<p>Polar navigation with low-cost strapdown inertial navigation systems (SINS) is challenged by weak heading observability, GPS velocity degradation, and visually difficult scenes containing weak snow texture, strong ice reflection, low visibility, and motion blur. This paper proposes a learning-based quality-aware visual-inertial GPS/SINS coarse alignment method in the transverse frame. Image feature tracking and IMU preintegration are used to form short-window visual-inertial relative-motion constraints, which are mapped into auxiliary transverse-frame velocity and attitude observations. A closed-loop Kalman filtering structure jointly estimates the time-varying body-frame attitude relationship and gyroscope bias, and feeds the corrected states back to the velocity-vector construction process. To prevent unreliable visual observations from degrading the alignment filter, a lightweight learned quality model predicts the visual-inertial measurement weight from feature statistics, inlier ratio, reprojection error, image-degradation indicators, and preintegration residuals. Under degraded GPS observations, the proposed method reduces the heading RMSE from <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(0.650^\circ \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0</mn> <mo>.</mo> <msup> <mn>650</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation> for the GPS/SINS closed-loop baseline to <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(0.251^\circ \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0</mn> <mo>.</mo> <msup> <mn>251</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation>, and reduces the maximum heading error from <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(1.286^\circ \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>1</mn> <mo>.</mo> <msup> <mn>286</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation> to <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(0.612^\circ \)</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0</mn> <mo>.</mo> <msup> <mn>612</mn> <mo>∘</mo> </msup> </mrow> </math></EquationSource> </InlineEquation>. Direct comparisons with VINS-Mono, OpenVINS, and ORB-SLAM3 in the same calibrated data protocol further show the advantage of transverse-frame closed-loop bias compensation for coarse alignment. Numerical simulations, semi-physical vehicle validation using virtual polar-region transformation, and VR-based polar scene simulations demonstrate robust alignment performance under GPS fluctuation and visual degradation. Code, simulation scripts, selected VR scene assets, and example data are available at <a href="https://github.com/huangling8920/Coarse-Alignment.git.">https://github.com/huangling8920/Coarse-Alignment.git.</a></p>

错误:搜索内容不能为空,请输入英文关键词
错误:关键词超出字数限制,请精简
高级检索

Quality-aware visual-inertial GPS/SINS coarse alignment in a transverse frame for robust polar navigation

  • Ling Huang,
  • Nan Luo,
  • Dongyan Wu

摘要

Polar navigation with low-cost strapdown inertial navigation systems (SINS) is challenged by weak heading observability, GPS velocity degradation, and visually difficult scenes containing weak snow texture, strong ice reflection, low visibility, and motion blur. This paper proposes a learning-based quality-aware visual-inertial GPS/SINS coarse alignment method in the transverse frame. Image feature tracking and IMU preintegration are used to form short-window visual-inertial relative-motion constraints, which are mapped into auxiliary transverse-frame velocity and attitude observations. A closed-loop Kalman filtering structure jointly estimates the time-varying body-frame attitude relationship and gyroscope bias, and feeds the corrected states back to the velocity-vector construction process. To prevent unreliable visual observations from degrading the alignment filter, a lightweight learned quality model predicts the visual-inertial measurement weight from feature statistics, inlier ratio, reprojection error, image-degradation indicators, and preintegration residuals. Under degraded GPS observations, the proposed method reduces the heading RMSE from \(0.650^\circ \) 0 . 650 for the GPS/SINS closed-loop baseline to \(0.251^\circ \) 0 . 251 , and reduces the maximum heading error from \(1.286^\circ \) 1 . 286 to \(0.612^\circ \) 0 . 612 . Direct comparisons with VINS-Mono, OpenVINS, and ORB-SLAM3 in the same calibrated data protocol further show the advantage of transverse-frame closed-loop bias compensation for coarse alignment. Numerical simulations, semi-physical vehicle validation using virtual polar-region transformation, and VR-based polar scene simulations demonstrate robust alignment performance under GPS fluctuation and visual degradation. Code, simulation scripts, selected VR scene assets, and example data are available at https://github.com/huangling8920/Coarse-Alignment.git.