Improved bounding box regression loss for weapon detection systems using deep learning

Rising number of crime rate using firearms (such as open firing, robbery, suicides, mass shootings, homicides, threatening at gun point, etc.), has underscored the growing importance of timely detection of weapons. Bounding box regression, an essential element in object detection, plays a crucial role in accurately identifying and localizing these firearms. This paper introduces a Manhattan-Complete IoU (MCIoU) loss for bounding boxes, which demonstrates significantly faster convergence during training compared to other IoU losses. By incorporating MCIoU into one of the advanced object detection framework, (You Only Look Once) YOLOv7, the proposed work demonstrates consistent improvement on their performance across popular weapon detection benchmarks datasets such as Granda, Synthetic, Cam157, Internet Movie Firearms Database (IMFDB), Gun Movie and Monash. The most encouraging outcomes were obtained on a Gun movie dataset with a precision and recall value of 98.2% and 96.3% respectively, which is an appreciable improvement compared to the baseline YOLOv7 model. Experiments show that the best precision achieved was at 98.2% and mAP@.50:.95 of 42.5% over other existing IoUs.