TreeLoc: LiDAR Localization in Forest Environments

• TreeLoc is a learning-free, geometry-driven framework for 6-DoF LiDAR localization that leverages tree stem features to address repetitive, occluded forest scenes.
• It extracts and geometrically describes tree stems using a compact Tree Distribution Histogram and invariant triangle matching for robust place recognition and pose estimation.
• Empirical results show TreeLoc outperforms urban-centric methods with higher recall, lower localization errors, and efficient storage for large-scale forest mapping.

TreeLoc is a learning-free, geometry-driven framework for 6-DoF LiDAR global localization in forest environments, designed to address the unique perceptual and algorithmic challenges posed by repetitive, occluded, and structurally ambiguous tree-dominated scenes where traditional, urban-centric localization methods fail. Its pipeline centers around the detection and geometric description of tree stems, combining a compact global descriptor (the Tree Distribution Histogram, or TDH) for place recognition with robust local verification via inter-tree triangle features and precise geometric alignment, resulting in accurate pose estimation even under degraded GPS and cross-session misalignments (Jung et al., 2 Feb 2026).

1. Challenges of Forest LiDAR Localization

Forested environments introduce several challenges to LiDAR-based global localization:

• Repetitive and Ambiguous Structure: Trees are often closely packed, cylindrical, and exhibit little local uniqueness, minimizing the availability of distinctive features.
• Occlusion and Seasonal Variation: Dense undergrowth and variable foliage produce frequent occlusions and appearance changes across time, complicating feature correspondence.
• Terrain Complexity: Irregular terrain and the absence of large planar surfaces render ground-plane-based methods unreliable.
• Degraded GPS: Dense canopies attenuate GPS signals, leading to significantly misaligned SLAM trajectories across repeated traversals.

Urban-centric LiDAR localization approaches, such as Scan Context, RING++, BTC, and STD, rely on assumptions (planar roads, building edges, distinctive reflectivity) that do not generalize to natural, forested settings. As a result, these methods often produce ambiguous place recognition and inaccurate 6-DoF pose estimation in forests (Jung et al., 2 Feb 2026).

2. Scene Representation: Tree-Centric Parameterization

2.1 Payload Aggregation and Tree Extraction

LiDAR scans are partitioned into overlapping payload windows $\mathcal{P}_u$, aggregated using transformations $\mathbf{T}_{t\leftarrow u}$ (from SLAM, e.g., FAST-LIO2), yielding submaps: $$\mathcal{Z}_t = \bigcup_{u\in\mathcal{W}_t}\mathbf{T}_{t\leftarrow u}\,\mathcal{P}_u$$ RealtimeTrees segments vertical clusters, identifying tree stems, and fits geometric parameters for each tree $j$:

• Axis orientation $\mathbf{A}_j\in SO(3)$, trunk direction $\mathbf{a}_j$
• Stem center $\mathbf{p}'_j=(c'_{j,x},c'_{j,y},b'_j)$ (with $b'_j$ as base height)
• Diameter at breast height $d_j$ via circular fit

The forest scene at time $t$ is summarized as: $$\mathbf{M}_t = (\mathbf{T}_t, \mathcal{I}_t),\quad \mathcal{I}_t = \{(\mathbf{A}_j, \mathbf{p}'_j, d_j)\}_{j=1}^{n_t}$$

2.2 Roll–Pitch Correction and 2D Projection

To eliminate roll and pitch variance, a rotation $\mathbf{R}^A_t\in SO(3)$ aligns all stem axes $\{\mathbf{a}_j\}$ to the world-up vector $\mathbf{e}_z$: $$\mathbf{R}^A_t = \arg\min_{R \in SO(3)} \sum_j \Bigl(1 - \bigl|\mathbf{e}_z^\top R\,\mathbf{a}_j\bigr|\Bigr)^2$$ Projected horizontal centers are: $$\mathbf{c}_j = (I - \mathbf{e}_z\mathbf{e}_z^\top)(\mathbf{R}^A_t\mathbf{p}'_j) \in \mathbb{R}^2$$

2.3 Tree Distribution Histogram (TDH)

TDH summarizes the spatial and size distribution of trees:

• Radial bins $i=1,\ldots,n_r$ (width $r_{\mathrm{res}}$), DBH bins $k=1,\ldots,n_d$ (width $w_{\mathrm{dbh}}$)
• Tree $j$ assigned via

$$i_j = \Bigl\lfloor\|\mathbf{c}_j\|/r_{\mathrm{res}}\Bigr\rfloor,\quad k_j = \Bigl\lfloor(d_j - d_{\min})/w_{\mathrm{dbh}}\Bigr\rfloor$$

• Histogram elements:

$H_t(i,k) = |\{j : i_j=i, k_j=k\}|$

• Optionally smoothed by a $2\times2$ filter. With typical settings ($n_r=5$, $n_d=8$), $H_t$ is a 40-dimensional descriptor.

