Differentiable Factor Graph Optimization with Intelligent Covariance Adaptation for Accurate Smartphone Positioning

Factor graph optimization (FGO) has recently become a popular approach for Global Navigation Satellite System (GNSS) positioning. Weighting being used in the FGO is crucial to the performance. Great efforts have been spent in deciding the weighting with various approaches, including uncertainties estimated from devices, statistical modeling, empirical model, etc. In this paper, we proposed a differentiable FGO (DFGO) for intelligent covariance adaptation to determine the weighting for state estimation. Adaptive weightings are obtained via the neural network, which is trained with ground truth location via backpropagation. The end-to-end training for the DFGO can make use of the information from the data while maintaining the benefits of the estimation framework. The weighting from the DFGO is demonstrated to perform better in positioning than the weighting calculated from the measurement uncertainties, which shows the feasibility of the DFGO approach.