电子测量与仪器学报

赵普,武一.改进 KAPAO 的人体关键点检测[J].电子测量与仪器学报,2023,37(7):177-185

改进 KAPAO 的人体关键点检测

Improved human keypoints detection for KAPAO

DOI：

英文关键词:keypoints detection non-maximum suppression attention feature fusion data augmentation

基金项目:国家自然科学基金(51977059)、河北省自然科学基金(E2020202042)项目资助

作者	单位
赵普	1. 河北工业大学电子信息工程学院
武一	1. 河北工业大学电子信息工程学院，2. 河北工业大学电子与通信工程国家级实验教学示范中心

Author	Institution
Zhao Pu	1. School of Electronics Information Engineering, Hebei University of Technology
Wu Yi	1. School of Electronics Information Engineering, Hebei University of Technology，2. National Demonstration Center for Experimental,Electronic & Communication Engineering,Hebei University of Technology

摘要点击次数: 808

全文下载次数: 815

中文摘要:

针对人体关键点检测存在检测精确度低的不足,在 KAPAO( keypoints and pose as objects)网络的基础上进行改进。使用 PoseTrans(pose transformation)进行数据增强,提高网络的泛化性;针对特征融合能力的不足,设计融合注意力机制的 BiFPN (Bi-directional feature network)模块充分融合不同语义特征,提高网络对深层语义信息和浅层语义信息的融合能力;在网络输出阶段设计自适应扩张卷积模块,将不同扩张率的输出分支进行自适应融合,有效获得图像的全局信息;在网络的后处理部分设计 SDR-NMS(soft DIOU relocation non-maximum suppression)替代传统的 NMS,保留最优的关键点预测框。实验结果表明,网络的 AP 分数提高了 4. 8%,AP 为 68. 6%,检测速度为 19. 1 ms。网络精确度和检测速度均具有较好的表现性。

英文摘要:

For the lack of detection accuracy for human keypoints, it is improved on the basis network of KAPAO (keypoints and pose as objects). The generalization of network is improved by the enhance data method of PoseTrans ( pose transformation); for the lack of characteristic fusion capabilities, the BiFPN (Bi-directional feature network) module is designed to fully integrate different semantic characteristic to improve the integration ability of deep semantics information and shallow semantic information; the adaptive expansion convolution module is designed to adaptive fusion different expansion rates of output branch during the network output phase, it effectively obtains the global information of the image; in order to retain the optimal key point prediction box, the traditional NMS is replaced by SDR-NMS ( soft DIOU relocation non-maximum suppression ) during the post-processing part of the network. The experimental results show that the AP score was increased by 4. 8%, the AP was 68. 6%, and the detection speed was 19. 1 ms. The accuracy and detection speed of network have better performance.

查看全文查看/发表评论下载PDF阅读器