Block Attention Module (CBAM) and a Bi-directional Feature Pyramid Network (BiFPN). The model was trained on a diverse dataset of 18,500 images and 6,200 negative samples to reduce false alarms. It achieves high accuracy with a 95.5% [email protected] and operates at 48 FPS. The false-positive rate was reduced to 5.2%, which was significantly lower than the baseline. For real-world use, the system was optimized for NVIDIA Jetson Nano using pruning and INT8 TensorRT, enabling real-time performance at 28 frames per second (FPS). When a firearm is detected, the system sends instant alerts through MQTT to smart city platforms, allowing for a quick response and automation. Overall, this approach combines Deep Learning, Edge AI, and IoT to enable efficient and reliable real-time surveillance, thereby making smart environments safer.