Real-Time Firearm and Knife Detection Using Fine-Tuned YOLOv8m with Timestamped Alerting and Evidence Capture for CCTV-Style Video Surveillance
Year: 2026
Type: preprint
Abstract: This paper presents the design, implementation, and validation of a firearm and knife detection system using a fine-tuned YOLOv8m object detection model for CCTV-style video surveillance. The system uses a weapon image dataset sourced from Kaggle, converts Pascal VOC annotations into YOLO format, and performs dataset preparation using Roboflow. Model training and inference are implemented using Python, PyTorch, OpenCV, and NVIDIA RTX 4050 GPU acceleration. The reported validation results include a mean precision of 0.9174, mean recall of 0.7530, mAP@50 of 0.7936, mAP@75 of 0.7019, and mAP@50–95 of 0.6020. The application supports confidence-based weapon detection, timestamped terminal alerts, and automatic evidence-frame capture. The paper discusses practical limitations, including false detections, occlusion, low-light conditions, and the need for further evaluation before real-world deployment. This manuscript describes an applied undergraduate research project and has not been independently peer-reviewed.
Institution: MIT Art, Design and Technology University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7236348528
