LiDAR-only HD Map Construction

• LiDAR-only HD map construction is the creation of detailed digital maps using solely LiDAR data, emphasizing geometric precision and semantic annotations.
• It leverages robust SLAM techniques and deep learning methods, including multi-scale BEV feature extraction and knowledge distillation, to achieve competitive accuracy.
• Emerging solutions eliminate GNSS reliance and enhance performance in diverse conditions, making them crucial for autonomous driving and smart urban infrastructure.

LiDAR-only high-definition (HD) map construction refers to the process of creating detailed, accurate, and semantically rich digital representations of roadways and urban environments using only LiDAR sensor data, without reliance on camera imagery or Global Navigation Satellite System (GNSS) measurements for georeferencing or semantic augmentation. These maps combine geometric precision with semantic annotations to support downstream tasks in autonomous driving, robotics, and smart mobility infrastructure. With advances in deep learning, multi-scale BEV (bird’s-eye view) representation, knowledge distillation, and robust SLAM, LiDAR-only pipelines now achieve accuracy and robustness competitive with or superior to camera-based or multimodal systems.

1. Core Methodologies in LiDAR-only HD Map Construction

LiDAR-only HD map systems encompass two principal paradigms: (i) geometric/SLAM-based pipelines that focus on producing globally consistent, georeferenced metric maps (often as large point clouds), and (ii) data-driven semantic mapping pipelines that yield semantic segmentation or vectorized (polyline) representations in BEV. Both rely on the inherent 3D structure and reflectivity information captured by LiDAR sensors.

Geometric approaches, exemplified by OpenLiDARMap (Kulmer et al., 19 Jan 2025), leverage iterative closest point (ICP) algorithms, pose-graph optimization, and external structural priors (e.g., building footprints, digital elevation models) to generate dense, drift-free, globally anchored maps even in GNSS-denied environments. These pipelines often maintain local submaps for scan-to-scan registration, aligning frames using a combination of scan-to-map and scan-to-scan ICP objectives, and solve the trajectory using non-linear least squares with robust loss functions to suppress outliers.

Semantic segmentation and vectorization methods, as advanced by HDMapNet (Li et al., 2021), LiDAR2Map (Wang et al., 2023), LIE (Mazumder et al., 2 May 2026), MGMap (Liu et al., 2024), and MambaMap (Yang et al., 27 Jul 2025), involve converting raw point clouds into pillar or voxel representations, extracting BEV features via deep backbones (SECOND, PointPillars), and decoding to semantic labels and/or vector primitives. These systems frequently employ multi-scale feature decoders, mask guidance, deformable attention, or temporal state-space models for online, robust instance prediction.

2. Geometric Consistency and Global Georeferencing

A core innovation in LiDAR-only mapping is the removal of GNSS for global anchoring. OpenLiDARMap (Kulmer et al., 19 Jan 2025) constructs a sparse, georeferenced reference map from publicly available building outlines (e.g., OSM) and digital surface models (GeoTIFF, airborne LiDAR), tessellating façades and ground at meter-level resolution. Each live LiDAR scan is downsampled and matched via ICP both to this static reference (providing an absolute pose) and to a local rolling submap (providing a relative constraint). The resulting factor graph combines these absolute and relative pose constraints:

• For scan-to-scan: $$f_{\text{ss}}(X_{k-1}, X_k)=\log((X_{k-1})^{-1}X_k)\ominus\Delta T_k$$
• For scan-to-map: $$f_{\text{sm}}(X_k)=\log(X_k\ominus T_k^{\text{map}})$$

Optimization is performed over poses $X_k$ with a Cauchy loss for relative and aggressive Tukey loss for absolute constraints, admitting a robust, zero-drift estimate of trajectory and map points. The final dense map inherits global coordinates directly from the prior map's reference frame (UTM/lat-lon). Experimental results show mean absolute trajectory error (ATE) below 1m for urban and mixed environments (e.g., KITTI: 0.66m, EDGAR: 0.52m), and centimeter-level global overlays on aerial orthophotos, outperforming prior SLAM/odometry (Kulmer et al., 19 Jan 2025).

3. Semantic Map Construction via Pure LiDAR Deep Learning

Semantic segmentation and map vectorization from LiDAR faces the challenge of weak texture and low semantic richness, especially for elements with poor 3D cues (painted lanes, crosswalks). Purely LiDAR-based BEV semantic mapping is addressed in HDMapNet (Li et al., 2021), LiDAR2Map (Wang et al., 2023), and LIE (Mazumder et al., 2 May 2026).

• BEV Feature Extraction: PointPillars or VoxelNet encoders partition point clouds into non-overlapping pillars or voxels, aggregate per-pillar features with small PointNet or sparse 3D CNNs, and scatter to 2D BEV maps (e.g., $$F_\text{LiDAR}^{\text{BEV}} \in \mathbb{R}^{C\times H\times W}$$).
• Semantic Decoding: Multi-scale feature pyramid decoders (BEV-FPD) or U-Net style upsampling fuse spatial cues across scales. Only the LiDAR branches are retained at inference for real-time deployment (e.g., 35–36 FPS on RTX 2080/4090; model size $\approx$8–9M parameters) (Mazumder et al., 2 May 2026, Wang et al., 2023).
• Knowledge Distillation: To overcome low semantic density, knowledge distillation (KD) schemes borrow semantic context from camera or intensity modality during training, but deploy only LiDAR at inference. In LIE (Mazumder et al., 2 May 2026), an online KD pipeline fuses LiDAR pillars and high-resolution intensity map tiles (precomputed from dense LiDAR accumulation, with normalized reflectivity) via a position-guided cross-modal fusion (PGxMF), distills global and local affinity at feature level, and softmax logits at output. This strategy yields $+8.2$ mIoU advantage vs. SOTA camera-based models on nuScenes and $+1.0$ mIoU over prior LiDAR KD methods (Mazumder et al., 2 May 2026).
• Vectorization: HDMapNet (Li et al., 2021), MGMap (Liu et al., 2024), and MambaMap (Yang et al., 27 Jul 2025) segment polylines or instances using clustering, mask-activated decoders, transformer or state-space sequence models, and directional/positional prediction for polyline points. MGMap achieves $+12.3$ mAP over MapTR on nuScenes (67.9 vs 55.6), while MambaMap attains 67.3 mAP (nuScenes, 60×30m), surpassing prior transformer-based temporally fused systems at lower complexity.

