Efficient deep learning frameworks for LiDAR point cloud analysis: from classification to 3D object detection

Abstract

Reliable three-dimensional (3D) perception from Light Detection and Ranging (LiDAR) point clouds is essential for autonomous driving, where vehicles must recognize and localize traffic participants from sparse, irregular, and range-dependent point returns to plan and act safely. Existing methods struggle to meet this need efficiently: many high-accuracy classifiers and detectors rely on heavy transformers or deep attention stacks that are poorly matched to real-time or resource-constrained deployment, pillar-based Bird’s-Eye-View (BEV) detectors apply a uniform receptive field regardless of local point density—so sparse, far-range objects receive no more context than dense, near-range ones—and single-resolution BEV grids force a trade-off in which fine cells preserve localization precision but lack scene context, while coarse cells capture context but destroy sub-meter precision in the process. This thesis addresses these limitations through a phased research built around a single key idea: adaptive gating and multi-resolution fusion, rather than uniformly heavier computation to let LiDAR perception be both accurate and efficient. In the first phase, SiGF-Net performs point cloud object classification using a dual-stream architecture in which semantic-derived sigmoid gating adaptively fuses multi-scale geometric features with semantic context (Chapter 3). In the second phase, AG-DPE extends this principle to scene-level LiDAR 3D object detection, using a lightweight density estimator to guide continuous gating across fine, mid, and coarse BEV scales and to condition attention on local point sparsity within a pillar-based pipeline (Chapter 4). In the third phase, DPCA-Net resolves the single-grid detail–context trade-off by encoding each LiDAR sweep into complementary fine and coarse BEV grids and fusing them through bidirectional cross-stream attention at mid-level, followed by semantic-gated injection (Chapter 5). Experiments on standard benchmarks validate each phase. SiGF-Net reaches 94.17% accuracy on ModelNet40 and 81.49% on ScanObjectNN with only 0.15M parameters and 0.52 GFLOPs, giving an efficiency score of 9.56 that exceeds competing methods. AG-DPE improves KITTI validation mAP over a PointPillars baseline from 63.05% to 66.53% on 3D and from 64.08% to 68.81% on BEV with only 24.5K additional parameters (+0.5%), and remains competitive on nuScenes with 0.696 mAP / 0.737 NDS and Waymo with 0.8320 AP/L2 / 0.8170 APH/L2. DPCANet achieves the highest overall Moderate mAP among compared methods on the KITTI validation set holding 75.58% 3D, 73.26% BEV, +11.33 over PointPillars, with the largest gains on Pedestrian and Cyclist, and remains competitive on the Waymo while scoring 88.05 AP L1 / 82.52 APH L2 and nuScenes by 70.0% mAP / 73.9 NDS test leaderboards. Collectively, these results demonstrate that explicit density- and context-aware gating provides an effective, generalizable alternative to uniformly heavy architectures for LiDAR perception, delivering competitive or state-of-the-art accuracy from object-level classification through scene-level 3D detection while remaining efficient enough.