[Paper Review] Sensor-Aided Learning for Wi-Fi Positioning with Beacon Channel State Information

Jeongsik Choi|arXiv (Cornell University)|Jul 13, 2020

Indoor and Outdoor Localization Technologies4 citations

TL;DR

This paper proposes a sensor-aided unsupervised learning framework for Wi-Fi positioning that leverages channel state information (CSI) from beacon frames to improve ranging accuracy. By fusing inertial sensor data with Wi-Fi CSI, the method trains a ranging module to align estimated trajectories with sensor-derived paths, achieving higher precision than RSS-only methods, especially in dynamic indoor environments with multipath fading.

ABSTRACT

Because each indoor site has its own radio propagation characteristics, a site survey process is essential to optimize a Wi-Fi ranging strategy for range-based positioning solutions. This paper studies an unsupervised learning technique that autonomously investigates the characteristics of the surrounding environment using sensor data accumulated while users use a positioning application. Using the collected sensor data, the device trajectory can be regenerated, and a Wi-Fi ranging module is trained to make the shape of the estimated trajectory using Wi-Fi similar to that obtained from sensors. In this process, the ranging module learns the way to identify the channel conditions from each Wi-Fi access point (AP) and produce ranging results accordingly. Furthermore, we collect the channel state information (CSI) from beacon frames and evaluate the benefit of using CSI in addition to received signal strength (RSS) measurements. When CSI is available, the ranging module can identify more diverse channel conditions from each AP, and thus more precise positioning results can be achieved. The effectiveness of the proposed learning technique is verified using a real-time positioning application implemented on a PC platform.

Motivation & Objective

• To reduce reliance on manual site surveys and ground-truth data in Wi-Fi-based indoor positioning by enabling autonomous environment characterization.
• To improve ranging accuracy by exploiting fine-grained CSI from beacon frames, which capture multipath and frequency-selective fading better than RSS alone.
• To develop an unsupervised learning approach that uses sensor trajectories as a proxy for ground truth to train Wi-Fi ranging models without labeled coordinates.
• To enable real-time, on-device adaptation of ranging models using CSI and inertial sensor fusion, minimizing human intervention.

Proposed method

• Uses inertial sensor data (accelerometer and gyroscope) from a mobile device to reconstruct a reference trajectory via pedestrian dead reckoning (PDR).
• Employs an extended Kalman filter (EKF) with multiple initializations of the reference direction to estimate device position and orientation, using PDR as the state transition model.
• Trains a Wi-Fi ranging module using unsupervised learning: the module learns to map CSI from beacon frames to range estimates that minimize the error between Wi-Fi-based and sensor-based trajectories.
• Leverages a new CSI tool capable of capturing CSI from legacy OFDM beacon frames (20 MHz bandwidth) on commodity Wi-Fi chips, enabling simultaneous multi-AP CSI collection without AP modifications.
• Uses innovation (residual) minimization across multiple EKF candidates to select the best state estimate, where the innovation is the difference between predicted and measured ranges.
• Integrates CSI and RSS measurements to improve channel condition identification, enabling more robust distance estimation under fading and shadowing.

Experimental results

Research questions

• RQ1 Can unsupervised learning using sensor trajectories effectively train Wi-Fi ranging models without ground-truth coordinates?
• RQ2 To what extent does incorporating CSI from beacon frames improve positioning accuracy compared to RSS-only ranging?
• RQ3 How does the proposed method handle the challenges of limited CSI bandwidth, asynchronous beacon transmission, and low sampling rate in beacon frames?
• RQ4 Can the fusion of CSI and inertial sensor data reduce drift and improve trajectory consistency in real-time indoor positioning?

Key findings

• The proposed unsupervised learning framework significantly improves positioning accuracy by aligning Wi-Fi-based trajectories with sensor-derived trajectories, reducing error without requiring labeled training data.
• Incorporating CSI from beacon frames enables the system to identify multipath components and frequency-selective fading, leading to more stable and accurate ranging than RSS alone.
• The method achieves better performance than RSS-based fingerprinting in environments with high multipath and shadowing, due to CSI's ability to capture fine-grained channel characteristics.
• The use of multiple EKF candidates with innovation-based selection improves robustness to initial state uncertainty and measurement noise.
• The CSI tool enables simultaneous collection of CSI from multiple APs using legacy beacon frames, overcoming prior limitations of HT-only CSI capture.
• The system demonstrates real-time feasibility on a PC platform, validating the approach for practical deployment in commercial devices.