4. Temporal Fusion and Online Map Updates

Temporal modeling of LiDAR sequences enhances robustness to occlusion, sensor noise, and sparsity by propagating and fusing context over time:

• Short-term Fusion: Warping and max/mean pooling of BEV features from prior frames using ego-motion transformations as in HDMapNet, resulting in up to 3.5–4 percentage point IoU gain (Li et al., 2021).
• Sequence Modeling: MambaMap (Yang et al., 27 Jul 2025) introduces state-space models (SSMs) and a memory bank to store BEV and instance features over a sliding window, fusing context with multi-directional (left/right/up/down) and spatial-temporal scanning. This architecture achieves linear computational complexity ($O(L+D)$ per sequence), supporting real-time online construction with mAP up to 67.3 (nuScenes, 60×30 m) and maintaining strong performance under geographical distributional shifts.

5. HD Map Maintenance and Change Detection from LiDAR

Long-term viability of HD maps for smart mobility applications requires robust change detection and maintenance, all from LiDAR measurements. Urban 3D change detection pipelines operate as follows (Albagami et al., 24 Oct 2025):

• Multi-resolution NDT Alignment & Point-to-Plane ICP: Accurate bi-temporal registration (across years), capturing rigid and non-rigid changes at meter-to-centimeter scales.
• Height Normalization & Level of Detection (LoD): Per-tile ground normalization and site-specific LoD estimates propagate registration and roughness uncertainty into statistical change thresholds.
• Instance-Centric Association: Geometry- and class-constrained bipartite assignment with “dummy” penalties enables accurate tracking of split/merge events and resolves object correspondences under uncertainty.
• Confidence Gating & Tiling: Changes are scored and accepted only if exceeding local LoD, suppressing false positives from partial overlap or sampling variation. Scalable tiling avoids memory issues and preserves narrow ground changes.
• Evaluation: Achieves 95.2% accuracy, 90.4% macro F1, 82.6% mIoU on city-scale blocks, with >7 point IoU gain on hardest classes vs. neural baselines (Albagami et al., 24 Oct 2025).

6. Comparative Quantitative Performance

The emergence of robust KD strategies and mask-guided learning has enabled LiDAR-only methods to outperform camera-only, and even camera–LiDAR fusion approaches on standard HD map benchmarks. Table: Sample mIoU/mAP results on nuScenes (val, 60×30m, three classes) (Mazumder et al., 2 May 2026, Wang et al., 2023, Liu et al., 2024, Yang et al., 27 Jul 2025, Li et al., 2021):

Method	Modality	Backbone	mIoU / mAP
HDMapNet-LiDAR	LiDAR	PointPillars	29.5 mIoU
LiDAR2Map	LiDAR	PointPillars	57.4 mIoU
LIE	LiDAR	PointPillars	58.0 mIoU
MGMap	LiDAR	SECOND	67.9 mAP
MambaMap	LiDAR	ResNet50	67.3 mAP

LiDAR-only pipelines maintain real-time throughput (12–36 FPS, hardware dependent), strong generalization under environmental variation (rain, night), and effective cross-dataset transfer with minimal fine-tuning (e.g., LIE achieves mIoU 50.8 on Argoverse2 with only 10% fine-tuning) (Mazumder et al., 2 May 2026).

7. Strengths, Limitations, and Future Directions

LiDAR-only HD map construction is characterized by:

Strengths:

• Immunity to lighting and weather; not reliant on camera calibration.
• High geometric fidelity; direct 3D structure measurement.
• Robust global georeferencing via public priors without GNSS (Kulmer et al., 19 Jan 2025).
• Real-time online mapping and map maintenance; scalable to city scale (Albagami et al., 24 Oct 2025, Yang et al., 27 Jul 2025).
• Superior or comparable semantic performance to camera-based alternatives via intensity enhancement and KD (Mazumder et al., 2 May 2026, Wang et al., 2023).

Limitations:

• Semantic ambiguity for purely appearance-based map elements (faint paint, severe occlusion).
• Need for dense aerial map priors (for global anchoring) in geometric pipelines (Kulmer et al., 19 Jan 2025).
• Single-sweep semantic pipelines (LIE, LiDAR2Map) lack temporal fusion, limiting context in highly dynamic environments.
• Mask- and KD-based methods require complex training infrastructure and precomputation for intensity/affinity maps.

Emerging directions include 4D map sequence prediction, self-supervised pretraining on large-scale unlabeled data, online joint camera–LiDAR hybridization at inference for maximum robustness, and explicit incorporation of road network priors (e.g., curb lines, vector maps).

LiDAR-only HD map construction now constitutes a mature and high-performing alternative for both metric and semantic mapping tasks, advancing the autonomy, robustness, and scalability of intelligent transportation and urban digital twin systems (Kulmer et al., 19 Jan 2025, Mazumder et al., 2 May 2026, Albagami et al., 24 Oct 2025, Liu et al., 2024, Yang et al., 27 Jul 2025, Li et al., 2021, Wang et al., 2023).