基于门控渐进高斯滤波网络的人体运动估计

杨旭升; 杨基业; 张文安

doi:10.16383/j.aas.c250250

基于门控渐进高斯滤波网络的人体运动估计

doi: 10.16383/j.aas.c250250 cstr: 32138.14.j.aas.c250250

杨旭升^{1, 2,},
杨基业^{1, 2,},
张文安^{1, 2,}

1.
浙江工业大学信息工程学院杭州 310023
2.
浙江全省复杂系统智能感知与控制重点实验室杭州 310023

基金项目: 国家自然科学基金(62473335, U25A20456, W2421117), 杭州市科技发展计划(2022AIZD0080) 资助

详细信息

作者简介:
杨旭升：浙江工业大学信息工程学院副教授. 主要研究方向为信息融合估计, 人体运动估计和目标定位. E-mail: xsyang@zjut.edu.cn

杨基业：浙江工业大学信息工程学院硕士研究生. 主要研究方向为人体运动估计, 信息融合估计. E-mail: yangjiye@zjut.edu.cn

张文安：浙江工业大学信息工程学院教授. 主要研究方向为多源信息融合估计及应用. 本文通信作者. E-mail: wazhang@zjut.edu.cn

计量
- 文章访问数: 5
- HTML全文浏览量: 3
- 被引次数: 0
出版历程
- 收稿日期: 2025-06-09
- 录用日期: 2025-12-31
- 网络出版日期: 2026-03-16

Gated Progressive Gaussian Filtering Network for Human Motion Estimation

YANG Xu-Sheng^{1, 2
,},
YANG Ji-Ye^{1, 2
,},
ZHANG Wen-An^{1, 2
,}

1.
College of Information Engineering, Zhejiang University of Technology, Hangzhou 310023
2.
Zhejiang Key Laboratory of Intelligent Perception and Control for Complex Systems, Hangzhou 310023

Funds: Supported by National Natural Science Foundation of China (62473335, U25A20456, W2421117) and Science and Technology Development Program of Hangzhou (2022AIZD0080)

More Information

Author Bio:
YANG Xu-Sheng　Associate professor at the College of Information Engineering, Zhejiang University of Technology. His research interests include information fusion estimation, human motion estimation, and target positioning

YANG Ji-Ye　Master student at the College of Information Engineering, Zhejiang University of Technology. His research interests include human motion estimation and information fusion estimation

ZHANG Wen-An　Professor at the College of Information Engineering, Zhejiang University of Technology. His research interests include multi-source information fusion estimation and its applications. Corresponding author of this paper

摘要

摘要: 针对表面肌电信号和人体运动等模型误差引起的运动估计性能下降问题, 提出一种基于门控渐进高斯滤波网络(gated progressive Gaussian filtering network, GPGF-net) 的人体运动估计方法, 以实现对模型误差的补偿以及提升运动估计的精度. 首先, 设计门控记忆机制来调控信息流, 以学习出系统状态的长期依赖与观测信号的时变特性, 从而动态调整误差补偿项的分布参数. 其次, 通过融合贝叶斯滤波与深度学习的优势, 引入渐进式观测更新策略到GPGF-net, 以减小非线性近似误差以及增强模型鲁棒性. 最后, 通过人体肢体运动估计实验表明, 相较于现有方法, GPGF-net显著提高了估计精度, 均方根误差降低15.62 %, 相关系数R²提升5.08 %, 验证了所提方法的有效性.
- 高斯滤波网络 /
- 模型误差补偿 /
- 人体运动估计 /
- 表面肌电信号
Abstract: To address the issue of degraded motion estimation performance caused by surface electromyography model and human motion model errors, this paper proposes a human motion estimation method based on gated progressive Gaussian filtering network (GPGF-net), aiming to effectively compensate for model errors and improve the accuracy of motion estimation. First, a gated memory mechanism is designed to regulate the flow of information, enabling the model to learn long-term dependencies in system states and time-varying characteristics of observation signals, thereby dynamically adjusting the distribution parameters of the error compensation term. Secondly, by integrating the advantages of Bayesian filtering and deep learning, a progressive multi-step observation update strategy is introduced into the GPGF-net to reduce nonlinear approximation errors and enhance model robustness. Finally, experiments on human limb motion estimation demonstrate that, compared with existing methods, GPGF-net significantly improves estimation accuracy, reducing the root mean square error by 15.62 % and increasing the correlation coefficient R² by 5.08 %, thus validating the effectiveness of the proposed method.
- Gaussian filtering network /
- model error compensation /
- human motion estimation /
- surface electromyography

