Hierarchical Self Attention Based Autoencoder for Open-Set Human Activity Recognition

Wearable sensor based human activity recognition is a challenging problem due to difficulty in modeling spatial and temporal dependencies of signals. Recognition models closed-set assumption are forced yield members known classes as prediction. However, can encounter an unseen body-worn malfunction or disability the subject performing activities. This be addressed through solution according open-set recognition. Hence, proposed self attention approach combines data hierarchically from different placements across time classify activities it obtains notable performance improvement over state-of-the-art on five publicly available datasets. The decoder this autoencoder architecture incorporates self-attention feature representations encoder detect setting. Furthermore, maps generated by hierarchical model demonstrate explainable selection features We conduct extensive leave one out validation experiments that indicate significantly improved robustness noise specific variability source code at: github.com/saif-mahmud/hierarchical-attention-HAR.