An efficient visual sensing method for electric bicycle detection in complex elevator environments based on VGWS YOLO

Abstract As electric bicycles (e-bikes) pose severe lithium-ion thermal runaway fire risks within residential elevators, vision-based edge sensing has become a critical safety measure for smart building Internet of Things (IoT). However, precise e-bike detection in elevator cabins is severely hindered by heavy target occlusion and the tight computing limits of embedded sensing hardware. To address these challenges, this paper proposes VGWS-YOLO , an ultra-lightweight detection network built upon the You Only Look Once version 11 nano (YOLOv11n) architecture for resource-constrained elevator visual perception. Three coordinated optimizations are integrated into the network: the novel VoV-Ghost Cross Stage Partial (VoV-GSCSP) module strengthens deep feature representation while cutting computational redundancy; Wise-IoU version 3 (WIoU-v3) loss optimizes bounding box regression under heavy occlusion; and customized Soft Non-Maximum Suppression (Soft-NMS) effectively reduces missing detections in crowded pedestrian-vehicle scenes. Comprehensive experiments were conducted on a self-built dataset containing 4,864 real elevator surveillance images. VGWS-YOLO requires only 2.428 M parameters and 6.098 GFLOPs , achieving a frame rate of 157.38 FPS . Its mean average precision at IoU = 0.5 (mAP50) and recall reach 93.9% and 95.3% , surpassing the YOLOv11n baseline by 3.3% and 5.2% , respectively. Ablation studies further validate the synergistic gains of all three proposed modules. Balancing an ultra-lightweight architecture with high detection precision, VGWS-YOLO shows promising potential as an edge-oriented computing solution for real-time e-bike fire risk monitoring in intelligent building systems; practical inference performance on embedded hardware requires further physical on-device measurement.