HTML全文

图 1 GPGF-net总体框架

Fig. 1 Overall framework of GPGF-net

下载: 全尺寸图片幻灯片

图 2 门控渐进高斯滤波网络结构

Fig. 2 Architecture of the gated progressive Gaussian filter network

下载: 全尺寸图片幻灯片

图 3 补偿渐进观测更新

Fig. 3 Compensated progressive observation update

下载: 全尺寸图片幻灯片

图 4 实验系统与设备布置((a)数据采集场景; (b)传感器佩戴布局)

Fig. 4 Experimental system and equipment setup((a) Data acquisition scenario; (b) Sensor placement layout)

下载: 全尺寸图片幻灯片

图 5 特征提取模块

Fig. 5 ECA feature extraction module

下载: 全尺寸图片幻灯片

图 6 角度估计和误差曲线

Fig. 6 Angle estimation and error curves

下载: 全尺寸图片幻灯片

图 7 五个模型的RMSE和R²

Fig. 7 RMSE and R² of the five models

下载: 全尺寸图片幻灯片

表 1 模型性能评价

Table 1 Evaluation of model performances

测试者	RMSE					R²
测试者	EKF	TCN-LSTM	LSTM-KF	PUKF-net	GPGF-net	EKF	TCN-LSTM	LSTM-KF	PUKF-net	GPGF-net
S1	20.332	16.632	13.654	12.940	11.349	0.753	0.823	0.892	0.906	0.920
S2	22.351	22.787	18.678	16.980	15.436	0.622	0.631	0.765	0.824	0.846
S3	20.662	20.735	16.996	15.434	14.031	0.718	0.712	0.821	0.842	0.876
S4	18.624	17.238	14.353	13.048	11.851	0.772	0.811	0.886	0.898	0.910
S5	20.648	20.895	20.366	18.502	17.995	0.710	0.699	0.750	0.765	0.765
S6	22.992	16.642	13.644	12.956	11.778	0.614	0.821	0.898	0.905	0.913
S7	19.965	17.965	17.164	15.669	14.245	0.721	0.802	0.819	0.836	0.872
S8	18.356	17.356	15.241	13.958	12.689	0.776	0.815	0.853	0.886	0.903
S9	20.162	19.235	18.468	16.753	15.256	0.718	0.754	0.774	0.828	0.842
S10	20.169	21.240	18.655	17.235	15.658	0.716	0.657	0.770	0.795	0.802
S11	21.362	20.109	17.869	15.668	14.245	0.651	0.723	0.782	0.832	0.870
S12	18.623	15.543	13.458	12.487	11.351	0.770	0.843	0.901	0.909	0.914
均值	21.462	19.698	16.504	15.376	13.926	0.710	0.751	0.827	0.844	0.869
标准差	3.297	3.533	2.494	1.925	2.145	0.057	0.072	0.053	0.047	0.049

下载: 导出CSV

表 2 四种模型的计算复杂度与性能

Table 2 The computational complexity and performance of four models

	TCN-LSTM	LSTM-KF	PUKF-net	GPGF-net
FLOPs	912186	1219448	719448	699656
Params	320133	342337	256511	245718
推理时间/ms	1.096	1.469	0.865	0.811

下载: 导出CSV

表 3 不同信噪比下的模型性能

Table 3 Model performance under different SNRs

模型	指标	无噪声	SNR=15 dB	SNR=10 dB
LSTM-KF	RMSE	16.504	19.236	21.874
LSTM-KF	$ {\rm{R}}^{2} $	0.827	0.752	0.677
GPGF-net	RMSE	13.926	15.683	17.721
GPGF-net	$ {\rm{R}}^{2} $	0.869	0.801	0.788

下载: 导出CSV

参考文献(29)

