Rolling-Shutter Aware LiDAR-Inertial-Visual Odometry
Abstract
This paper presents FAST-LIVO-RS, a rolling-shutter (RS) aware direct LiDAR-inertial-visual odometry system that addresses the row-dependent geometric inconsistency arising when RS images are processed under global-shutter (GS) assumptions. In direct LIVO pipelines, this inconsistency degrades photometric optimization, point-cloud colorization, and robustness under fast motion. To capture intra-frame motion over the RS readout time, FAST-LIVO-RS introduces adaptive state augmentation within a sequential error-state iterated Kalman filter (ESIKF), dynamically inserting clone states during IMU propagation based on estimated angular and linear acceleration to provide a motion-adaptive piecewise-linear trajectory approximation. Building on this representation, a row-consistent direct visual update is formulated through fixed-point row-time projection and Jacobian distribution to clone-segment endpoints, ensuring that each photometric residual is linearized at the correct row-specific camera pose. RS-consistent visual map management keeps patch maintenance and reference-frame updates consistent with row-specific RS geometry rather than single end pose assumptions. We evaluate FAST-LIVO-RS on a self-collected RS dataset and four GrandTour tasks, together with a comparison using synchronized RS and GS cameras and 3D reconstruction. The results show that row-time modeling improves robustness on RS sequences and provides localization and reconstruction quality comparable to FAST-LIVO2 using the synchronized GS camera.