Algorithm for Safety Helmet Recognition and Parameter Optimization in Ship Engine Control Rooms

Abstract

Ship engine rooms present difficult visual conditions, including low illumination, oil mist or steam, equipment occlusion, overhead viewpoints and densely distributed hazardous zones. To link safety-helmet recognition with hazard-zone warning under these conditions, this study proposes a YOLOv13s model-selection and electronic-fence fusion method. YOLOv5s, YOLOv8s, YOLOv11s, YOLOv13s and YOLOv13-EEAttne were compared on the same ship-engine-room helmet dataset under unified training settings. Scene subsets were further constructed from image brightness, target-box pixel area, inter-box overlap and target count, covering normal and low illumination, near-, mid- and far-range targets, and no, partial and severe occlusion. YOLOv13s was re-evaluated on these subsets, and its detection outputs were connected to foot-point electronic-fence rules and multi-frame confirmation. YOLOv13s achieved a final Precision of 0.9559, mAP@0.5 of 0.9201 and mAP@0.5:0.95 of 0.6795, outperforming the other compared models overall. The model contained approximately 9.55 million parameters, required 21.62 GFLOPs at an input size of 640 and reached 43.25 FPS on an RTX 4060 Laptop GPU. Scene-subset validation showed mAP@0.5:0.95 values of 0.672 and 0.597 under normal and low illumination, and 0.714, 0.673 and 0.470 for near-, mid- and far-range targets. These results indicate that distant small targets remain the main challenge and that YOLOv13s, combined with electronic-fence rules, provides a feasible detection basis for safety-helmet warning and hazardous-zone monitoring in ship engine rooms.