轻量且高精度增强的姿态检测网络HG-YOLO

doi:10.11772/j.issn.1001-9081.2024121819

《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (12): 4004-4011.DOI: 10.11772/j.issn.1001-9081.2024121819

• 多媒体计算与计算机仿真 • 上一篇下一篇

轻量且高精度增强的姿态检测网络HG-YOLO

崔家礼¹, 刘永基¹, 李子贺¹, 郑瀚¹^,²

^1.北方工业大学信息学院，北京 100144
^2.广西高校人工智能与信息处理重点实验室（河池学院），广西河池 546300

收稿日期:2024-12-27 修回日期:2025-04-09 接受日期:2025-04-14 发布日期:2025-04-28 出版日期:2025-12-10
通讯作者: 郑瀚
作者简介:崔家礼（1975—），男，山东枣庄人，副研究员，博士，CCF会员，主要研究方向：图像处理、模式识别
刘永基（1999—），男，山东枣庄人，硕士研究生，主要研究方向：图像处理、模式识别
李子贺（2000—），男，天津人，硕士研究生，主要研究方向：图像处理、模式识别
郑瀚（1987—），女，江苏泰州人，副教授，博士研究生，主要研究方向：图像处理、计算机视觉。
基金资助:
广西高校人工智能与信息处理重点实验室（河池学院）开放研究基金资助项目(2024GXZDSY006)

HG-YOLO： lightweight and high-precision enhancement pose detection network

Jiali CUI¹, Yongji LIU¹, Zihe LI¹, Han ZHENG¹^,²

^1.School of Information Science and Technology，North China University of Technology，Beijing 100144，China
^2.Key Laboratory of AI and Information Processing of Guangxi Colleges and Universities （Hechi University），Hechi Guangxi 546300，China

Received:2024-12-27 Revised:2025-04-09 Accepted:2025-04-14 Online:2025-04-28 Published:2025-12-10
Contact: Han ZHENG
About author:CUI Jiali， born in 1975， Ph. D.， associate research fellow. His research interests include image processing， pattern recognition.
LIU Yongji， born in 1999， M. S. candidate. His research interests include image processing， pattern recognition.
LI Zihe， born in 2000， M. S. candidate. His research interests include image processing， pattern recognition.
ZHENG Han， born in 1987， Ph. D. candidate， associate professor. Her research interests include image processing， computer vision.
Supported by:
Open Fund of Key Laboratory of AI and Information Processing of Guangxi Colleges and Universities （Hechi University）(2024GXZDSY006)

摘要/Abstract

摘要：

在人体姿态检测任务中，现有的深度学习网络存在检测精度不足、网络参数复杂和计算成本高等问题，严重限制了它们的应用。为了解决这些问题，提出一种轻量且高精度的姿态检测改进网络HG-YOLO （High-precision and Ghost YOLO）。针对检测精度不足的问题，在HG-YOLO的主干网络，融合基于Transformer的检测网络RT-DETR （Real-Time DEtection TRansformer），并将大型可分离核注意力（LSKA）模块嵌入主干网络中，以在不增加内存占用和计算复杂性的基础上，提高网络应对复杂场景的特征提取能力，从而提高人体姿态的检测精度。针对网络参数复杂和计算成本高的问题，引入轻量化的Ghost卷积模块替换部分标准卷积，此外，在HG-YOLO的检测头部分，设计一种共享卷积检测头，以通过参数和权重共享机制减少卷积计算，从而降低网络的参数量和计算复杂度。在COCO （Common Objects in COntext） 2017-Keypoints数据集和CrowdPose数据集上的实验结果表明，与基准的YOLOv8-Pose网络相比，HG-YOLO的参数量减少了32%，浮点运算量减少了18%；在规模为小型（s）时，在COCO 2017-Keypoints数据集上，AP50 （Average Precision at OKS （Object Keypoint Similarity） of 0.50）提升了0.8个百分点，在CrowdPose数据集上，AP提升了2.9个百分点。可见，HG-YOLO不仅轻量，而且检测精度高，是人体姿态检测领域的优秀网络模型。

