Intelligent Surveillance System for Knife Detection Based on YOLOv8 Neural Networks
TL;DR
The design and implementation of an intelligent surveillance system based on the YOLOv8nano architecture for real-time detection of sharp objects, specifically knives, in images and video streams is presented.
Abstract
Cold weapon violence in urban environments represents a growing problem that demands innovative technological solutions for its prevention and control. This article presents the design and implementation of an intelligent surveillance system based on the YOLOv8nano architecture for real-time detection of sharp objects, specifically knives, in images and video streams. The model was trained through transfer learning on an annotated dataset of 2,078 images obtained from the public platform DatasetNinja, employing a two-phase strategy: freezing the ImageNet-pretrained backbone followed by full fine-tuning. Experimental results obtained on the test set demonstrate outstanding performance with a precision of 97.04%, a recall of 91.35%, and an mAP@0.5 value of 97.62%, outperforming several comparable approaches reported in the literature. The system achieves an inference speed of 79.9 FPS with an average time of 12.5 ms per image on an NVIDIA RTX 3050 GPU, which enables its integration with an early warning module based on temporal persistence to automatically notify security personnel. Shopping centers, stadiums, and high-risk urban areas are proposed as target deployment scenarios, although their validation in these specific environments remains as future work, given that the current evaluation is limited to static images from the dataset used.