Robust 3D LIDAR Point Cloud Registration Using Uncertainty-Aware Generalized Iterative Closest Point with Voxel-Based Efficiency

Point cloud registration for LIDAR data presents unique challenges due to distance-dependent uncertainty and complex noise characteristics that violate the uniform error assumptions of conventional methods. This paper introduces a novel enhancement to the Generalized Iterative Closest Point (GICP) framework that fundamentally addresses these limitations through three key innovations: First, we derive a physics-based spatial variance term that explicitly models LIDAR-specific uncertainty propagation, enabling sensor-aware weighting of point correspondences based on their physical reliability. Second, we develop a hierarchical outlier rejection strategy combining robust estimation with voxel-density filtering to handle LIDAR-specific artifacts. Third, we implement an adaptive voxelization framework that preserves uncertainty characteristics while accelerating computation. Extensive validation on real-world automotive LIDAR datasets demonstrates that our unified approach reduces absolute trajectory error by 3.52% compared to state-of-the-art methods while operating faster than standard GICP implementations. Furthermore, the spatial variance term provides the theoretical foundation for uncertainty-aware registration, proving particularly effective in challenging scenarios where conventional approaches fail.