关键词: 姿态检测, 高精度增强, 大型可分离核注意力, 共享卷积检测头, 轻量化网络

Abstract:

In human pose detection task， the existing deep learning networks have the problems of insufficient detection precision， complex network parameters and high computational cost， which seriously limit their applications. To solve these problems， a lightweight and high-precision enhancement pose detection network HG-YOLO （High-precision and Ghost YOLO） was proposed. Aiming at the problem of insufficient detection precision， the Transformer-based detection network RT-DETR （Real-Time DEtection TRansformer） was integrated into the backbone part of HG-YOLO， and the Large Separable Kernel Attention （LSKA） module was embedded into the backbone， which improved feature extraction ability of the network to cope with the complex scenarios without increasing the memory occupation and computational complexity， thus improving the human pose detection precision. Aiming at the problem of complex network parameters and high computational cost， the lightweight Ghost convolution module was introduced to replace some of the standard convolutions， and furthermore， a shared convolution detection head was designed in the detection head part of HG-YOLO， which reduced the convolution computation through the parameter and weight sharing mechanism， thus reducing number of parameters and computational complexity of the network. Experimental results on the COCO （Common Objects in Context） 2017-Keypoints dataset and the CrowdPose dataset show that compared to the benchmark YOLOv8-Pose network， HG-YOLO reduces the parameters by 32% and the floating-point operations by 18%； and when the scale is s （small）， on the COCO 2017-Keypoints dataset， HG-YOLO has the AP50 （Average Precision at OKS （Object Keypoint Similarity） of 0.50） improved by 0.8 percentage points， on the CrowdPose dataset， HG-YOLO has the AP improved by 2.9 percentage points. It can be seen that HG-YOLO is not only lightweight but also has high detection precision， which is an excellent network model in the field of human pose detection.

Key words: pose detection, high-precision enhancement, Large Separable Kernel Attention (LSKA), shared convolutional detection head, lightweight network

中图分类号:

TP391.4

崔家礼, 刘永基, 李子贺, 郑瀚. 轻量且高精度增强的姿态检测网络HG-YOLO[J]. 计算机应用, 2025, 45(12): 4004-4011.

Jiali CUI, Yongji LIU, Zihe LI, Han ZHENG. HG-YOLO： lightweight and high-precision enhancement pose detection network[J]. Journal of Computer Applications, 2025, 45(12): 4004-4011.

