YOLOv11-Based Object Detection for Personnel Operation Procedures on Offshore Platforms

Abstract

Offshore platforms need reliable visual monitoring, but their operating spaces are crowded, lighting conditions change quickly, and edge devices often have limited computing resources. This paper presents an improved YOLOv11 detector for recognizing workers, key equipment, and operation-related actions in offshore platform procedures. The feature extraction layer is rebuilt with a C3k2-SMPCGLU module. SMPConv is used to shift sampling positions toward irregular worker and equipment contours, while the CGLU gate keeps useful local responses and suppresses background interference from pipes, decks, and machinery. A task-aligned dynamic detection head (TADDH) is then introduced. It uses shared convolution with GroupNorm and learnable Scale layers to reduce redundant parameters and handle features at different scales. Multi-level interaction features further guide the offset generation of DCNv2 and the dynamic feature selection of the classification branch, making classification confidence and localization quality more consistent. Experiments on an in-house offshore-platform dataset show that the proposed model reaches 0.861 precision, 0.743 recall, 0.798 F1-score, and 0.784 mAP@0.5, outperforming the compared YOLOv9m, YOLOv10m, YOLOv11m, and YOLOv12m models. The results indicate that the method can provide a practical visual basis for procedure monitoring in resource-constrained offshore scenarios.