Multi-Granularity and Multi-Modal Feature Fusion for Indoor Positioning

Lijuan Ye, Yi Wang, Shenglei Pei, Yu Wang, Hong Zhao, Shi DongSymmetry 2025 · DOI

TL;DR

This paper introduces a novel indoor positioning system that fuses Wi-Fi Channel State Information (CSI) with visual data from monocular cameras. By employing a coarse-to-fine-grained approach, it overcomes the limitations of individual methods, achieving centimeter-level accuracy without requiring specialized hardware.

What Problem Does This Paper Address?

Accurate indoor positioning is crucial for applications like navigation in large public spaces and efficient management in industrial settings. While GPS is effective outdoors, it fails indoors due to signal obstruction. Existing indoor solutions face challenges: Wi-Fi-based methods struggle with accuracy and architectural complexity, while vision-based methods often require expensive stereo cameras for depth information. Channel State Information (CSI) from Wi-Fi can be susceptible to periodic errors due to multipath interference, and image-based methods can suffer from spatial ambiguity in repetitive textures. This paper aims to bridge this gap by developing a robust, accurate, and accessible indoor positioning system using readily available smartphone technology.

Methods

The proposed method employs a two-stage approach:

  • Coarse-Grained Positioning: A lightweight neural network, MobileViT, is used to process image data and extract initial location information. This stage leverages the geometric relationships within images to provide a rough localization.
  • Fine-Grained Positioning: This stage fuses CSI features from Wi-Fi signals with Scale-Invariant Feature Transform (SIFT) features extracted from images. This fusion creates a rich feature fingerprint library. A Support Vector Machine (SVM) regression model is then trained on this fused library to achieve high-precision localization.

Key novelties include:

  • Multi-Granularity: The system uses both coarse and fine-grained approaches for positioning.
  • Multi-Modal Fusion: It synergistically combines Wi-Fi CSI and visual data, using each to compensate for the weaknesses of the other. Visual data helps mitigate CSI errors, while CSI's global stability corrects image-based local ambiguities.
  • SIFT for Image Features: SIFT is used for its robustness to scale, rotation, and illumination changes, making it suitable for diverse indoor environments.
  • CSI Amplitude: Only the amplitude of CSI subcarriers is used, as phase information is more susceptible to noise and frequency offsets.

Datasets

The paper describes a custom dataset collected for training and evaluation. This dataset consists of:

  • 100 unique location points within a designated indoor area.
  • At each location point, 8 images were captured, representing different orientations.
  • A total of 8800 high-quality images were collected.
  • Wi-Fi data (CSI) was also collected at these points.

The paper does not mention using publicly available indoor positioning datasets.

Key Insights and Ideas

  • Synergistic Fusion: The core insight is that combining Wi-Fi CSI and visual data in a multi-modal, multi-granularity framework significantly outperforms individual methods. The strengths of each modality effectively address the weaknesses of the other.
  • Overcoming Depth Limitations: By fusing CSI with monocular image features, the system bypasses the need for expensive stereo cameras or depth sensors, making it accessible via standard smartphones.
  • Robustness to Indoor Challenges: The approach is designed to handle common indoor issues like texture repetition (addressed by CSI) and multipath interference (addressed by visual cues).
  • Scalability and Portability: The system's design allows for easy updates to the fingerprint database and the creation of new databases for different indoor environments, enhancing its practical applicability.

Key Results

The experimental results demonstrate the effectiveness of the proposed method:

  • High Accuracy: The system achieved an accuracy of 0.4 m for 45% of positioning points and 0.8 m for 67% of points.
  • Comparison with Wi-Fi Only: The proposed method achieved a 50% localization error within 1 m, significantly better than a Wi-Fi-only approach which had a 50% error within 2 m.
  • Impact of Sample Size: Increasing the sample size for training improved accuracy, with 50% of errors narrowing to 0.4 m with 20,000 samples.
  • Antenna Configuration: Using multiple antennas (specifically three) significantly improved localization accuracy, reducing the 50% error to 0.4 m.
  • Comparison with Popular Algorithms: The proposed method outperformed existing algorithms like DFPhaseFL, MSDFL, SCNN, SVM, and KNN, achieving a 43% probability of error within 0.4 m and a 64% probability within 0.8 m, with the smallest average error of 0.67 m.

Limitations and Future Work

Limitations: * Dataset Specificity: The custom dataset, while comprehensive, might not fully represent the diversity of all indoor environments. * Computational Cost: While efforts are made to optimize, complex feature extraction and fusion can still be computationally intensive.

Future Work: * Deep Learning Enhancement: Further exploration of deep learning techniques for indoor positioning to improve accuracy and reliability. * Dynamic Fusion Strategies: Developing more sophisticated methods for dynamic fusion of multi-source information. * User Privacy: Prioritizing user privacy protection in the development and utilization of indoor positioning technology. * Real-time Updates: Enabling users to upload location data for real-time database updates and create new fingerprint databases for new environments.

Summary updated

BibTeX
@article{ye2025multi,
  author = {Lijuan Ye and Yi Wang and Shenglei Pei and Yu Wang and Hong Zhao and Shi Dong},
  title = {{Multi-Granularity and Multi-Modal Feature Fusion for Indoor Positioning}},
  journal = {Symmetry},
  year = {2025},
  doi = {10.3390/sym17040597},
}
Figure