Deep Smartphone Sensors-WiFi Fusion for Indoor Positioning and Tracking

Leonid Antsfeld,Boris Chidlovskii,Emilio Sansano-Sansano

DOI: https://doi.org/10.48550/arXiv.2011.10799

2020-11-21

Abstract:We address the indoor localization problem, where the goal is to predict user's trajectory from the data collected by their smartphone, using inertial sensors such as accelerometer, gyroscope and magnetometer, as well as other environment and network sensors such as barometer and WiFi. Our system implements a deep learning based pedestrian dead reckoning (deep PDR) model that provides a high-rate estimation of the relative position of the user. Using Kalman Filter, we correct the PDR's drift using WiFi that provides a prediction of the user's absolute position each time a WiFi scan is received. Finally, we adjust Kalman Filter results with a map-free projection method that takes into account the physical constraints of the environment (corridors, doors, etc.) and projects the prediction on the possible walkable paths. We test our pipeline on IPIN'19 Indoor Localization challenge dataset and demonstrate that it improves the winner's results by 20\% using the challenge evaluation protocol.

Machine Learning,Networking and Internet Architecture,Robotics

What problem does this paper attempt to address?

The problem that this paper attempts to solve is the indoor positioning problem, and the specific goal is to predict the user's trajectory from the data collected by smart phones. The paper proposes a new sensor - fusion framework, which utilizes the inertial sensors of smart phones, Wi - Fi measurement data and landmarks to achieve accurate indoor positioning and tracking. This framework mainly includes the following aspects:

1. **Deep Pedestrian Dead - Reckoning (Deep PDR)**:
- Use deep - learning methods to process the sensor data stream, and extract the hidden associations between different sensors through Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN) to learn the model of user local displacement.
- This method can deal with sensor noise and replace manual feature extraction. The latter is usually vulnerable to data noise and requires complex threshold adjustments for different pedestrian characteristics.

2. **Landmarks and Pseudo - Labels**:
- Since the truly labeled data is scarce and limited to a few landmarks, while the original sensor data is generated in large quantities at a high rate, a method has been developed to label the sensor data using pseudo - labels to generate a large - scale training set.
- Generate pseudo - labels based on simple tasks such as user activity recognition and landmark detection, as well as speculation on user behavior between landmarks.

3. **Semi - supervised Variational Auto - Encoder (VAE) for Wi - Fi Positioning**:
- Use a semi - supervised learning method to construct a radiomap from Wi - Fi data, and can infer the approximate position of the Wi - Fi fingerprint even without accurate position labels.
- This method combines unlabeled and labeled sensor data and improves the accuracy of positioning.

4. **Kalman Filter and Map - Independent Projection Algorithm**:
- Use the Kalman filter to fuse the different - frequency predictions of the Deep PDR and Wi - Fi components to provide an estimate of the user's position.
- Finally, use the map - independent projection algorithm to project the results of the Kalman filter onto the paths allowed by the environmental physical constraints, avoiding the prediction results from passing through impossible areas (such as walls).

The paper demonstrates the superiority of this framework over existing methods through tests on the IPIN'19 Indoor Positioning Challenge dataset, with a 20% reduction in positioning error.