SuperiorGAT: Graph Attention Networks for SparseLiDAR Point Cloud Reconstruction in AutonomousSystems
DOI: 10.21203/rs.3.rs-8930448/v1
archive: archived pipeline: cataloged verified
Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)
Summary
This paper introduces SuperiorGAT, a graph attention network framework designed to reconstruct missing elevation (z-coordinate) data in sparse LiDAR point clouds caused by structured beam dropout. The research addresses the critical vulnerability of autonomous vehicle perception systems, where hardware faults, occlusions, or low-cost sensors with fewer scanning beams degrade spatial continuity. Existing methods face a trade-off between accuracy and efficiency: compressive sensing is computationally prohibitive for real-time use, while voxel-based CNNs suffer from memory overhead, and standard Graph Neural Networks often fail to prioritize critical local geometric features. SuperiorGAT aims to bridge this gap by providing high-fidelity reconstruction without increasing network depth or relying on additional sensor modalities. The methodology models LiDAR scans as beam-aware graphs, where nodes represent points enriched with beam index features to capture vertical scanning structure. SuperiorGAT augments standard Graph Attention Networks with gated residual fusion and lightweight feed-forward refinement. This architecture dynamically optimizes point-to-point interactions through learned attention weights while stabilizing feature propagation in sparse regions. The model was evaluated on the KITTI dataset (64-beam sensor) across five environments (Residential, City, Person, Campus, Road) and validated on the nuScenes dataset (32-beam sensor). Experiments simulated structured beam dropout at rates of 12.5%, 25%, and 37.5% to mimic realistic hardware degradation. Performance was measured using vertical Root Mean Square Error (RMSEz), Chamfer distance, and inference latency. Results demonstrate that SuperiorGAT achieves superior reconstruction accuracy and geometric consistency compared to interpolation-based methods, PointNet variants, and deeper GAT baselines. On the KITTI dataset, the model achieved an RMSEz of 0.181 m and an RMSExyz below 0.11 m across all environments, with inference latency reduced to approximately 8.6 ms per frame, down from 13 ms for the GAT baseline. A sensitivity analysis identified a neighborhood size of k=10 as the optimal balance between accuracy and computational cost. Ablation studies confirmed that beam index encoding, gated residuals, and feed-forward refinement are essential components, with the removal of beam indices causing the largest performance degradation. Furthermore, SuperiorGAT maintained robust performance on the lower-resolution nuScenes dataset, exhibiting slower error degradation rates under increasing dropout levels than baseline methods. The significance of this work lies in its provision of a computationally efficient, real-time solution for enhancing LiDAR vertical reconstruction. By focusing on architectural refinement rather than depth, SuperiorGAT enables autonomous systems to maintain high-resolution perception capabilities despite sensor limitations or failures. The findings suggest that targeted graph attention mechanisms can effectively recover structured sparsity patterns, offering a practical pathway for deploying robust perception systems in cost-constrained autonomous driving applications without requiring hardware upgrades.
Provenance
The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.
| Stage | Outcome | Tool | Model | Prompt | Attempts | Completed |
|---|---|---|---|---|---|---|
| discover | success | Crossref | — | — | 1 | 2026-06-25 |
| archive | success | canonical_url | — | — | 1 | 2026-06-26 |
| extract | success | cached | — | — | 2 | 2026-06-26 |
| clean | success | clean | — | — | 1 | 2026-06-25 |
| chunk | success | chunk | — | — | 1 | 2026-06-25 |
| embed | success | embed | Qwen/Qwen3-Embedding-8B | — | 1 | 2026-06-25 |
| promote | success | — | — | — | 1 | 2026-06-25 |
| summarize | success | llm | qwen3.6-27b-prismaquant | summ-v5 | 1 | 2026-06-26 |
| tag | success | vector_similarity | — | — | 6 | 2026-06-25 |
| verify | success | — | — | — | 1 | 2026-06-26 |
Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.
Topics
Ranked by relevance to this paper. Hover a topic for its definition.