SuperiorGAT: graph attention networks for sparse LiDAR point cloud reconstruction in autonomous systems

Area of Science:

• Robotics and Autonomous Systems
• Computer Vision
• Sensor Fusion

Background:

• LiDAR perception is limited by sparse vertical sampling and structured beam dropout, degrading autonomous system performance.
• Existing reconstruction methods lack the accuracy-computational efficiency balance for real-time applications.
• Disrupted vertical geometric continuity negatively impacts object detection, localization, and scene understanding.

Purpose of the Study:

• To present SuperiorGAT, a novel graph attention framework for reconstructing missing elevation information in sparse LiDAR point clouds.
• To address structured beam loss caused by occlusions, sensor faults, or hardware limitations.
• To improve vertical reconstruction fidelity without increasing network depth.

Main Methods:

• Modeled LiDAR scans as beam-aware graphs.
• Enhanced standard graph attention networks (GAT) with gated residual fusion and feed-forward refinement.
• Evaluated on KITTI datasets (Person, Road, Campus, City) and nuScenes with cross-dataset validation.

Main Results:

• SuperiorGAT achieved lower overall reconstruction error and improved geometric consistency compared to interpolation, PointNet, and standard GAT.
• Demonstrated robustness under severe structured sparsity (16-beam-equivalent).
• Maintained computational efficiency suitable for real-time autonomous perception.

Conclusions:

• SuperiorGAT effectively reconstructs missing LiDAR elevation data under structured beam loss.
• The framework offers a superior balance of accuracy and efficiency for real-time autonomous systems.
• SuperiorGAT enhances downstream perception tasks by restoring vertical geometric continuity.