LDGC-VIS: A Lightweight Differentiable Geometric-Constraint Framework for Visual-Inertial Relative Pose Estimation on Micro-Nano Uncrewed Vehicles
Abstract
Visual-inertial relative pose estimation in Global Positioning System (GPS)-denied environments remains challenging for micro-nano uncrewed vehicles under strict size, weight, and power (SWaP) constraints. We propose LDGC-VIS, a lightweight differentiable geometric-constraint framework for monocular visual-inertial relative pose estimation on embedded platforms. LDGC-VIS combines an edge-oriented monocular depth module to provide dense geometric priors, a feature and confidence weights module to emphasize reliable image regions, a differentiable inverse compositional pose module with learnable residual-adaptive damping, and an optional inertial measurement unit (IMU)-assisted post-fusion module for residual pose refinement. Compared with Reloc3r-L, the proposed approach obtains lower end-point error (EPE) on unseen TUM KF4 in zero-shot evaluation, remains close on TUM KF8 and ScanNet1500, and uses only 12.1 M parameters. In offline evaluation on recorded indoor micro aerial vehicle (MAV) flight data, it maintains sub-meter accuracy with 0.74%–1.52% drift over trajectories up to 36.55 m across 224.84 s of recorded flight data. In a separate deployment benchmark, LDGC-VIS runs at 23 Hz on RK3588 s with mixed-precision quantization.