Utilization of unlabeled data for smartphone-robot localization

This paper studies an indoor localization for smartphone-based mobile robot. Due to the difficulty for collecting labeled training data in realistic applications, we propose the learning approach using only a small amount of labeled training data. The key aspect is the utilization of unlabeled data, by combining core concepts of pseudolabelling and time-series learning. The experimental result shows that the developed learning algorithm is the most accurate and robust to the varying number of training data, when compared with other state-of-the-art semi-supervised learning methods.