FT-Loc: A Fine-Grained Temporal Features-Based Fusion Network for Indoor Localization

Indoor location-based services (LBSs) are critical for enhancing social and commercial activities that require accurate and efficient localization techniques. Existing deep-learning-based indoor localization methods mainly focus on predefined global features to learn local discriminative representations, which increases learning difficulty and is not efficient or robust to scenarios with small variations. To address the above issues, we propose a novel fine-grained temporal features-based localization (FT-Loc) framework that utilizes multiple subsignal features to provide accurate location estimation, and each subsignal represents a piece of clue for a specific position. Specifically, the proposed framework takes multiple local signal sequences as input, and deep networks considering temporal correlations are designed for extracting features from the corresponding location clues, respectively. Then, a lightweight attention generation scheme is used to learn the importance of each temporal representation. Guided by the obtained attention values, we fuse multiple local features to generate more distinguishing ones for accurate localization. The experimental results show that FT-Loc significantly outperforms existing localization schemes with accuracy improvements of at least 43.36%.