| 18 | 0 | 30 |
| 下载次数 | 被引频次 | 阅读次数 |
短时人体行为信号信息量有限，通常呈现非平稳性，且缺乏足够的判别信息。通常可通过变分模态分解结合的希尔伯特变换（VMD-HT）构建三维的希尔伯特谱以解决非平稳问题。然而，三维表示在识别过程中易出现特征重叠，从而造成短时特征的损失。因此，可将三维的希尔伯特谱解析并投影到二维子空间，以避免遮挡并实现对短时行为信号的更精细表达。此外，单一二维投影难以充分反映三维的希尔伯特谱中复杂的时频结构，而多视图解析能够从不同表示角度刻画其时频结构。基于此，本文提出一种基于三视图解析的希尔伯特谱方法，通过投影生成三个互补的二维特征子空间，实现对短时人体行为信号的精准有效表征。考虑到不同子空间对行为识别的贡献存在差异，引入Squeeze-and-Excitation注意力模块（SE-Layer）结合ResNet-34网络，构建SE-ResNet-34模型，对三视图特征进行自适应加权与深层特征学习。该方法在PAMAP2、NOITOM和WARD数据集上分别取得95.07%、98.09%和98.57%的识别率，验证了其在一定程度上能够增加短时信号的利用率，识别性能可与长时行为识别方法相当。
Abstract：Short-term human activity signals contain limited information and often exhibit nonstationarity and lack of sufficient discriminative features. A common approach to addressing nonstationarity is to construct a three-dimensional Hilbert spectrum using variational mode decomposition combined with the Hilbert transform(VMD-HT). However, in the recognition process, the three-dimensional representation is prone to feature overlap, which may result in the loss of short-term characteristics. To alleviate this issue, the three-dimensional Hilbert spectrum is analyzed and projected into two-dimensional subspaces to avoid occlusion and enable more fine-grained representation of short-term activity signals. Moreover, a single two-dimensional projection is often insufficient to fully capture the complex time-frequency structure inherent to the three-dimensional Hilbert spectrum, whereas multi-view projection can characterize such structures from different representational perspectives. Motivated by this observation, this paper proposes a three-view Hilbert spectrum analysis method, in which three complementary two-dimensional feature subspaces are generated through projection to achieve accurate and effective representation of short-term human activity signals. Considering that different subspaces contribute unequally to activity recognition, a Squeeze-and-Excitation attention module(SE-Layer)is incorporated into a ResNet-34 network to construct an SE-ResNet34 model, enabling adaptive weighting and deep feature learning of the three-view features. Experimental results on the PAMAP2,NOITOM, and WARD datasets yield recognition accuracies of 95.07%, 98.09%, and 98.57%, respectively, demonstrating that the proposed method can, to a certain extent, enhance the utilization of short-term signals and achieve recognition performance comparable to that of long-term activity recognition methods.
[1]SHUO X, SHENGZHI W, ZHENZHEN H, et al. Two-stream Transformer Network for Sensor-Based Human Activity Recognition[J], Neurocomputing, 2022, vol. 512:253-268.
[2]XIAOKANG Z, WEI L, KAI I K W, et al. Deep-learning-enhanced Human Activity Recognition For Internet of Healthcare Things[J]. IEEE Internet of Things Journal, 2020, 7(7):6429-6438.
[3]SAKORN M, ANUCHIT J, MEKRUKSAVANICH S, et al. Lstm Networks Using Smartphone Data for Sensor-Based Humanactivity Recognition in Smart Homes[J]. Sensors, 2021, 21(5):1636.
[4]YAGUCHI K, IKARIGAWA K, KAWASAKI R, et al. Human Activity Recognition Using Multi-Input Cnn Model with Fft Spectrograms[C]//Adjunct Proceedings of the 2020 ACM International Joint Conference on Pervasive and Ubiquitous Computing and Proceedings of the 2020 ACM International Symposium on Wearable Computers, Mexico:Quintana Roo, 2020:364-367.
[5]SHIBO Z, YAXUAN L, SHEN Z, et al. Deep Learning in Human Activity Recognition with Wearable Sensors:A Review on Advances[J]. Sensors 2022, 22:1476.
[6]LINTAO D, MICHAEL L, ZHIGUO W, et al. Human Lower Limb Motion Capture and Recognition Based on Smartphones[J].Sensors, 2022, 22(14):5273.
[7]苏本跃,张利,何清旋,等.基于小波特征匹配的短时人体行为识别[J].系统仿真学报,2023,35(1):158-168.
[8]XU H, LIU J, HU H, et al. Wearable Sensor-Based Human Activity Recognition Method with Multi-Features Extracted from Hilbert-Huang Transform[J]. Sensors 2016, 16:2048.
[9]SUN S. A Survey of Multi-View Machine Learning[J]. Neural Computing and Applications, 2013, 23(7):2031-2038.
[10]HE K, ZHANG X, REN S, et al. Deep Residual Learning for Image Recognition[C]//2016, IEEE Conference On Computer Vision and Pattern Recognition. 2016:770-778, LasVegas,NV,USA:IEEE.
[11]HU J, SHEN L, SUN G. Squeeze-and-Excitation Networks[J]//IEEE Conference on Computer Vision and Pattern Recognition.2020,82(8):2011-2023.
[12]REISS A, STRICKER D. Introducing a New Benchmarked Dataset for Activity Monitoring[C]//In 2012 16th International Symposium on Wearable Computers, Newcastle, United Kingdom:IEEE, 2012:108-109.
[13]YANG Y A, JAFARI R, SASTRY S S, et al. Distributed Recognition of Human Actions Using Wearable Motion Sensor Networks[J]. Journal of Ambient Intelligence and Smart Environments, 2009, 1(2):103-115.
[14]HYUGA T, KEI K, KOKI T, et al. Sensor-Based Activity Recognition Using Frequency Band Enhancement Filters and Model Ensembles[J]. Sensors, 2023, 23(3):1465.
[15]YUAN R, ZHANG Y, WANG L, et al. Human Activity Recognition with a Multibranch Network Based on Cnn and Lstm[C]//Proc. Spie 12987, Third International Conference On Computer Technology, Information Engineering, and Electron Materials(CTIEEM 2023), Shenzhen,China:SPZE,2024:2.
[16]VUONG H T, DOAN T, TAKASU A. Deep Wavelet Convolutional Neural Networks for Multimodal Human Activity Recognition Using Wearable Inertial Sensors[J]. Sensors , 2023, 23-24:9721.
[17]WIELAND C, PANKRATIUS V. Enhancing Sensor-Based Human Activity Recognition with Multiresolu-tion Wavelet-Attention[J]. IEEE Sensors Journal, 2025, 25(14):26920-26930.
[18]王善荣,罗彪,盛敏.基于改进的希尔伯特谱的短时人体行为识别[J].安庆师范大学学报（自然科学版）,2025,31(1):69-74.
基本信息:
中图分类号:TP391.41
引用信息:
[1]陈欣,张清雅,葛志新,等.基于三视图解析希尔伯特谱的短时人体行为识别方法[J].合肥大学学报,2026,43(02):83-92.
基金信息:
安徽省高校优秀科研创新团队（2023AH010056）
2025-12-30
2025
2026-01-22
2026-02-27
2026
1
2026-04-28
2026-04-28