[1]	Guo W, Zhao Z, Zhou Z, et al. Hand kinematics, high-density sEMG comprising forearm and far-field potentials for motion intent recognition. Scientific Data, 2025, 12(1): 445 doi: 10.1038/s41597-025-04749-8
[2]	Hu F, Qian M, He K, Zhang W A, Yang X. A novel multi-feature fusion network with spatial partitioning strategy and cross-attention for armband-based gesture recognition. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2024, 32: 3878−3890 doi: 10.1109/TNSRE.2024.3487216
[3]	张鋆豪, 何百岳, 杨旭升, 张文安. 基于可穿戴式惯性传感器的人体运动跟踪方法综述. 自动化学报, 2019, 45(8): 1439−1454 doi: 10.16383/j.aas.c180367 Zhang Jun-Hao, He Bai-Yue, Yang Xu-Sheng, Zhang Wen-An. A Review on Wearable Inertial Sensor Based Human Motion Tracking. Acta Automatica Sinica, 2019, 45(8): 1439−1454 doi: 10.16383/j.aas.c180367
[4]	李自由, 赵新刚, 张弼, 等. 基于表面肌电的意图识别方法在非理想条件下的研究进展. 自动化学报, 2021, 47(5): 955−969 doi: 10.16383/j.aas.c200263 Li Zi-You, Zhao Xingang, Zhang Bi, et al. Review of sEMG-based motion intent recognition methods in non-ideal conditions. Acta Automatica Sinica, 2021, 47(5): 955−969 doi: 10.16383/j.aas.c200263
[5]	Buchanan T S, Lloyd D G, Manal K, et al. Neuromusculoskeletal modeling: estimation of muscle forces and joint moments and movements from measurements of neural command. Journal of Applied Biomechanics, 2004, 20(4): 367−395 doi: 10.1123/jab.20.4.367
[6]	Sartori M, Reggiani M, Farina D, et al. EMG-driven forward-dynamic estimation of muscle force and joint moment about multiple degrees of freedom in the human lower extremity. PLoS One, 2012, 7(12): e52618 doi: 10.1371/journal.pone.0052618
[7]	Zhao Y, Zhang Z, Li Z, et al. An EMG-driven musculoskeletal model for estimating continuous wrist motion. IEEE Transactions on Neural Systems and Rehabilitation Engineering, 2020, 28(12): 3113−3120 doi: 10.1109/TNSRE.2020.3038051
[8]	Tang G, Sheng J, Wang D, et al. Continuous estimation of human upper limb joint angles by using PSO-LSTM model. IEEE Access, 2020, 9: 17986−17997 doi: 10.1109/access.2020.3047828
[9]	张文安, 林安迪, 杨旭升, 俞立, 杨小牛. 融合深度学习的贝叶斯滤波综述. 自动化学报, 2024, 50(8): 1502−1516 doi: 10.16383/j.aas.c230457 Zhang Wen-An, Lin An-Di, Yang Xu-Sheng, Yu Li, Yang Xiao-Niu. A survey on Bayesian filtering with deep learning. Acta Automatica Sinica, 2024, 50(8): 1502−1516 doi: 10.16383/j.aas.c230457
[10]	Avian C, Prakosa S W, Faisal M, et al. Estimating finger joint angles on surface EMG using manifold learning and long short-term memory with attention mechanism. Biomedical Signal Processing and Control, 2022, 71: 103099 doi: 10.1016/j.bspc.2021.103099
[11]	Zhao D, Ma Y, Meng J, et al. MCR-ALS-based muscle synergy extraction method combined with LSTM neural network for motion intention detection. Frontiers in Neurorobotics, 2023, 17: 1174710 doi: 10.3389/fnbot.2023.1174710
[12]	Chen C, Guo W, Ma C, et al. sEMG-based continuous estimation of finger kinematics via large-scale temporal convolutional network. Applied Sciences, 2021, 11(10): 4678 doi: 10.3390/app11104678
[13]	Fan J, Zhang K, Huang Y, et al. Parallel spatio-temporal attention-based TCN for multivariate time series prediction. Neural Computing and Applications, 2023, 35(18): 13109−13118 doi: 10.1007/s00521-021-05958-z
[14]	Hu F, Zhang L, Yang X, Zhang W A. EEG-based driver fatigue detection using spatio-temporal fusion network with brain region partitioning strategy. IEEE Transactions on Intelligent Transportation Systems, 2024, 25(8): 9618−9630 doi: 10.1109/TITS.2023.3348517