图/表 10

参考文献 37

[1]	HUNG J M， CHIANG J Y， WANG K. Tennis player pose classification using YOLO and MLP neural networks［C］// Proceedings of the 2021 International Symposium on Intelligent Signal Processing and Communication Systems. Piscataway： IEEE， 2021： 1-2.
[2]	唐鲁婷，黄洪琼. 基于YOLOv7的轻量化水下目标检测算法［J］. 电光与控制， 2024， 31（9）： 92-97.
	TANG L T， HUANG H Q. A YOLOv7 based lightweight underwater target detection algorithm［J］. Electronics Optics and Control， 2024， 31（9）： 92-97.
[3]	杨锦辉，李鸿，杜芸彦，等. 基于改进YOLOv5s的轻量化目标检测算法［J］. 电光与控制， 2023， 30（2）： 24-30.
	YANG J H， LI H， DU Y Y， et al. A lightweight object detection algorithm based on improved YOLOv5s［J］. Electronics Optics and Control， 2023， 30（2）： 24-30.
[4]	TERVEN J， CÓRDOVA-ESPARZA D M， ROMERO-GONZÁLEZ J A. A comprehensive review of YOLO： from YOLOv1 to YOLOv8 and YOLO-NAS［J］. Machine Learning and Knowledge Extraction， 2023， 5（4）： 1680-1716.
[5]	NEWELL A， YANG K， DENG J. Stacked hourglass networks for human pose estimation［C］// Proceedings of the 2016 European Conference on Computer Vision， LNCS 9912. Cham： Springer， 2016： 483-499.
[6]	CHEN Y， WANG Z， PENG Y， et al. Cascaded pyramid network for multi-person pose estimation［C］// Proceedings of the 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2018： 7103-7112.
[7]	CAO Z， SIMON T， WEI S E， et al. Realtime multi-person 2D pose estimation using part affinity fields［C］// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2017： 1302-1310.
[8]	CHENG B， XIAO B， WANG J， et al. HigherHRNet： scale-aware representation learning for bottom-up human pose estimation［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2020： 5385-5394.
[9]	DONG C， TANG Y， ZHANG L. HDA-Pose： a real-time 2D human pose estimation method based on modified YOLOv8［J］. Signal， Image and Video Processing， 2024， 18（8/9）： 5823-5839.
[10]	ZHENG B， ZHANG H， JIN L. Research on multi-person pose estimation based on YOLO and decoupled multi-level feature layers fusion［C］// Proceedings of the 5th ACM International Conference on Multimedia in Asia. New York： ACM， 2023： No.58.
[11]	MOU F， REN H， WANG B， et al. Pose estimation and robotic insertion tasks based on YOLO and layout features［J］. Engineering Applications of Artificial Intelligence， 2022， 114： No.105164.
[12]	ZHANG Z， LU X， CAO G， et al. ViT-YOLO： Transformer-based YOLO for object detection［C］// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision Workshops. Piscataway： IEEE， 2021： 2799-2808.
[13]	ZHANG F， CAO W， WANG S， et al. Improved YOLOv4 recognition algorithm for pitaya based on coordinate attention and combinational convolution［J］. Frontiers in Plant Science， 2022， 13： No.1030021.
[14]	WU J， DONG J， NIE W， et al. A lightweight YOLOv5 optimization of coordinate attention［J］. Applied Sciences， 2023， 13（3）： No.1746.
[15]	GONG C， ZHANG Y， WEI Y， et al. Multicow pose estimation based on keypoint extraction［J］. PLoS ONE， 2022， 17（6）： No.e0269259.
[16]	LIU J， CAI Q， ZOU F， et al. BiGA-YOLO： a lightweight object detection network based on YOLOv5 for autonomous driving［J］. Electronics， 2023， 12（12）： No.2745.
[17]	JIANG Y， YANG K， ZHU J， et al. YOLO-RlePose： improved YOLO based on Swin Transformer and RLE-OKS loss for multi-person pose estimation［J］. Electronics， 2024， 13（3）： No.563.
[18]	MAJI D， NAGORI S， MATHEW M， et al. YOLO-Pose： enhancing YOLO for multi-person pose estimation using object keypoint similarity loss［C］// Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. Piscataway： IEEE， 2022： 2636-2645.
[19]	DAI Y， LIU W. GL-YOLO-Lite： a novel lightweight fallen person detection model［J］. Entropy， 2023， 25（4）： No.587.
[20]	LI X， GUO Y， PAN W， et al. Human pose estimation based on lightweight multi-scale coordinate attention ［J］. Applied Sciences， 2023， 13（6）： No.3614.
[21]	ZHANG Y， WANG Z， LI M， et al. SP-YOLO： an end-to-end lightweight network for real-time human pose estimation［J］. Signal， Image and Video Processing， 2024， 18（1）： 863-876.
[22]	DONG C， DU G. An enhanced real-time human pose estimation method based on modified YOLOv8 framework［J］. Scientific Reports， 2024， 14： No.8012.
[23]	ZHAO Y， LV W， XU S， et al. DETRs beat YOLOs on real-time object detection［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 16965-16974.
[24]	HAN K， WANG Y， TIAN Q， et al. GhostNet： more features from cheap operations［C］// Proceedings of the 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2020： 1577-1586.
[25]	WANG A， CHEN H， LIU L， et al. YOLOv10： real-time end-to-end object detection［C］// Proceedings of the 38th International Conference on Neural Information Processing Systems. Red Hook： Curran Associates Inc.， 2024： 107984-108011.
[26]	LAU K W， PO L M， REHMAN Y A UR. Large separable kernel attention： rethinking the large kernel attention design in CNN［J］. Expert Systems with Applications， 2024， 236： No.121352.
[27]	GUO M H， LU C Z， LIU Z N， et al. Visual attention network［J］. Computational Visual Media， 2023， 9（4）： 733-752.
[28]	LIN T Y， MAIRE M， BELONGIE S， et al. Microsoft COCO： common objects in context ［C］// Proceedings of the 2014 European Conference on Computer Vision， LNCS 8693. Cham： Springer， 2014： 740-755.
[29]	LI J， WANG C， ZHU H， et al. CrowdPose： efficient crowded scenes pose estimation and a new benchmark［C］// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2019： 10863-10872.
[30]	TIAN Z， CHEN H， SHEN C. DirectPose： direct end-to-end multi-person pose estimation［EB/OL］. ［2024-04-06］..
[31]	CAO Z， HIDALGO G， SIMON T， et al. OpenPose： realtime multi-person 2D pose estimation using part affinity fields［J］. IEEE Transactions on Pattern Analysis and Machine Intelligence， 2018， 43（1）： 172-186.
[32]	SUN K， XIAO B， LIU D， et al. Deep high-resolution representation learning for human pose estimation［C］// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2019： 5686-5696.
[33]	McNALLY W， VATS K， WONG A， et al. Rethinking keypoint representations： modeling keypoints and poses as objects for multi-person human pose estimation［C］// Proceedings of the 2022 European Conference on Computer Vision， LNCS 13666. Cham： Springer， 2022： 37-54.
[34]	LU P， JIANG T， LI Y N， et al. RTMO： towards high-performance one-stage real-time multi-person pose estimation ［C］// Proceedings of the 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway： IEEE， 2024： 1491-1500.
[35]	KUOK K， LIU X， YE J， et al. GDE-pose： a real-time adaptive compression and multi-scale dynamic feature fusion approach for pose estimation［J］. Electronics， 2024， 13（23）： No.4847.
[36]	DING J， NIU S， NIE Z， et al. Research on human posture estimation algorithm based on YOLO-Pose［J］. Sensors， 2024， 24（10）： No.3036.
[37]	LU T， CHENG K， HUA X， et al. KSL-POSE： a real-time 2D human pose estimation method based on modified YOLOv8-pose framework［J］. Sensors， 2024， 24（19）： No.6249.

