ELYRA: Edge-Optimized Lightweight YOLO with ROI Alignment for Real-Time SOP Compliance Monitoring in Crowded Environments

Automated monitoring of attire SOP (standard operating procedure) compliance in crowded environments remains challenging due to occlusions, transient detections, and computational constraints on edge devices. This paper presents ELYRA (Edge-Optimized Lightweight YOLO with ROI Alignment), a real-time framework integrating an optimized YOLO11 (You Only Look Once version 11) pipeline with adaptive ROI (region of interest) segmentation for efficient PPE (personal protective equipment) compliance monitoring in dense laboratory settings. ELYRA employs dynamic ROI segmentation to isolate individuals within complex scenes, minimizing unnecessary computations. The framework, optimized with NCNN (Nihui Convolutional Neural Network) inference, enables lightweight deployment on embedded IoT (Internet of Things) devices without compromising detection fidelity. A persistent tracking module correlates compliance status temporally, reducing transient errors until individuals exit the surveillance zone. The system is trained on a domain-specific, annotated dataset and supports synchronized file sharing via NFSv4 (Network File System version 4). Our experimental results show that ROI-based detection improves mean FPS (frames per second) by 102% over full-frame analysis, while NCNN optimization further boosts FPS by 77.2%, achieving 82.21% CPU utilization. These findings validate ELYRA’s effectiveness for scalable, real-time compliance monitoring under edge device constraints.