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

LiDAR-based perception in autonomous systems is fundamentally limited by sparse vertical sampling and further degraded by structured beam dropout caused by occlusions, sensor faults, or reduced-cost LiDAR hardware. These degradations disrupt vertical geometric continuity and negatively affect downstream perception tasks such as object detection, localization, and scene understanding. Existing reconstruction approaches often struggle to balance reconstruction accuracy with the computational efficiency required for real-time autonomous operation. This paper presents SuperiorGAT, a graph attention-based framework for reconstructing missing elevation information in sparse LiDAR point clouds under structured beam loss. The proposed method models LiDAR scans as beam-aware graphs and enhances standard graph attention networks using gated residual fusion and lightweight feed-forward refinement to improve vertical reconstruction fidelity without increasing network depth. The effectiveness of SuperiorGAT is evaluated on multiple KITTI environments, including Person, Road, Campus, and City, as well as through cross-dataset validation on nuScenes with lower vertical resolution. Additional experiments under severe structured sparsity further evaluate robustness in a 16-beam-equivalent sensing condition. Results demonstrate that SuperiorGAT achieves lower overall reconstruction error and improved geometric consistency compared to interpolation-based methods, PointNet-based models, and standard GAT baselines while maintaining computational efficiency suitable for real-time perception pipelines.