Author(s) :

Sri Medha Pedaprolu.

Volume/Issue :

Abstract :

The recent increase in gun violence, especially in underprivileged communities, highlights the need for affordable concealed weapon detection to enhance public safety. While there are existing commercial solutions, such as millimeter wavelength imaging, they are expensive and may not be easily deployed in under-resourced communities. Thermal imaging can be used to identify concealed firearms through heat pattern analysis, highlighting hidden metallic objects based on temperature contrast. This study investigated four classification models: CNN, MobileNetV2 Transfer Learning model, YOLOv8 Classification, and a Vision Transformer, and one object detection model (YOLOv8), to compare their effectiveness on the Concealed Pistol Detection Dataset. The classification models were evaluated based on accuracy, precision, recall, and F1-scores. Among them, the YOLOv8 and Vision Transformer both achieved accuracy exceeding 96%, with the Vision Transformer appearing to be slightly more conservative when classifying images as ‘withgun’. The MobileNetV2 Transfer Learning model and CNN also performed extremely well, with accuracies exceeding 0.94. The detection models (both before and after augmentation) were evaluated based on precision, recall, mAP50, and mAP50-95, and the model without data augmentation performed better for concealed weapon detection for the dataset. The YOLOv8 models can be executed on a Raspberry Pi, while the Transformer model may require additional computing power, such as the NVIDIA Jetson Nano.