基于大模型的钻井现场人体姿态估计方法研究  

Research on Human Pose Estimation Method in Drilling Site Based on Large Model

作　　者：刘兆年 连远锋[2] 师印亮 王宁 姜彬 LIU Zhaonian;LIAN Yuanfeng;SHI Yinliang;WANG Ning;JIANG Bin(CNOOC Research Institute Ltd.,Beijing 100028,China;China University of Petroleum(Beijing),Beijing 102249,China)

机构地区：[1]中海油研究总院有限责任公司 [2]中国石油大学(北京)

出　　处：《钻采工艺》2025年第1期104-112,共9页Drilling & Production Technology

基　　金：中海油集团公司“十四五”重大科技项目“数据质量和安全自动化与数据分析预处理技术研究”(编号:KJGG-2024-15-0501)。

摘　　要：准确的人体姿态估计对钻井现场员工行为的监测和安全预警至关重要。针对钻井平台现场监控视频中存在高反光、高模糊和遮挡问题,提出一种基于双向特征融合的人体姿态估计模型,通过构建一种高效的双向特征融合机制,在ViT预训练模型的基础上引入空洞金字塔池化技术捕捉的图像多尺度空间特征。该机制可同时关注ViT预训练模型内部特征、多尺度空间特征以及两者间的交互特征,实现多类特征的高效集成。实验结果表明,通过与基准模型HRNet的对比,文章方法在KAP和KAR上分别实现了3.6%和4.1%的显著提升。同时,在南海某平台的智能监控系统中对所提出的模型进行应用测试,仍然显示出较高的准确性,为后续深入研究员工不安全行为的智能分析提供了精确的动作估计基础。Accurate human pose estimation is important for monitoring the drilling crew behavior and providing safety warnings.Aiming at the issues of high reflection,high blurring and occlusion in the on-site monitoring videos of drilling platform,a human pose estimation model based on bidirectional feature fusion is proposed.By constructing an efficient two-way feature fusion mechanism,the atrous spatial pyramid pooling(ASPP)is introduced based on the Vision Transformer(ViT)pre-training model to capture the multi-scale spatial features of the image.This mechanism can simultaneously focus on the internal features of the ViT pre-training model,multiscale spatial features,and interactive features between the two,achieving efficient integration of multiple types of features.According to experimental results,this approach significantly improved KAP and KAR by 3.6%and 4.1%,respectively,when compared to the benchmark model HRNet.Moreover,the application test of the proposed model in an intelligent monitoring system on a platform in Nanhai continued to demonstrate high accuracy,providing a precise action estimation foundation for further in-depth research on intelligent analysis of employees'risky behaviors.

关 键 词：人体姿态估计 预训练大模型 空洞金字塔池化 双向特征融合 

分 类 号：TE28[石油与天然气工程—油气井工程] TP391.41[自动化与计算机技术—计算机应用技术]