CredibleDFGO: Differentiable Factor Graph Optimization with Credibility Supervision
Researchers have introduced CredibleDFGO (CDFGO), a novel differentiable GNSS factor graph framework designed to enhance positioning reliability in urban environments. While Global Navigation Satellite System (GNSS) positioning is standard for urban navigation, existing solvers often report unreliable covariance data in urban canyons. Previous differentiable factor graph optimization methods improved mean estimates but failed to correct covariance shapes or scales. CDFGO addresses this by making covariance credibility an explicit training target. It utilizes a Weighting Generation Network to predict per-satellite reliability weights, which are mapped to position estimates and posterior covariance via a differentiable Gauss-Newton solver. The system is supervised end-to-end using proper scoring rules like Negative Log-Likelihood and Energy Score. Testing on UrbanNav scenes demonstrated significant improvements in uncertainty credibility and positioning accuracy. In harsh-urban Mong Kok scenarios, CDFGO reduced mean horizontal error from 13.77m to 11.68m and drastically lowered uncertainty metrics, proving its effectiveness in delivering more credible local covariance ellipses and consistent axis-wise performance.
Wire timeline
CredibleDFGO: Differentiable Factor Graph Optimization with Credibility Supervision
Researchers have introduced CredibleDFGO (CDFGO), a novel differentiable GNSS factor graph framework designed to enhance positioning reliability in urban environments. While Global Navigation Satellite System (GNSS) positioning is standard for urban navigation, existing solvers often report unreliable covariance data in urban canyons. Previous differentiable factor graph optimization methods improved mean estimates but failed to correct covariance shapes or scales. CDFGO addresses this by making covariance credibility an explicit training target. It utilizes a Weighting Generation Network to predict per-satellite reliability weights, which are mapped to position estimates and posterior covariance via a differentiable Gauss-Newton solver. The system is supervised end-to-end using proper scoring rules like Negative Log-Likelihood and Energy Score. Testing on UrbanNav scenes demonstrated significant improvements in uncertainty credibility and positioning accuracy. In harsh-urban Mong Kok scenarios, CDFGO reduced mean horizontal error from 13.77m to 11.68m and drastically lowered uncertainty metrics, proving its effectiveness in delivering more credible local covariance ellipses and consistent axis-wise performance.
cs.AI updates on arXiv.org