Toward Personalized Human Activity Recognition Model with Auto-Supervised Learning Framework
Ala Mhalla, Jean-Marie Favreau · 2021
The performance of any Human Activity Recognition (HAR) model trained from a general population (subject-independent), will often decreases when tested on subjects unseen during training. This is due to inter-subject variability such as differences in activity patterns, emotional status, gait or posture between users. To solve this problem, we propose a novel auto-supervision formalism based on the theory of a Particle Filter (PF) or Sequential Monte Carlo (SMC) method to automatically build a personalized HAR model. The suggested approach uses different steps based on the SMC filter method to automatically and iteratively approximate the target distribution as a set of temporal samples in order to personalize the HAR model towards a target user. These temporal samples are selected from the target user based on their importance score, which indicates the likelihood of the temporal samples belong to the target distribution. We validated our framework on several public HAR datasets, and prediction accuracy was consistently improved for new subjects in the test set.