网络	参数量/10⁶	AP50/%	AP75/%	APM/%	APL/%	AR/%
DirectPose^［30］	—	86.4	68.2	56.7	69.8	—
OpenPose^［31］	—	84.9	67.5	57.1	68.2	66.5
HRNet^［32］	28.5	86.3	70.4	59.6	73.9	—
KAPAO-s^［33］	12.6	88.4	70.4	58.6	71.7	71.2
YOLOv5-s	10.8	87.8	69.5	57.6	72.6	67.6
RTMO-s^［34］	9.9	88.8	73.6	61.1	75.7	70.9
YOLOv8-s	11.6	86.2	65.9	53.9	70.3	66.5
HG-YOLO-s	7.9	87.0	67.2	55.7	70.9	67.4
KAPAO-m	35.8	90.5	76.5	64.3	76.0	76.3
YOLOv5-m	29.3	90.7	75.8	63.4	77.1	72.8
RTMO-m	22.6	90.6	77.1	65.1	78.1	74.2
YOLOv8-m	26.4	88.8	72.3	60.5	73.6	71.8
HG-YOLO-m	19.6	90.5	73.7	61.7	74.7	72.4

网络	参数量/10⁶	AP50/%	AP75/%	APM/%	APL/%	AR/%
DirectPose^［30］	—	86.4	68.2	56.7	69.8	—
OpenPose^［31］	—	84.9	67.5	57.1	68.2	66.5
HRNet^［32］	28.5	86.3	70.4	59.6	73.9	—
KAPAO-s^［33］	12.6	88.4	70.4	58.6	71.7	71.2
YOLOv5-s	10.8	87.8	69.5	57.6	72.6	67.6
RTMO-s^［34］	9.9	88.8	73.6	61.1	75.7	70.9
YOLOv8-s	11.6	86.2	65.9	53.9	70.3	66.5
HG-YOLO-s	7.9	87.0	67.2	55.7	70.9	67.4
KAPAO-m	35.8	90.5	76.5	64.3	76.0	76.3
YOLOv5-m	29.3	90.7	75.8	63.4	77.1	72.8
RTMO-m	22.6	90.6	77.1	65.1	78.1	74.2
YOLOv8-m	26.4	88.8	72.3	60.5	73.6	71.8
HG-YOLO-m	19.6	90.5	73.7	61.7	74.7	72.4

