A Dual-Modality CNN Approach for RSS-Based Indoor Positioning Using Spatial and Frequency Fingerprints.

Area of Science:

• Indoor positioning systems
• Wireless sensor networks
• Signal processing

Background:

• Received signal strength (RSS) based indoor positioning relies on spatial fingerprint features.
• Low-resolution fingerprint images limit the extraction of subtle features, impacting accuracy and robustness.
• Existing methods struggle with effectively utilizing spatial information for precise localization.

Purpose of the Study:

• To propose an improved RSS-based indoor positioning method using enhanced spatial and frequency domain fingerprint representations.
• To leverage fusion learning with a convolutional neural network (CNN) for more accurate 3D coordinate prediction.
• To address the limitations of low-resolution spatial RSS fingerprint images.

Main Methods:

• Applied bicubic interpolation to enhance spatial resolution of RSS fingerprint images.
• Utilized 2D Fast Fourier Transform (2D FFT) to convert spatial images into frequency domain representations.
• Developed a parallel CNN with efficient multi-scale attention (EMA) modules for dual-modality (spatial and frequency) feature extraction and fusion.
• Implemented a coordinate optimization strategy for final 3D coordinate estimation.

Main Results:

• The proposed method demonstrated significant improvements in positioning accuracy across seven public datasets.
• Achieved a reduction in mean positioning error by up to 47.1%.
• Reduced root mean square error (RMSE) by up to 54.4% compared to traditional and advanced methods.

Conclusions:

• The combination of enhanced spatial-frequency fingerprint representation and fusion learning effectively extracts richer features for indoor positioning.
• The proposed method offers a substantial advancement in RSS-based indoor positioning accuracy and robustness.
• This approach provides a promising solution for overcoming the challenges posed by low-resolution spatial RSS fingerprint images.