An Audio Fingerprinting Based Indoor Localization System: From Audio to Image

Audio signal has emerged as a localization source in the popular indoor fingerprinting positioning (FPP) field in recent years, thanks to its rich time-frequency domain features. However, the numerical role of signals is usually considered and the temporal context of the relative time-of-arrival (ToA) of signals is ignored, which leads to a low device and non-line-of-sight (NLoS) adaptability for consumer applications. In this work, a chirp signals group is designed to strengthen the signal spatial correlation. The energy density map (EDM) of the received signal is generated and analyzed as a whole. First, the signal response domain range is narrowed by a global normalization to reduce the device differences. Second, a simple yet effective audio-chirp-attention network (ACANet) is constructed to fuse the edge detection map with a normalized EDM, strengthening the attention to the relative ToA and associating the fingerprint dataset with the corresponding spatial position for real-time coordinates estimations. Last, evaluations of three smartphones are carried out over three typical scenes under both soft- and hard-NLoS. These evaluations show that the proposed system achieves an average localization root-mean-square error (RMSE) of 1.69 m in the scenes of 1–3 anchor NLoS with different smartphones and motion postures. The ACANet trained with the normalized EDM data is proved to reduce the device difference by 43.7% when using the same model, as well as decrease the absolute error by 63.7% compared to using the raw EDM data.