Deep Adversarial Learning with Activity-Based User Discrimination Task for Human Activity Recognition

TL;DR

The paper tackles inter-person variability in HAR from inertial sensors by introducing an activity-based adversarial discrimination task that integrates activity information into the discriminator training. It builds a four-component framework (feature extractor $F$, reconstructor $R$, activity classifier $C$, and discriminator $D$) and uses a specially constructed dataset $A'$ of activity-consistent pairs to learn a shared latent space that generalizes across users. Evaluated with Leave-One-(Person)-Out Cross-Validation on PAMAP2, MHEALTH, and REALDISP, the method achieves state-of-the-art accuracy and F1 scores (e.g., up to $0.9710$ accuracy on REALDISP) and shows reduced inter-subject variability compared to prior discriminators. The results demonstrate improved robustness and privacy-preserving representations, with future work extending to larger datasets and cross-dataset validation to further establish generalization.

Abstract

We present a new adversarial deep learning framework for the problem of human activity recognition (HAR) using inertial sensors worn by people. Our framework incorporates a novel adversarial activity-based discrimination task that addresses inter-person variability-i.e., the fact that different people perform the same activity in different ways. Overall, our proposed framework outperforms previous approaches on three HAR datasets using a leave-one-(person)-out cross-validation (LOOCV) benchmark. Additional results demonstrate that our discrimination task yields better classification results compared to previous tasks within the same adversarial framework.

Figures (3)

• Figure 1: In our adversarial framework blue-dashed lines indicate the path followed by samples in $A$, green-full lines indicate the path for samples in $A'$, and red-dotted lines indicate the frozen parts during the adversarial learning process. The testing phase assesses the activity recognition performance for new, unseen individuals using only the previously trained feature extractor and classifier. .
• Figure 2: The upper row shows Accuracy while the lower row depicts $F1-Score_{W}$ for datasets PAMAP2 (left), MHEALTH (center), and REALDISP (right). The black dashed lines in the boxes represent the average value.
• Figure 3: $F-Score_W$ for the different discrimination tasks in PAMAP2 (left), MHEALTH (center), and REALDISP (right). The black dashed lines in the boxes represent the average value.