MorphCL: Morphological Contrastive Learning for Inertial-based Human Activity Recognition
MorphCL is a self-supervised pretraining method for Human Activity Recognition using inertial data that explicitly models global movement structure. It achieves this by first discovering recurring motion patterns (motifs) in time series, clustering these motifs into morphological groups, and then using a group-based contrastive loss to ensure segments with similar motion patterns are learned together. This approach significantly boosts downstream performance.