Cross-Domain HAR · Megha Thukral

The problem. Inertial measurement units (IMUs) in smartphones and smartwatches make it easy to capture human movement, but for specific HAR applications the cost and difficulty of ground-truth annotation severely limits dataset scale and diversity. Transfer learning from publicly available labeled datasets can help, yet these approaches fail when the gap between source and target conditions is large or when only a few samples from the target domain are available, both typical in real-world HAR.

Cross-Domain HAR teacher-student self-training framework — Cross-Domain HAR: a teacher-student self-training framework that bridges large source-to-target gaps in users, sensor placements, and activity vocabularies, using only a few seconds of labeled target data per activity class.

Method

We introduce Cross-Domain HAR, a transfer-learning framework that follows the teacher-student self-training paradigm to recognize activities with very limited target labels. It explicitly bridges conceptual gaps between source and target domains, including different sensor locations on the body and different activity vocabularies. The workflow is:

Train teacher on the labeled source domain.
Generate pseudo-labels on the unlabeled target data using the teacher.
Filter top-K confident predictions per class to suppress noisy pseudo-labels.
Pre-train student in a multi-task setting on the augmented, filtered data, with self-supervised loss on unlabeled target streams and consistency regularization (matching teacher predictions on weakly-augmented inputs to student predictions on strongly-augmented versions of the same signal).
Fine-tune the student with the few labeled target examples.

We use eight different signal augmentations to artificially expand much smaller activity-recognition datasets and to drive the consistency regularization. The combination of teacher-student self-training, confidence filtering, multi-task self-supervision on target data, and strong/weak augmentation consistency is what allows transfer to succeed even when source and target cover disjoint activities and different sensor placements.

Why it matters

Most real-world HAR deployments involve target domains where only a few seconds of labeled sensor data per activity are realistic to collect, with sensor placements and activity sets that differ from existing public datasets (e.g. healthcare-specific movements on novel body locations). Cross-Domain HAR enables substantial performance improvements over the state-of-the-art in such few-shot, large-gap settings without requiring expensive new annotation campaigns, helping bootstrap HAR models quickly for new applications.

Results

Through extensive evaluation across a range of benchmark HAR datasets, Cross-Domain HAR delivers significant accuracy improvements over state-of-the-art transfer-learning baselines in practically relevant few-shot scenarios. We also conduct detailed component analyses to identify which parts of the framework drive successful transfer, and provide practical suggestions for applying the framework to real-world HAR applications.

Cite

@article{thukral2025cross,
  title  = {Cross-Domain HAR: Few-Shot Transfer Learning for Human Activity Recognition},
  author = {Thukral, Megha and Haresamudram, Harish and Pl{\"o}tz, Thomas},
  journal= {ACM Transactions on Intelligent Systems and Technology},
  volume = {16}, number = {1}, pages = {1--35},
  year   = {2025},
  doi    = {10.1145/3704921}
}