DeepLocNet: Deep Observation Classification and Ranging Bias Regression for Radio Positioning Systems

Maani Ghaffari, Sahib Singh Dhanjal, Maani Ghaffari Jadidi, Ryan M. Eustice

TL;DR

This paper introduces DeepLocNet, a novel deep learning framework for accurate indoor localization using only commodity WiFi signals. It tackles the challenge of distinguishing Line-of-Sight (LOS) from Non-Line-of-Sight (NLOS) propagation, achieving 85% classification accuracy and significantly improving localization performance in both simulated and real-world environments.

What Problem Does This Paper Address?

Accurate indoor localization is crucial for many applications, from augmented reality to robotics. While WiFi is ubiquitous and low-cost, its signals are notoriously unreliable for precise positioning due to multipath effects and obstructions, leading to Line-of-Sight (LOS) and Non-Line-of-Sight (NLOS) propagation. Existing methods often require specialized hardware or extensive pre-mapping (fingerprinting), limiting their practicality for everyday mobile devices. This paper aims to bridge this gap by developing a system that leverages readily available WiFi Received Signal Strength Indicator (RSSI) data, combined with a deep learning approach, to reliably classify LOS/NLOS signals and improve localization accuracy without custom hardware or tedious site surveys.

Methods

The core of the DeepLocNet framework is a deep neural network (DNN) designed to classify WiFi signals as either LOS or NLOS. This network, inspired by autoencoders, learns to extract relevant features from RSSI data. The authors experimented with various architectures, settling on a "light-weight" design that can run on embedded devices.

This DNN is then integrated into a particle filtering framework for localization. Specifically, it's used within the measurement model of the particle filter. The DNN's LOS/NLOS classification output is used to adjust the ranging bias, making the measurements more reliable. For scenarios where access point (AP) locations are not precisely known, the framework utilizes FastSLAM, a particle filter variant that simultaneously estimates the robot's pose and the map of AP locations. The paper explores three measurement model variations: No Classification (NC), Hard Classification (HC), and Soft Classification (SC), with SC leveraging the probabilities from the DNN.

Datasets

The paper describes two primary data sources:

  • Simulated Data: Over 10 million data points were generated using the Pylayers simulator. This tool allowed for the creation of diverse indoor environments with varying parameters (temperature, humidity, AP locations, floor plans) to simulate wave propagation.
  • Real-World Data: Data was collected in a building on the University of Michigan campus using a Fetch Mobile Manipulator. This involved manually moving the robot, recording WiFi signatures, and using visual odometry (ORB-SLAM) for ground truth.

Key Insights and Ideas

  • Commodity WiFi is Sufficient: The research demonstrates that with the right approach, accurate indoor localization can be achieved using only standard WiFi RSSI signals, eliminating the need for specialized hardware.
  • Deep Learning for LOS/NLOS is Powerful: A deep neural network can effectively learn to distinguish between LOS and NLOS propagation, a critical step for improving the reliability of RSSI-based ranging.
  • Soft Classification Improves Robustness: Using the probabilities from the DNN (soft classification) in the measurement model leads to more robust localization compared to hard classification or no classification at all.
  • FastSLAM Handles Unknown AP Locations: The integration of the DeepLocNet classifier with FastSLAM allows for effective localization even when the exact positions of WiFi access points are not perfectly known, by simultaneously estimating both the robot's path and the AP map.
  • System is Robust to Map Errors: The proposed framework shows resilience to errors in the initial map of access points, with localization accuracy maintained even with up to 0.8m of error in AP locations.

Key Results

  • LOS/NLOS Classification Accuracy: The DeepLocNet classifier achieves approximately 85% accuracy in distinguishing between LOS and NLOS signal propagation.
  • Localization Accuracy (Simulation): In simulated 2D and 3D office environments, the DeepLocNet framework (using FastSLAM with soft classification) achieved a mean Root Mean Square Error (RMSE) of 1.6271m (3D) and 0.5258m (2D), significantly outperforming methods without classification or with hard classification.
  • Localization Accuracy (Hardware): In real-world hardware experiments, the FastSLAM with soft classification achieved an RMSE of 1.6271m in a 2D environment.
  • Improved Performance: The paper shows substantial RMSE improvements for localization using both Particle Filter and FastSLAM when employing the DeepLocNet classifier, particularly with soft classification.

Limitations and Future Work

  • PHY Layer Information: The authors acknowledge that incorporating Physical Layer (PHY) information (beyond RSSI) could further enhance localization accuracy. However, this would require mobile devices to be equipped with the necessary hardware.
  • Hardware Requirements for PHY: The reliance on custom hardware for PHY layer information is a significant barrier to widespread adoption for ubiquitous mobile devices.
  • Future Work: The paper suggests exploring the integration of PHY layer information as a key direction for future research.

Summary updated

Abstract

WiFi technology has been used pervasively in fine-grained indoor localization, gesture recognition, and adaptive communication. Achieving better performance in these tasks generally boils down to differentiating Line-Of-Sight (LOS) from Non-Line-Of-Sight (NLOS) signal propagation reliably which generally requires expensive/specialized hardware due to the complex nature of indoor environments. Hence, the development of low-cost accurate positioning systems that exploit available infrastructure is not entirely solved. In this paper, we develop a framework for indoor localization and tracking of ubiquitous mobile devices such as smartphones using on-board sensors. We present a novel deep LOS/NLOS classifier which uses the Received Signal Strength Indicator (RSSI), and can classify the input signal with an accuracy of 85\%. The proposed algorithm can globally localize and track a smartphone (or robot) with a priori unknown location, and with a semi-accurate prior map (error within 0.8 m) of the WiFi Access Points (AP). Through simultaneously solving for the trajectory and the map of access points, we recover a trajectory of the device and corrected locations for the access points. Experimental evaluations of the framework show that localization accuracy is increased by using the trained deep network; furthermore, the system becomes robust to any error in the map of APs.
BibTeX
@inproceedings{ghaffari2019deeplocnet,
  author = {Maani Ghaffari and Sahib Singh Dhanjal and Maani Ghaffari Jadidi and Ryan M. Eustice},
  title = {{DeepLocNet: Deep Observation Classification and Ranging Bias Regression for Radio Positioning Systems}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2019},
  doi = {10.1109/iros40897.2019.8967767},
}
Figure