LiDAR odometry method based on multi-scale fusion and semantic enhancement

In this paper, we propose an end-to-end deep learning-based LiDAR odometry framework addressing key challenges such as point cloud information loss, density variability, and dynamic scene uncertainty. By directly using raw point clouds, our method avoids dimensionality reduction loss and introduces a light-weight geometrically adaptive convolution to improve feature extraction based on local geometric structures. Additionally, a multi-scale fusion and semantic enhancement strategy is employed to incorporate semantic context and optimize pose estimation from coarse to fine. Experimental results on the KITTI dataset show that our approach is competitive with existing methods in accuracy and robustness.