Local and Global Alignments for Generalizable Sensor-Based Human Activity Recognition

Sensor-based human activity recognition (HAR) plays an important role in our daily life. Most work on HAR often assumes that training and test samples follow the same data distribution, which is not realistic in practice. For example, activity patterns usually vary from person to person, which will hinder the generalization ability of the model. In this paper, we propose Local And Global alignment (LAG) for generalized sensor-based HAR. Our method is able to alleviate distribution shifts among training and test samples without touching test data. Specially, the proposed method learns domain-invariant features from both the local and global perspectives and utilizes combined features to classify. Comprehensive experimental evaluations are conducted on two benchmarks to demonstrate the superiority of the proposed method over state-of-the-art approaches.