Self-supervised learning for human activity recognition using 700,000 person-days of wearable data

Summary

This Oxford OxWearables group paper demonstrates that self-supervised learning (SSL) applied to the UK Biobank's massive wrist accelerometer corpus — over 100,000 participants, 700,000+ person-days of unlabelled free-living data — produces generalisable HAR models that substantially outperform supervised baselines trained on small labelled datasets. Using a ResNet backbone with three complementary self-supervised pretext tasks (Arrow of Time, Permutation detection, Time-Warp detection), the authors pretrain on UK Biobank and fine-tune on labelled benchmarks including CAPTURE-24, Rowlands, Opportunity, PAMAP2, and others. The key finding is that self-supervised representations generalise across cohorts, living environments, and sensor devices in a way that purely supervised models do not.