Weapon Detection System

Abstract: Rising incidents of weapon-related violence in public spaces have intensified demand for automated surveillance systems capable of real-time weapon identification. This paper presents a complete end-to-end weapon detection and alerting pipeline built on the YOLOv8 nano architecture. The system identifies three weapon categories — knives, pistols, and rifles — from live camera streams, webcam feeds, and static images. To address the well-known false-positive problem in weapon detection, we introduce a structured hard-negative mining strategy: background images sourced from diverse indoor environments (offices, homes, classrooms, public spaces) are automatically screened by the model; frames that generate spurious detections are harvested and re-introduced into the training corpus with empty labels. The dataset, comprising approximately 11,500 annotated images across three weapon classes, is further augmented with 1,500+ negative samples. After five iterative fine-tuning rounds with mosaic, mixup, and copy- paste augmentation, the final model achieves an overall mAP@50 of 0.727 (knife 0.75, pistol 0.55, rifle 0.88). An accompanying alert subsystem enforces temporal debouncing — requiring N consecutive high-confidence frames before triggering — and dispatches annotated evidence images to operators via Telegram or configurable HTTPS webhooks. A Flask-based local dashboard provides a browser interface for image analysis, video processing, and live webcam monitoring. Evaluation results, qualitative analyses, known limitations (notably a pistol precision of approximately 0.41), and recommendations for production deployment are discussed in detail.