Accurate and Efficient Human Activity Recognition Through Semi-Supervised Deep Learning

In recent years, deep learning has been widely used for human activity recognition (HAR) based on wearable sensor data due to its excellent performance. However, deep learning approaches often need a large amount of labeled data to train the model. The collection of huge amounts of labeled activity data is usually labor-intensive and even cost-prohibitive, limiting the practical use of these methods. An effective way to address this issue is to use semi-supervised learning methods. Yet, most existing semi-supervised approaches fail to fully make use of unlabeled data to improve deep learning models’ accuracy and generalizability. In this work, we propose an innovative Adaptive Confidence Semi-supervised HAR (ACS-HAR) method that can dynamically adjust the selective confidence threshold based on the model’s learning state, thereby making more efficient use of unlabeled data. Experimental results on six public datasets demonstrate that our method outperforms traditional supervised methods by up to 5.20% accuracy and state-of-the-art semi-supervised methods by up to 3.30% accuracy.We release our code at https://github.com/objwww/ACS–HAR.