Enhancing Human Activity Recognition With Smartphone Sensors: A Deep Semi-Supervised Approach

Human activity recognition (HAR) has emerged as a crucial field with diverse applications in areas such as healthcare, sports, and security. The majority of available HAR datasets have collected noise-free data, which do not represent real-world applications due to limitations in the data collection phase. Deep learning is revolutionizing several fields, including HAR. However, the scarcity of labeled data poses a considerable challenge in training deep HAR models. Semi-supervised learning (SSL) strategies effectively utilize both labeled and unlabeled data to enhance performance. These approaches improve the accuracy of supervised learning (SL) models that rely solely on labeled data by leveraging unlabeled data. In this article, we present a new dataset collected from real-world situations, consisting of both labeled and unlabeled samples. We propose a novel semi-supervised approach for HAR based on the MixUp augmentation and adversarial noise. Our approach outperforms SL by 8.89% in accuracy by exploiting latent patterns in unlabeled data. The results indicate that the proposed deep semi-supervised learning (DSSL) approach effectively addresses the challenges posed by noisy real-world datasets.