Application of Machine Learning to GNSS/IMU Integration for High Precision Positioning on Smartphone

Abstract

This paper describes our solution for the Google smartphone decimeter challenge (GSDC), which was held from May to August 2022. The GSDC is a competition for improving positioning accuracy of smartphones. The global navigation satellite system (GNSS) data from smartphones have lower signal levels and higher noise in GNSS observations compared to commercial GNSS receivers. Therefore, it is difficult to directly apply the existing GNSS high-precision positioning methods like precise point positioning (PPP) and real-time kinematic (RTK). The smartphones used to collect the raw GNSS data have multi-constellation, dual-frequency GNSS receivers, and Inertial Measurement Unit (IMU) sensors. Multi-sensor fusion technology has become very prominent for seamless navigation systems due to its complementary capabilities to GNSS positioning. In this work, we developed a machine learning (ML) based adaptive positioning approach to estimate the positions of the smartphone by utilizing post-processed kinematic (PPK) precise positioning techniques to process the GNSS datasets. The ML model is used to predict the driving paths (highways, tree-lined streets, or downtown areas). Depending on the predicted driving path, PPK technique uses the carrier phase to compute the user position using differential corrections from known GNSS base stations. We then use of the Rauch–Tung–Striebel (RTS) smoother, which consists of a forward pass Kalman Filter (KF) and a backward recursion smoother to achieve a loosely coupled integration of GNSS and IMU measurements for positioning estimation of the smartphone. We refer to this method as LC-GNSS/IMU/ML using ML based adaptive positioning (MAP) real-time kinematic (RTK) post-processing algorithm (MAP RTK). This method is validated using reference data from GNSS survey-grade receivers provided with the training datasets. The final validation of this proposed method is done on Kaggle.com, the host of the GSDC competition. Using the proposed method, we estimated the location of the smartphone and tackled the competition. The final public score was 2.61 m, while the final private score was 2.29 m.