[15]	Wang F, Liang W, Afzal H M R, et al. Estimation of lower limb joint angles and joint moments during different locomotive activities using the inertial measurement units and a hybrid deep learning model. Sensors, 2023, 23(22): 9039 doi: 10.3390/s23229039
[16]	Du J, Liu Z, Dong W, et al. A Novel TCN-LSTM Hybrid Model for sEMG-Based Continuous Estimation of Wrist Joint Angles. Sensors, 2024, 24(17): 5631 doi: 10.3390/s24175631
[17]	Han J, Ding Q, Xiong A, et al. A state-space EMG model for the estimation of continuous joint movements. IEEE Transactions on Industrial Electronics, 2015, 62(7): 4267−4275 doi: 10.1109/TIE.2014.2387337
[18]	杨旭升, 吴江宇, 胡佛, 张文安. 基于渐进高斯滤波融合的多视角人体姿态估计. 自动化学报, 2024, 50(3): 607−616 doi: 10.16383/j.aas.c230316 Yang Xu-Sheng, Wu Jiang-Yu, Hu Fo, Zhang Wen-An. Multi-view human pose estimation based on progressive Gaussian filtering fusion. Acta Automatica Sinica, 2024, 50(3): 607−616 doi: 10.16383/j.aas.c230316
[19]	Yang X, Zhang W A, Yu L. Fractional Kalman filters. Automatica, 2025, 178: 112383 doi: 10.1016/j.automatica.2025.112383
[20]	Miao Z, Shi H, Zhang Y, et al. Neural network-aided variational Bayesian adaptive cubature Kalman filtering for nonlinear state estimation. Measurement Science and Technology, 2017, 28(10): 105003 doi: 10.1088/1361-6501/aa7d70
[21]	Yang X, Zhang W A, Liu A, et al. Linear fusion estimation for range-only target tracking with nonlinear transformation. IEEE Transactions on Industrial Informatics, 2019, 16(10): 6403−6412 doi: 10.1109/tii.2019.2955931
[22]	Coskun H, Achilles F, DiPietro R, et al. Long short-term memory kalman filters: Recurrent neural estimators for pose regularization. Proceedings of the IEEE International Conference on Computer Vision, 20175524−5532
[23]	Zhao C, Sun L, Yan Z, et al. Learning Kalman Network: A deep monocular visual odometry for on-road driving. Robotics and Autonomous Systems, 2019, 121: 103234 doi: 10.1016/j.robot.2019.07.004
[24]	Liu J, Wang Z, Xu M. DeepMTT: A deep learning maneuvering target-tracking algorithm based on bidirectional LSTM network. Information Fusion, 2020, 53: 289−304 doi: 10.1016/j.inffus.2019.06.012
[25]	Yang X, Yin S, Zhang W A, Hu F, Yu L. Asynchronous Gaussian filtering fusion for human motion estimation based on RGB-D cameras. IEEE Sensors Journal, 2023, 23(22): 28044−28054 doi: 10.1109/JSEN.2023.3323869
[26]	杨旭升, 李福祥, 胡佛, 张文安. 基于肌电-惯性融合的人体运动估计: 高斯滤波网络方法. 自动化学报, 2024, 50(5): 991−1000 Yang Xu-Sheng, Li Fu-Xiang, Hu Fo, Zhang Wen-An. Human motion estimation based on EMG-inertial fusion: A Gaussian filtering network approach. Acta Automatica Sinica, 2024, 50(5): 991−1000
[27]	杨旭升, 王雪儿, 汪鹏君, 张文安. 基于渐进无迹卡尔曼滤波网络的人体肢体运动估计. 自动化学报, 2023, 49(8): 1723−1731 doi: 10.16383/j.aas.c220523 Yang Xu-Sheng, Wang Xue-Er, Wang Peng-Jun, Zhang Wen-An. Estimation of human limb motion based on progressive unscented Kalman filter network. Acta Automatica Sinica, 2023, 49(8): 1723−1731 doi: 10.16383/j.aas.c220523
[28]	Yan S, Liang Y, Zheng L, et al. Explainable gated Bayesian recurrent neural network for non-Markov state estimation. IEEE Transactions on Signal Processing, 2024, 72: 4302−4317 doi: 10.1109/TSP.2024.3390139
[29]	Julier S, Uhlmann J, Durrant-Whyte H F. A new method for the nonlinear transformation of means and covariances in filters and estimators. IEEE Transactions on Automatic Control, 2000, 45(3): 477−482 doi: 10.1109/9.847726