Skip to content
Open access

GRNet: Global Resolution LiDAR--Camera Drift Correction Using Mixed Loss Network

Nov 2026 · IEEE Sensors Journal · Vol 26, pp. 32802-32815 · 0 citations · 27 references

Abstract

Multisensor fusion of LiDAR and cameras plays a crucial role in advancing technologies such as environmental perception and 3-D reconstruction in autonomous vehicles and robotic systems. Achieving precise extrinsic calibration for heterogeneous sensors is essential, since it is the foundational step for effective and precise data fusion. In autonomous systems, sensor extrinsics can drift over time due to temperature changes and mechanical vibrations, leading to misalignment between modalities and degraded perception performance. In this article, we propose GRNet, an end-to-end network for LiDAR–camera extrinsic drift correction. The proposed method corrects sensor misalignment by estimating extrinsic deviations between the LiDAR and camera. We introduce improvements to three critical components of the baseline design, achieving superior performance compared to existing methods. First, our model introduces a feature pyramid network (FPN)-based global resolution module, which extracts and fuses features of each modality with both fine details and global semantic information. Second, a cross-modal attention mechanism obtains the alignment information of the same pixel in the RGB image and the depth map. Finally, a mixed loss concept is proposed. It comprehensively accounts for both the optical imaging characteristics of the camera and the imaging characteristics of spatial points in 2-D images. This helps the model minimize error more effectively, leading to more accurate extrinsic drift correction. During inference, the extrinsic drift is corrected in a coarse-to-fine manner using three networks with identical architectures but different precision levels, resulting in accurate LiDAR–camera extrinsic estimation. Our model is evaluated on the KITTI odometry dataset. The results demonstrate that GRNet outperforms existing baseline models, achieving a mean absolute error of 0.3487 cm in translation and 0.0382° in rotation. In terms of efficiency, the three-stage cascade corrects one frame in 148.56 ms (6.73 FPS) on an RTX 3090, faster end-to-end than the four-stage baseline, and GRNet is intended for periodic online drift correction.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.