A Systematic Study of Unsupervised Domain Adaptation for Robust Human-Activity Recognition
summary
Wearable IMU HAR models suffer up to 90% accuracy drop when sensor placement changes. This paper evaluates three unsupervised domain adaptation methods on four HAR datasets: Feature Matching (MMD), Confusion Maximization (domain adversarial), and Data Augmentation. Feature Matching gives the best target-domain gains, Confusion Maximization often hurts both source and target, and Data Augmentation improves adaptability but reduces persistence; performance critically depends on dataset size and class-distribution mismatch.
AI-written summary · about
related talks
06mSelfHAR: Improving Human Activity Recognition through Self-training with Unlabeled Data
06mContrastive Predictive Coding for Human Activity Recognition
06mUnsupervised Human Activity Representation Learning with Multi-task Deep Clustering
06mDeriving Effective Human Activity Recognition Systems through Objective Task Complexity Assessment
11mDetection and Differentiation of Obstacles in Repeated Adaptive Human-Computer Interactions from ...
05mLeveraging Activity Recognition to Enable Protective Behavior Detection in Continuous Data
06mApproaching the Real-World: Supporting Activity Recognition Training with Virtual IMU Data
09mAn Activity Recognition System for Taking Medicine Using In-The-Wild Data to Promote Medication ...