基于多尺度增量学习的单人体操动作中关键点检测方法  

Keypoint Detection Method for Single Person Gymnastics Actions Based on Multi-Scale Incremental Learning

作　　者：江佳鸿 夏楠 李长吾[1] 周思瑶 于鑫淼 JIANG Jia-hong;XIA Nan;LI Chang-wu;ZHOU Si-yao;YU Xin-miao(School of Information Science and Engineering,Dalian Polytecnic University,Dalian,Liaoning 116034,China)

机构地区：[1]大连工业大学信息科学与工程学院,辽宁大连116034

出　　处：《电子学报》2024年第5期1730-1742,共13页Acta Electronica Sinica

基　　金：教育部产学合作协同育人项目(No.220603231024713)。

摘　　要：人体关键点检测是计算机视觉的热点研究领域.目前,对于体操动作关键点检测,仍存在检测精度不足及缺乏细节部位检测能力等问题.为了提升检测精度,本文设计了一种多分辨率网络,该网络在浅层具备较大感受野,同时能够利用高分辨率通道增强细节特征的提取能力.为实现对手部及脚部关键点的检测,设计了一种增量学习网络.该网络融合了多分辨率网络的浅层特征并利用自建数据集计算深层特征以提升网络对手部及脚部关键点的检测能力.最后对两个网络输出结果进行合并.计算机仿真表明,多分辨率网络在COCO2017关键点检测数据集上达到了94.4%的准确率,并且增量学习网络能够在训练数据较少的情况下实现对细节部位关键点的准确检测.Keypoint detection of human body is a hot research area in computer vision.At present there exist some problems for keypoint detection in gymnastics actions,such as insufficient detection accuracy and lack of capability to detect detailed body parts.In order to improve the detection accuracy,this paper proposes a multi-resolution network that has a larger receptive field in the shallow layers and can utilize high-resolution channel to enhance the extraction of detailed features.To achieve the detection of keypoints of hands and feet,an incremental learning network is designed.The network fuses the shallow features of the multi-resolution network and computes deep features using a gymnastics actions self-built dataset,so that the detection ability of keypoints on hands and feet is improved.Finally,the output results of the two subnetworks are concated.Computer simulations demonstrate that the multi-resolution network achieves an accuracy rate of 94.4% on the COCO2017 keypoint detection dataset,and the incremental learning network can accurately detect keypoints of detailed body parts with fewer training data.

关 键 词：人体关键点检测 体操动作 多分辨率网络 增量学习 权重融合 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]