Activity-Conditioned Cross-Subject Alignment via Time–Frequency Cross-View Learning for Human Activity Recognition
Cross-subject distribution shifts caused by differences in body morphology, motion patterns, and device-wearing conditions remain a major challenge for inertial sensor-based human activity recognition (HAR). This study proposes CSCV-TF, an activity-conditioned cross-subject time-frequency cross-view learning framework designed to improve recognition of unseen subjects. CSCV-TF employs independent temporal and frequency Transformers with fixed late fusion to learn complementary representations. During training, Subject-Class Balanced Sampling ensures the stable availability of same-activity, different-subject pairs, while a bidirectional contrastive objective constrains temporal and frequency representations under activity-semantic conditions. The sampling strategy, projection heads, and contrastive objective are removed during inference, leaving the inference architecture and computational complexity unchanged from those of TF-Late. Under the official subject-independent UCI HAR protocol, CSCV-TF achieves an Accuracy of 94.75 ± 0.55% and a Macro-F1 of 94.81 ± 0.55% across five random seeds, improving Macro-F1 by 1.03 percentage points over TF-Late, which shares the identical inference backbone. Controlled comparisons with conventional supervised contrastive learning indicate that general contrastive regularization accounts for a substantial proportion of the improvement, while the cross-subject positive-pair definition and cross-view alignment provide modest but complementary gains. Independent five-fold subject-disjoint evaluation on MotionSense, with models independently trained from scratch within this dataset, yields an Accuracy of 95.01 ± 0.16% and a Macro-F1 of 93.42 ± 0.22%, representing a 1.04-percentage-point improvement over TF-Late. This result demonstrates reproducibility under a second dataset-specific unseen-subject evaluation setting, rather than direct cross-dataset transfer. Representation analyses further indicate enhanced activity separability and reduced activity-conditioned subject clustering. These findings support the reproducibility of activity-conditioned cross-subject relationship learning across two controlled acquisition configurations, without implying global subject invariance or universal robustness across arbitrary devices and environments.
Authors
- Zhigang Tan
- Jiaxin Zhang (ORCID: https://orcid.org/0009-0000-1164-2269)
- Luping Tang
- Zhiming Liu
Institutions
- Hunan Normal University (CN)
- University of South China (CN)
Publication Details
- Journal
- Applied Sciences
- Published
- 2026-11-19
- DOI
- https://doi.org/10.3390/app162211459
- Primary Topic
- Context-Aware Activity Recognition Systems
- Type
- article
- Field-Weighted Citation Impact
- 0.00