BoundLocNet: Object Detection-Driven Indoor Localization Using CSI Fingerprint

With the development of the Internet of Things (IoT) and the popularity of wireless fidelity (WiFi), WiFi-based indoor fingerprint localization has raised increasing attention. However, in dynamic multipath environments, traditional regression-based deep learning (DL) localization models not only require extensive training data but also struggle to cope with signal distortions induced by obstacles and human motion. To address these limitations, we propose BoundLocNet, an object detection–driven framework for device localization using channel state information (CSI) fingerprints. By redefining spatial coordinates as bounding boxes and employing an improved training scheme, the potential correlation between CSI fingerprints and the physical region of a device is captured. Then, the region proposal network (RPN) iteratively refines the bounding box predictions to mitigate environmental interference. Experimental results show that BoundLocNet improves the mean distance error (MDE) by 31% compared to baselines, with enhanced robustness to environmental changes.