文献类型：期刊文献

中文题名：基于重参数化大核卷积的高分辨率姿态估计

英文题名：High-Resolution Pose Estimation Based on Reparameterized Large Kernel Convolution

作者：陈佳艺[1];黄晓宇[1];吴胜昔[1];王学武[1]

机构：[1]华东理工大学能源化工过程智能制造教育部重点实验室,上海200237

年份：2025

卷号：51

期号：3

起止页码：341

中文期刊名：华东理工大学学报(自然科学版)

外文期刊名：Journal of East China University of Science and Technology

收录：;北大核心:【北大核心2023】；

基金：国家自然科学基金(62076095)。

语种：中文

中文关键词：姿态估计;重参数化大核卷积;HRNet;感受野;特征融合

外文关键词：pose estimation;reparameterized large kernel convolution;HRNet;receptive field;feature fusion

摘要：尽管人体姿态估计领域的研究已取得显著进展,但面对动态场景变化、目标遮挡及背景复杂等难题,实现高精度、强鲁棒性的姿态估计依然面临巨大挑战。为解决这些问题,特别是关键点遮挡、重合及复杂环境干扰问题,本文提出了一种融合大核卷积技术的高分辨率人体姿态估计模型(RepLK-HRNet)。该模型的核心在于特征提取网络的独特设计,通过引入重参数化大核卷积策略,增强了模型捕捉多尺度、多层次特征信息的能力,同时通过调整网络结构,显著降低了参数量和计算复杂度。实验结果表明,相较于传统的高分辨率网络(HRNet)模型,RepLK-HRNet模型在标准数据集MS COCO2017上的精度提高了1.83%,在遮挡数据集OCHuman上的精度提高了23.7%,计算复杂度参数Params和GFLOPs分别下降了63.84%、37.69%。RepLK-HRNet模型在常规及遮挡、关键点混淆等条件下的人体姿态估计精度均实现了显著提升,展现了出色的鲁棒性和泛化能力,同时还满足了实际应用中对计算效率和存储空间的要求。
Although significant progress has been made in the field of human pose estimation,it still faces enormous challenges for achieving high-precision and robust pose estimation for the case of dynamic scene changes,occlusions,and complex backgrounds.To address these issues—particularly keypoint occlusion,overlap,and interference from complex environments—this paper proposes a high-resolution human pose estimation model incorporating large kernel convolution techniques,named RepLK-HRNet.The core innovation of the proposed model lies in its unique design of the feature extraction network,which introduces a reparameterized large kernel convolution strategy to enhance the model's ability in capturing multi-scale and multi-level feature information.Meanwhile,the network architecture is optimized to significantly reduce the number of parameters and computational complexity.Experimental results demonstrate that,compared to the traditional HRNet model,the RepLK-HRNet model achieves an improvement of 1.83%in accuracy on the standard MS COCO 2017 dataset and an increase of 23.7%in accuracy on the occlusion dataset OCHuman,while reducing Params by 63.84%and GFLOPs by 37.69%.These results indicate that RepLK-HRNet significantly improves pose estimation accuracy under general,occluded,and keypoint-confused conditions,showcasing excellent robustness and generalization capabilities.Moreover,it meets practical application demands in terms of computational efficiency and memory usage.