Augmenting Missing Sensor Data for Robust Human Activity Recognition

Human Activity Recognition (HAR) using wearable sensors has become critical in healthcare, sports, and smart environments. However, model robustness is challenged by missing sensor values, class imbalance, and inter-subject variability. We propose a complete HAR pipeline addressing these issues through temporal augmentation, GAN-PCA-based imputation, and a DeepSense architecture combining convolutional and recurrent layers. Evaluations on the Opportunity dataset using K-fold, leave-one-session-out, and leave-one-subject-out crossvalidation show significant improvements over baseline methods. This study demonstrates the effectiveness of hybrid deep learning for real-world HAR.