Comparison of YOLOv3-tiny and YOLOv4-tiny in the Implementation Handgun, Shotgun, and Rifle Detection Using Raspberry Pi 4B

Criminal activities frequently involve carryable weapons such as handguns, shotguns, and rifle classes. Frequently, the targets of these weapons that are captured are concealed from plain sight by the people of the crowd. The detection process for these weapons can be assisted by using deep learning. In this case, we intend to identify the model of the firearm that was detected. This research aims to apply one of the deep learning concepts, namely You Only Look Once (YOLO). The authors use versions of YOLOv3-tiny and Yolov4-tiny for the detection and classification of types of weapons, which are one of the fastest and most accurate methods of object detection, outperforming other detection algorithms. However, both require heavy computer architecture. Therefore, YOLOv3-tiny and YOLOv4-tiny, lighter versions of YOLOv3, can be solutions for smaller architectures. YOLOv3-tiny and YOLOv4-tiny have higher FPS, which is supposed to yield faster performance. Since YOLOv3-tiny and YOLOv4-tiny are modified versions of YOLOv3, the accuracy is improved, and YOLOv3 is already outperforming Faster Single Shot Detector (SSD) and Faster Region with Convolutional Neural Network (R-CNN). The authors employ YOLOv3-tiny and YOLOv4-tiny due to the fact that the Frame Per Second (FPS) and Mean Average Precision (mAP) performance of both approaches are superior in object detection. The study found that YOLOv3-tiny had a high FPS and low mAP performance: an average Intersection over Union (IoU) score of 71.54%, an accuracy of 90%, a recall score of 78%, an F1 score of 84%, and an mAP of 86.7%. While YOLOv4-tiny has low FPS and high mAP: an average IoU score of 73.19%, an accuracy of 90%, a recall score of 84%, an F1 score of 87%, and an mAP of 90.7%.