Geometric Sound Profile: Multipath-Time-of-Flight Fingerprint for High-Accuracy Acoustic Localization

High-accuracy indoor positioning systems provide various location-aware applications, enhancing our daily experiences. The Global Navigation Satellite System is difficult to use indoors, leading to the development of various indoor positioning methods. Among these, acoustic fingerprint-based positioning, which utilizes widely available speakers and microphones, provides robust performance in Non-Line-of-Sight (NLOS) environments that are common due to structural obstacles and furniture. These approaches typically use Received Signal Strength Indicator or Power Spectral Density as location fingerprints but face challenges in achieving high positioning accuracy. In this paper, we propose Geometric Sound Profile (GSP), a temporal feature of complex reflection waves, enabling high-accuracy localization with a single speaker. GSP, defined as the envelope of cross-correlation between the transmitted and received signals, starting from the transmission time, serves as a highly informative feature encapsulating the multipath-Time-of-Flight. Additionally, we implement a Convolutional Neural Network for the estimation of the user’s position using GSP. We generated a high-resolution pre-trained model in the simulation and fine-tuned it with measurement data, allowing accurate positioning with minimal measured training data. Our experiments demonstrated that the median positioning error was 0.14 m and the 90th percentile error was 1.23 m in the Line-of-Sight environment, and the median positioning error was 0.09 m and the 90th percentile error was 1.14 m in the NLOS environment.