网络	参数量/10⁶	AP/%	APE/%	APM/%	APH/%
OpenPose	—	—	62.7	48.7	32.3
HRNet	28.5	71.3	80.5	71.4	62.5
KAPAO-l	77.0	68.9	76.6	69.9	59.5
RTMO-s	9.9	67.3	73.7	68.2	59.1
YOLOv8-s	11.6	62.1	69.6	63.4	58.3
HG‑YOLO‑s	7.9	64.7	70.8	65.7	62.5
RTMO-m	22.6	71.1	77.4	71.9	63.4
YOLOv8-m	26.4	64.9	73.5	64.4	59.3
HG‑YOLO‑m	19.6	67.8	75.2	66.8	63.6

网络	参数量/10⁶	AP/%	APE/%	APM/%	APH/%
OpenPose	—	—	62.7	48.7	32.3
HRNet	28.5	71.3	80.5	71.4	62.5
KAPAO-l	77.0	68.9	76.6	69.9	59.5
RTMO-s	9.9	67.3	73.7	68.2	59.1
YOLOv8-s	11.6	62.1	69.6	63.4	58.3
HG‑YOLO‑s	7.9	64.7	70.8	65.7	62.5
RTMO-m	22.6	71.1	77.4	71.9	63.4
YOLOv8-m	26.4	64.9	73.5	64.4	59.3
HG‑YOLO‑m	19.6	67.8	75.2	66.8	63.6

网络	参数量/10⁶	AP50/%	AP75/%	GFLOPs/10⁹
HG‑YOLO‑s	7.9	87.0	67.2	24.7
GDE-POSE^［35］	8.8	77.3	—	27.5
HG‑YOLO‑m	19.6	90.5	73.7	76.2
YOLO-POSE^［36］	22.3	89.1	69.5	—
KSL-POSE^［37］	—	89.6	58.9	117.3

轻量且高精度增强的姿态检测网络HG-YOLO

HG-YOLO： lightweight and high-precision enhancement pose detection network

RichHTML

PDF

可视化

摘要/Abstract

引用本文

使用本文

图/表 10

参考文献 37

相关文章 5

编辑推荐

Metrics

组序	HG-Ghost	LSKA	LSCD	参数量/10⁶	GFLOPs/ 10⁹	AP（50：95）/ %
1	—	—	—	11.6	30.2	56.6
2	√	—	—	9.5	27.8	57.2
3	—	√	—	11.9	30.9	57.8
4	—	—	√	9.5	26.4	55.4
5	√	√	—	9.9	28.5	58.9
6	—	√	√	9.9	27.1	56.3
7	√	—	√	7.5	24.0	55.4
8	√	√	√	7.9	24.7	58.4

[1]	王宏, 钱清, 王欢, 龙永. 融合大核注意力卷积的轻量化图像篡改定位算法[J]. 《计算机应用》唯一官方网站, 2023, 43(9): 2692-2699.
[2]	钟志峰, 夏一帆, 周冬平, 晏阳天. 基于改进YOLOv4的轻量化目标检测算法[J]. 《计算机应用》唯一官方网站, 2022, 42(7): 2201-2209.
[3]	孙海燕, 陈云博, 封丁惟, 王通, 蔡兴泉. 基于注意力模型和轻量化YOLOv4的林业害虫检测方法[J]. 《计算机应用》唯一官方网站, 2022, 42(11): 3580-3587.
[4]	邓雄, 王洪春. 基于深度学习和特征融合的人脸活体检测算法[J]. 计算机应用, 2020, 40(4): 1009-1015.
[5]	汪检兵, 李俊. 基于OpenPose-slim模型的人体骨骼关键点检测方法[J]. 计算机应用, 2019, 39(12): 3503-3509.