UCT: Uncertainty-Based Consistency Training for Domain Adaptive Human Activity Recognition

Abstract

In the filed of mobile sensing and applied computing, sensor-based human activity recognition (HAR) constitutes a critical element. Labeling large datasets is typically resource-intensive and costly. Moreover, HAR models trained on one user show sub-optimal performance when generalized to other users due to data heterogeneity. This heterogeneity arises from variations in daily routines and environmental conditions among different data-gathering subjects. To achieve precise HAR with lower labeling costs, We propose a new approach to achieve domain adaptation using uncertainty-based consistency training (UCT). Aligning source and target domains is a significant challenge for traditional domain adaptation methods, due to the influence of domain-specific features, which hinders the extraction of domain-invariant features and leads to sub-optimal adaptation performance. In summary, we propose an uncertainty-based consistency training method. First, source and target domains are mapped to their respective intermediate domains through consistency training. Since the intermediate domain shares feature representations and contains minimal domain-specific features, this mapping reduces the influence of domain-specific features while also narrowing the discrepancy between the two domains. Subsequently, the discrepancies between source and target domains are mitigated through aligning them with an intermediate domain, thereby achieving more precise feature alignment. Additionally, to enable the model to learn multi-modal features from multiple sensors, we propose a multi-scale feature Attention network (MSFAN). Finally, we demonstrate that the domain adaptation performance of UCT significantly outperforms other methods by conducting extensive experiments on three public HAR datasets.