WaveletLoc: A Lightweight Deep Learning Model for Indoor Positioning Based on Wavelet Transform

Shilin Ruan, Jianga Shang, Zhe Jian

Abstract: Fingerprint-based positioning has become one of the most popular and promising techniques for indoor localization. Although deep learning-based fingerprint positioning has been widely adopted, existing studies have not fully explored the time-frequency joint features of fingerprints. Furthermore, constrained by the limited computational resources of edge devices, it is necessary to strike a balance between positioning accuracy and system overhead. In this study, we propose a lightweight deep learning positioning model, termed WaveletLoc, which integrates Convolutional Neural Networks with Wavelet Transform. Specifically, we introduce a wavelet-transform convolution approach for multi-scale extraction of Received Signal Strength Indicator (RSSI) features. We propose a Wavelet Domain Spatial-Channel Attention mechanism to enhance each component’s features after wavelet decomposition adaptively. In addition, we design a Nonlinear Gated Channel Attention Unit, which improves the model’s nonlinear representation and feature fusion capability by element-wise multiplication and half activation of channels. Experimental results on multiple datasets demonstrate that, compared to other deep learning-based indoor positioning methods, WaveletLoc offers superior positioning performance.

External IDs:dblp:conf/ijcnn/RuanSJ25