Speaker
Description
This work investigates path signatures as a principled representation for short-window inertial Human Activity Recognition (HAR) on the UCI-HAR benchmark. Under a leakage-aware protocol (subject-disjoint validation, train-only standardisation) and a fixed MLP classifier, we compare a raw-window baseline to signature features computed using time augmentation and lead–lag embedding, truncated at depths m = 2, 3, 4. Results show that performance generally peaks at depth m = 3, with lead–lag (m = 3) being the strongest signature variant, but raw flattened windows remain best overall. Error analysis suggests signatures capture locomotion well but struggle to separate static postures, highlighting key trade-offs between embedding choice, truncation depth, and computational cost.