A Deep Learning Framwork with Attention-Guided Feature Learning for Robust Object Detection in Surveillance Systems

Weapons detection in smart video surveillance systems is critical for ensuring public safety and security in various environments. This paper proposes a novel approach to enhancing weapon detection accuracy by integrating attention mechanisms with the YOLOv5 object detection framework. The purpose of this research is to address the challenges associated with accurate and efficient weapon detection, such as occlusions, varying lighting conditions, and the presence of multiple objects in complex scenes. Our methodology involves modifying the YOLOv5 archi-tecture to incorporate attention mechanisms, including, Convolutional Block Attention Module (CBAM), Squeeze-and- Excitation (SE) and Channel Attention (CA). We train the modified models on the SOHAS dataset and a custom dataset containing images of weapons and persons holding weapons. Key findings from our experiments demonstrate that integrating attention mechanisms with YOLOv5 significantly improves weapon detection accuracy compared to the baseline YOLOv5 model. Specifically, the models equipped with attention mechanisms exhibit enhanced capability in focusing on relevant regions of the input images while sup-pressing noise and irrelevant features, leading to more precise and reliable detections. Each module demonstrated enhancements in identifying dangerous weapons across various ranges. The CA module represents the highest mAP, achieving a 95.6% accuracy rate, marking a 3.1 % enhancement over the base YOLOv5s model. Additionally, there was a slight increase in recognition speed by 0.11 seconds compared to the original YOLOv5s model