Learning From Decoupled Sensor Axes: Enhancing Cross-Dataset Human Activity Recognition
Zhipeng Tang, Shibo He, Cong Wang, Qian Chen, Jun Pan, Chaojie Gu
Zhejiang University Zhejiang College of Sports
阅读操作
确认中在文库中上传 PDF 后可生成中文音频讲解。
摘要与影响
Human activity recognition using wearable IMU sensors achieves high accuracy within individual datasets, yet generalization across different data collection settings remains a significant challenge. Existing domain adaptation methods typically treat multi-axis IMU data as a unified entity, aligning source and target distributions in joint space, which requires simultaneously matching marginal distributions of individual axes and their inter-axis correlations. This paper proposes a decoupling strategy that treats each sensor axis as an independent adaptation unit, reducing alignment complexity by eliminating the need to match inter-axis correlations. Through empirical analysis, we demonstrate that individual axes preserve discriminative information while exhibiting smaller domain gaps compared to joint multi-axis approaches. We present a Decouple-and-Ensemble framework that establishes axis correspondence through statistical alignment, trains independent classifiers with adversarial domain adaptation, and aggregates predictions through confidence-weighted fusion. Comprehensive evaluation across public HAR datasets with 20 source-target transfer pairs demonstrates that the proposed method achieves best accuracy in 18 scenarios, with accuracy improvements of 9.15% to 10.63% over state-of-the-art methods across four source datasets. Domain shift analysis confirms that single-axis adaptation reduces proxy A-distance from 1.90 to 1.18, supporting that decoupling reduces domain shift in cross-dataset transfer.
逐年被引趋势
暂无年度引用数据
关键指标
同类平均 = 1
同领域 · 同年份 · 同类型
Google Scholar 与 OpenAlex 的被引统计范围不同,数值存在差异属正常。
AI 辅助阅读
依据:摘要
可就本文提问;依据不足时会说明。
学术脉络
学科主题
计算机 / AIDomain Adaptation and Few-Shot Learning
Context-Aware Activity Recognition Systems · Gait Recognition and Analysis
参考文献 57
此处列出前 3 条