3. Two-Stage Matching: Coarse TDH and Fine Triangle Features

3.1 Coarse Place Recognition via TDH

Candidate matches are measured by chi-square distance: $$D_{\chi^2}(q, m) = \tfrac12\sum_{i,k}\frac{\bigl[H_q(i,k)-H_m(i,k)\bigr]^2} {H_q(i,k)+H_m(i,k)+\varepsilon}$$ The $K_{\mathrm{coarse}}$ (e.g., 100) lowest-distance candidates advance to fine matching.

3.2 Fine Matching via 2D Triangle Descriptor

Given 2D centers $\{\mathbf{c}_j\}$, all unordered triples $(i, j, k)$ yield triangles with side-lengths $\ell_{ab}$ and centroid $\mathbf{q}_{ijk}$. Sorted side-lengths are hashed, producing translation- and rotation-invariant keys $h_{ijk}$: $$h_{ijk} = \mathrm{hash}\bigl(\mathrm{sort}(\ell_{ij}, \ell_{jk}, \ell_{ik})\bigr)$$ The similarity between query and candidate scenes is $S(q, c) = |\mathcal{K}_q \cap \mathcal{K}_c|$; the $K_{\mathrm{fine}}$ (e.g., 10) most similar are retained.

4. Geometric Verification and 6-DoF Pose Estimation

4.1 Initial 4-DoF Alignment

For each matched triangle, centroid correspondences $(\mathbf{q}^q_u, \mathbf{q}^c_u)$ are found. Considering planar transforms (SO(2) rotation, $\mathbb{R}^2$ translation), the closed-form SVD alignment yields: $$\min_{R_c\in SO(2),\,t_c\in\mathbb R^2} \sum_{u}\|\mathbf{q}^c_u - (R_c\mathbf{q}^q_u + t_c)\|^2$$

4.2 Refined Alignment and Vertical Offset Estimation

Query centers are transformed by the initial planar alignment, then candidate matches within 0.4 m (Euclidean) and 0.2 m (DBH) are identified using a 2D KD-tree. RANSAC on base heights estimates vertical offset $\Delta z$ by minimizing $|b^c_j - b^q_i - \Delta z|$. The SVD alignment is refit to the set of inlier pairs, and a 4-DoF SE(3) transform is assembled, with vertical offset as the $z$ translation.

4.3 Overlap Criterion and Final 6-DoF Transformation

Matched sets $\mathcal{M}_{q,c}$ are used to compute overlap ratio: $$\mathcal{O}(q, c) = \frac{|\mathcal{M}_{q,c}|}{|T_q| + |T_c| - |\mathcal{M}_{q,c}|}$$ Candidates are ranked by $\mathcal{O}$, and the best is selected. The final 6-DoF transform in the world frame is

$T^{6D} = (T^A_c)^{-1} T^{4D} T^A_q$

where $T^A_q$ and $T^A_c$ are axis-alignment transforms.

5. Empirical Evaluation and Ablation Studies

TreeLoc achieves superior performance over prior baselines, particularly BTC, across representative forest datasets and alignment tasks:

Task	TreeLoc (Best)	BTC (Baseline)
Place Recog. (R@1/F1/AUC, Oxford Evo)	0.907/0.966/0.992	0.626/0.804/0.868
Place Recog. (R@1/F1/AUC, Venman V-04)	0.890/0.942/0.974	0.353/0.661/0.673
6-DoF Loc. (R@[email protected]/5°, K-04)	0.970/0.987	0.638/0.780
Median TE/RE	0.053 m / 0.137°	0.225 m / 0.641°
Multi-sess. ATE/ARE	0.248 m / 0.492°	2.491 m / 3.484°

Ablation results confirm each pipeline stage is critical: omitting the TDH increases fine-matching search space and reduces R@1 by 3–5%, removing DBH from TDH drops Recall@1 by 4%, and omitting axis-alignment impacts R@1 by 3–4%. Reliance on ground-plane fitting instead of tree-axis alignment leads to a 5% decrease in F1 score, especially on uneven terrain (Jung et al., 2 Feb 2026).

6. Applications and Compact Global Tree Database

TreeLoc enables efficient, large-scale forest localization and inventory:

• Global Database: Aggregating all observed trees as $$(\mathrm{ID}_j, \mathbf{A}_j, \mathbf{c}_j, d_j, b_j)$$ across missions.
• Storage Efficiency: For three missions (1,462 scenes), TreeLoc’s database measures 267 KB versus BTC's 1.6 GB and raw point clouds' 4.9 GB.
• On-Demand Descriptor Generation: TDH and triangle features can be generated in approximately 1.4 ms per location.
• Map Updates and Multi-Session Alignment: Supports incremental updates and loop closures with $\mathcal{O}>0.2$.
• Long-Term Monitoring: Facilitates digital forest inventory, ecological monitoring, and under-canopy robotic navigation with minimal storage and computational requirements.

7. Distinctiveness and Interpretability

All components of TreeLoc are learning-free, interpretable, and constructed from explicit geometric equations and threshold criteria. The modular pipeline, reliance on stem-centric geometric primitives, and absence of data-driven fitting enable straightforward adaptation to new environments and transparent error analysis. This property distinguishes TreeLoc within the broader class of forest localization methods and supports long-term maintainability in mission-critical and ecological monitoring applications (Jung et al., 2 Feb 2026).