[Paper Review] Wireless Localisation in WiFi using Novel Deep Architectures

Peizheng Li, Han Cui+5 more|arXiv (Cornell University)|Oct 16, 2020

Indoor and Outdoor Localization Technologies26 references6 citations

TL;DR

This paper proposes novel deep learning architectures—shallow neural networks (SNN), convolutional neural networks (CNN), and long short-term memory (LSTM) networks—for WiFi-based indoor localisation using raw and handcrafted features from channel state information (CSI). It demonstrates that with three access points, all models achieve ~0.5 m localisation accuracy, while the CNN and LSTM reduce pre-processing time by 6.5 hours and inference time to 0.1 ms per sample, enabling real-time deployment with high generalisation across motion profiles.

ABSTRACT

This paper studies the indoor localisation of WiFi devices based on a commodity chipset and standard channel sounding. First, we present a novel shallow neural network (SNN) in which features are extracted from the channel state information (CSI) corresponding to WiFi subcarriers received on different antennas and used to train the model. The single-layer architecture of this localisation neural network makes it lightweight and easy-to-deploy on devices with stringent constraints on computational resources. We further investigate for localisation the use of deep learning models and design novel architectures for convolutional neural network (CNN) and long-short term memory (LSTM). We extensively evaluate these localisation algorithms for continuous tracking in indoor environments. Experimental results prove that even an SNN model, after a careful handcrafted feature extraction, can achieve accurate localisation. Meanwhile, using a well-organised architecture, the neural network models can be trained directly with raw data from the CSI and localisation features can be automatically extracted to achieve accurate position estimates. We also found that the performance of neural network-based methods are directly affected by the number of anchor access points (APs) regardless of their structure. With three APs, all neural network models proposed in this paper can obtain localisation accuracy of around 0.5 metres. In addition the proposed deep NN architecture reduces the data pre-processing time by 6.5 hours compared with a shallow NN using the data collected in our testbed. In the deployment phase, the inference time is also significantly reduced to 0.1 ms per sample. We also demonstrate the generalisation capability of the proposed method by evaluating models using different target movement characteristics to the ones in which they were trained.

Motivation & Objective

• To enable accurate, real-time indoor localisation using commodity WiFi hardware and standard channel sounding.
• To evaluate the effectiveness of shallow and deep neural networks in extracting localisation features from CSI without relying on manual feature engineering.
• To compare the performance, complexity, and robustness of SNN, CNN, and LSTM models in practical indoor environments.
• To assess model generalisation across varying target movement characteristics and data arrangement robustness.
• To reduce data pre-processing time and inference latency for deployable, low-resource localisation systems.

Proposed method

• A shallow neural network (SNN) is trained on handcrafted features extracted from CSI, including angle-of-arrival (AoA) estimation and subcarrier principal component analysis (PCA).
• A convolutional neural network (CNN) is designed to process raw CSI matrices directly, using 1×1, 1×3, 3×1, and 3×3 convolutional filters followed by max-pooling and fully connected layers.
• An LSTM network is developed to model temporal dependencies in sequential CSI data, processing time-series CSI inputs to improve tracking stability.
• A novel indoor dataset is collected using an optical tracking system for ground-truth positioning with millimetre accuracy.
• Data augmentation and ablation studies are conducted by randomly dropping training data and features to assess model robustness.
• Model training and evaluation are performed across multiple scenarios with varying numbers of access points (APs), including three-AP configurations.

Experimental results

Research questions

• RQ1 Can a lightweight shallow neural network (SNN) achieve accurate WiFi-based localisation using handcrafted CSI features?
• RQ2 Can deep learning models like CNN and LSTM directly utilise raw CSI data to achieve high localisation accuracy without manual feature engineering?
• RQ3 How does the number of anchor access points (APs) affect the localisation accuracy of different neural network architectures?
• RQ4 What is the trade-off between model complexity, training time, and inference latency in deep learning-based localisation systems?
• RQ5 How robust are the proposed models to variations in target movement patterns and data arrangement (e.g., missing data or features)?

Key findings

• With three anchor access points, all proposed models—SNN, CNN, and LSTM—achieve a mean localisation error of approximately 0.5 metres.
• The deeper CNN architecture, with additional 1×1, 1×3, 3×1, and 3×3 convolutional filters, improves accuracy by around 20% (from 0.5663 m to 0.4664 m) compared to the base CNN.
• The use of raw CSI data in CNN and LSTM models reduces data pre-processing time by 6.5 hours compared to the SNN method that requires manual feature extraction.
• Inference time for the deep learning models is reduced to 0.1 ms per sample, enabling real-time deployment on resource-constrained devices.
• The models demonstrate strong generalisation performance when tested on data with different target movement characteristics than those used during training.
• Feature deactivation has a more significant negative impact on model accuracy than data reduction, with 50% feature loss reducing accuracy by ~20%.