Detection method of PPE wearing and small target positioning for offshore operators: The improved YOLOv11 model and targets recognition

海上作业人员个人防护装备穿戴及小目标定位检测方法:改进的YOLOv11模型及目标识别

阅读:1

Abstract

Focusing on the practical challenges of insufficient samples, incomplete categories, and low detection accuracy (particularly for small targets) in Personal Protective Equipment (PPE) wearing condition monitoring for operators in offshore environments, this research investigates PPE targets detection for offshore operators using an improved YOLOv11 model. The optimized model integrates the time-frequency features enhancement module (Spatial Pyramid Pooling-Fast, SFEAF) into the model's backbone network, employs a statistical-driven dynamic gating attention module (Token Statistics Self-Attention, TSSA) to refine attention weight distribution in the original C2PSA module, and incorporates a Normalized Wasserstein Distance (NWD) loss function. These modifications collectively enhance the model's capability to detect PPE targets for offshore operators. To mitigate missed detection problem of small targets such as earplugs and gloves, a cascaded network of YOLOv11 and YOLOv11-Pose models is proposed for small targets detection. The solution involves extracting human key points through YOLOv11-Pose model, constructing spatial constraint regions via two-point area positioning method, enhancing small target features through localized region cropping and normalization, and performing secondary detection on refined regions using YOLOv11 model. The ablation experiments show that the mAP@0.5 value of the optimization model has been improved by 1.8 percentage points compared to the original model for all targets, and the precision rates for both positive and negative samples of small targets-earplugs and gloves-are respectively improved by 5.2%, 4.2%, 0.2%, and 3.7%. The superiority of the optimization method has been proved. Furthermore, secondary detection experiments on small targets yielded an average Missed Detection Recovery Rate (MRR) of 56.64%, and the effectiveness of the multi-model cascaded detection method has been verified.

特别声明

1、本页面内容包含部分的内容是基于公开信息的合理引用;引用内容仅为补充信息,不代表本站立场。

2、若认为本页面引用内容涉及侵权,请及时与本站联系,我们将第一时间处理。

3、其他媒体/个人如需使用本页面原创内容,需注明“来源:[生知库]”并获得授权;使用引用内容的,需自行联系原作者获得许可。

4、投稿及合作请联系:info@biocloudy.com。