Improved YOLOv5 with Backbone Replacement to MobileNetv3 for Weapons Detection: Application for Smart Video Surveillance

In the field of computer vision technology recent advancements have made improvements, in object detection impacting domains such as transportation and security surveillance. One important use case involves the real time identification of weapons for law enforcement and monitoring purposes. This study focuses on comparing the performance of two object detection models; YOLOv5s and MobileNetv3 YOLOv5s, based on factors like accuracy, computational efficiency and parameter count. Our models were trained using the Sohas Weapon Dataset and Gun Movies Dataset to address orientations and lighting conditions. The main objective was to optimize the models for efficiency while maintaining an Average Precision (mAP). The results indicate a reduction in parameters, weight and GFLOPs; however, there was also a decrease, in mAP and recall metrics.