Firearm Detection Using Deep Learning

Surveillance Cameras (SCs) are great support in crime investigation and proximity alarms which play a critical role in public safety and peace. The major drawback is their limited use in producing evidences in judicial court but they were not used in providing early firearms detection to stop attacks in crime scenes. Traditional firearm detection techniques use X-Rays for detecting weapons in limited coverage area that fail in detecting non-metallic weapons. The main motivation of this research is developing a deep learning object detection model to early detect firearms in crime scenes and alert the corresponding authorities. Designing an efficient and accurate object detection model that localize and classify the firearm classes with overcoming variations in shape, size, appearance and occlusions is a real challenge. We collected a dataset of different firearm classes from Kaggle and Google Open Images and manually annotated about 2300 samples. The dataset consists of variety of firearms classes handgun, revolver, rifle along with knife and person classes. YOLOv5 (You Only Look Once) is a unified object detector which detects objects without losing their precision and accuracy. All the YOLOv5 models are built from scratch and generate final models that achieved 89.4%, 70.1%, 80.5% of precision, recall and mAP@0.5 respectively and achieved the highest 94.1% precision per individual class.