A Wavelet Scattering Feature Extraction Approach for Deep Neural Network Based Indoor Fingerprinting Localization

Bedionita Soro, Chaewoo Lee

TL;DR

This paper introduces a novel feature extraction method using wavelet scattering transforms to improve the accuracy of indoor localization systems that rely on Wi-Fi Received Signal Strength Indicator (RSSI) fingerprinting. By making features more robust to signal fluctuations and device orientation, their approach significantly reduces localization errors compared to existing methods.

What Problem Does This Paper Address?

Indoor localization is crucial for many smart city applications, but GPS signals are unreliable indoors. Wi-Fi fingerprinting, which uses RSSI from access points to estimate location, is a popular alternative. However, RSSI is notoriously unstable due to environmental factors, multipath propagation, and device orientation, which severely degrades the performance of even advanced machine learning models like Deep Neural Networks (DNNs). Existing methods struggle to extract features that are stable enough to overcome this RSSI variability, leading to inaccurate location estimates.

Methods

The core innovation of this paper lies in its feature extraction technique: the wavelet scattering transform. This method is applied to the raw, one-dimensional RSSI time-series data. The process involves:

  • Convolution with Wavelets: The input RSSI signal is convolved with a bank of dilated wavelets at different scales and orientations.
  • Nonlinearity (Modulus): The absolute value (modulus) of the convolution results is taken. This step is crucial for making the features invariant to small signal variations.
  • Averaging (Low-pass Filter): The modulus is then convolved with a low-pass filter to produce a smoothed representation.

This process is cascaded across multiple layers, creating a hierarchical representation of the signal. The authors specifically utilize the zeroth-order scattering coefficients (representing smoothed signal energy) and sometimes the first-order coefficients (capturing more detailed signal variations) as features. These extracted features, which are inherently stable to small deformations and invariant to translation and rotation, are then fed into a Deep Neural Network (DNN) for classification. The DNN, with four hidden layers and a SoftMax output, predicts the probability of the signal belonging to each Reference Point (RP). Finally, a KNN-based strategy using the top predicted RPs is employed to estimate the final location.

Datasets

The paper evaluates its proposed method on two types of datasets:

  • Local Corridor Experiment: A custom dataset collected in a 50m x 1.95m corridor at Ajou University. It involved 21 Reference Points (RPs) with 100 samples per RP, collected using a Samsung Galaxy S8+. Only the six strongest APs were used, resulting in 36 APs in total.
  • Publicly Available Dataset: The dataset from Mendoza-Silva et al. (2018), which contains 25 subsets collected over several months, designed to test robustness to RSSI variation. This dataset includes training and test sets with varying degrees of overlap in collected locations.

Key Insights and Ideas

  • Wavelet Scattering as a Robust Feature Extractor: The fundamental insight is that wavelet scattering transforms can effectively capture essential signal characteristics while discarding noise and variations that plague RSSI data. Its inherent properties of multiscale analysis, linearization, and sparse representation make it ideal for this task.
  • Invariance to Signal Perturbations: The key advantage of wavelet scattering is its ability to produce features that are invariant to small deformations, translations, and rotations. This directly addresses the problem of RSSI instability caused by device orientation and minor environmental changes.
  • Zeroth-Order Coefficients are Sufficient (Sometimes): For certain datasets (like the corridor experiment), the zeroth-order scattering coefficients alone provided excellent results, indicating that a significant portion of the discriminative information for localization is captured in the smoothed signal energy.
  • DNNs Benefit Greatly from Good Features: The paper reinforces the idea that the performance of a DNN is highly dependent on the quality of the input features. By providing stable, discriminative features, the DNN can achieve much higher accuracy.

Key Results

The proposed wavelet scattering-based feature extraction method, combined with a DNN, demonstrated significant improvements in localization accuracy:

  • Corridor Experiment: Achieved a mean positioning error of 0.68 meters, a substantial reduction compared to SAE+DNN (1.27m) and MNN (1.58m), and a reported 46% reduction in error compared to prior work [21].
  • Publicly Available Dataset (Option 1): Achieved 100.0% floor rate and a mean positioning error of 0.0 meters when test points perfectly coincided with training points. Even with variations, errors were significantly reduced.
  • Publicly Available Dataset (Option 2): Achieved 99.6% floor rate and a mean positioning error of 4.15 meters, outperforming SAE+DNN (6-10m) and showing competitive results with KNN (3.17m).
  • Publicly Available Dataset (Option 3): Achieved a minimum error of 4.27 meters, outperforming SAE+DNN (5.43m).

Limitations and Future Work

  • Overfitting Concerns: The authors suggest that a higher positioning error in one specific scenario (second scenario of the second experiment) might be due to overfitting.
  • Choice of Scattering Coefficients: The optimal choice of scattering coefficients (zeroth vs. first order) was found to be dataset-dependent, requiring experimental validation.
  • Future Work Suggestions:
    • Investigate other machine learning classifiers with wavelet scattering features, such as Support Vector Machines (SVMs).
  • Explore the exploitation of time-frequency information, Shannon entropy, and autoregressive features.
  • Further investigate other functionalities of the wavelet framework.

Summary updated

Abstract

The performance of an Artificial Neural Network (ANN)-based algorithm is subject to the way the feature data is extracted. This is a common issue when applying the ANN to indoor fingerprinting-based localization where the signal is unstable. To date, there is not adequate feature extraction method that can significantly mitigate the influence of the receiver signal strength indicator (RSSI) variation that degrades the performance of the ANN-based indoor fingerprinting algorithm. In this work, a wavelet scattering transform is used to extract reliable features that are stable to small deformation and rotation invariant. The extracted features are used by a deep neural network (DNN) model to predict the location. The zeroth and the first layer of decomposition coefficients were used as features data by concatenating different scattering path coefficients. The proposed algorithm has been validated on real measurements and has achieved good performance. The experimentation results demonstrate that the proposed feature extraction method is stable to the RSSI variation.
BibTeX
@inproceedings{soro2019wavelet,
  author = {Bedionita Soro and Chaewoo Lee},
  title = {{A Wavelet Scattering Feature Extraction Approach for Deep Neural Network Based Indoor Fingerprinting Localization}},
  booktitle = {Sensors},
  year = {2019},
  url = {https://openreview.net/forum?id=JQrwHaqUBde},
}

Loading related content...

Figure