SIVR-VIF: System-Level Source-End Inertial Measurement Unit and Visual Reconstruction for Visual–Inertial Fusion

Reliable visual–inertial navigation is essential for autonomous uncrewed aerial vehicle flight in global-positioning-system-denied environments. However, source measurements on low-cost platforms often suffer from coupled inertial and visual degradation. High-frequency noise from the inertial measurement units accumulates during preintegration and enlarges short-term relative motion errors. Low illumination and weak texture reduce feature tracking quality and visual geometric constraints, thereby weakening the correction of global trajectory drift. To address these coupled errors, this study proposes a system-level source-end inertial and visual reconstruction method for visual–inertial fusion. The method jointly optimizes the inertial and image streams before they enter the back-end estimator, without changing the estimator itself. On the inertial side, dual-quaternion-guided resampling, spline reconstruction, and manifold projection suppress high-frequency disturbances in a unified rigid-body motion representation, reducing local relative motion errors. On the visual side, illumination-gated enhancement and grayscale-adaptive super-resolution recover contrast and high-frequency structures in low-light and weak-texture images, strengthening the visual correction of global drift. Experiments on the EuRoC dataset show that the proposed method reduces the macro-average global trajectory error by 5.9%, with a maximum per-sequence reduction of 14.3%, and reduces the average local relative error by about 7.5%. The front-end processing speed remains above 18 frames per second. These results indicate that source-end bimodal reconstruction can improve visual–inertial localization accuracy with low system intrusion and basic feasibility for online deployment.