SLOAM: Semantic Lidar Odometry and Mapping for Forest Inventory

Steven W. Chen; Guilherme V. Nardari; Elijah S. Lee; Chao Qu; Xu Liu; Roseli A. F. Romero; Vijay Kumar

Source-linked AI summary

SLOAM: Semantic Lidar Odometry and Mapping for Forest Inventory

Steven W. Chen, Guilherme V. Nardari, Elijah S. Lee, Chao Qu, Xu Liu, Roseli A. F. Romero, Vijay Kumar

arXiv:1912.12726v1Dec 29, 2019cs.ROcs.CVcs.LG

TL;DR

Forest DBH estimation is difficult because natural forests contain clutter and induce challenging sensor motion, while conventional inventory and mapping approaches are limited. The paper combines VR labeling, semantic segmentation, trellis-based instance detection, and semantic lidar odometry and mapping. SLOAM produces robust mapping and DBH estimates, detecting 29 of 35 manually measured trees with 0.67 in average error.

Problem
Forest DBH estimation requires accurate mapping and tree measurement despite cluttered environments, extreme motion, and limited lidar training data.
Method
The paper builds an end-to-end pipeline combining VR point-cloud labeling, range-image FCN segmentation, trellis instance detection, and semantic lidar odometry and mapping.
Results
SLOAM detected 29 trees with 0.67 in average DBH error, while achieving the lowest UAV trajectory drift among compared methods.
Takeaways & Limitations
Semantic shape models support robust and scalable mapping in challenging natural environments while automatically providing tree diameter estimates.

Abstract

from arXiv · show

This paper describes an end-to-end pipeline for tree diameter estimation based on semantic segmentation and lidar odometry and mapping. Accurate mapping of this type of environment is challenging since the ground and the trees are surrounded by leaves, thorns and vines, and the sensor typically experiences extreme motion. We propose a semantic feature based pose optimization that simultaneously refines the tree models while estimating the robot pose. The pipeline utilizes a custom virtual reality tool for labeling 3D scans that is used to train a semantic segmentation network. The masked point cloud is used to compute a trellis graph that identifies individual instances and extracts relevant features that are used by the SLAM module. We show that traditional lidar and image based methods fail in the forest environment on both Unmanned Aerial Vehicle (UAV) and hand-carry systems, while our method is more robust, scalable, and automatically generates tree diameter estimations.

I. INTRODUCTION

The paper presents UAV-based robotic timber inventory using semantic lidar methods to estimate tree diameters while mapping forests. SLOAM addresses brittle conventional features with semantic landmarks and an end-to-end pipeline for challenging motion and natural environments.

Motivation: Accurate timber inventory supports forest management, but conventional TLS requires movement across multiple locations, increasing time and manpower.The proposed robotic timber cruise aims to improve data granularity for management and harvesting decisions.
Motivation: Forest inventory methods based on photography or TLS can be time-consuming and may produce inaccurate DBH estimates under dense forest canopy.
Approach: SLOAM extends semantic SLAM for forestry because LOAM’s texture-based features are brittle in forest environments.
Approach: The pipeline labels 3D scans in virtual reality, segments range images, detects landmark instances with trellis graphs, and performs semantic lidar odometry and mapping.
Contributions: The paper contributes a UAV-based forest DBH estimation pipeline and a semantic framework that estimates landmark parameters during aggressive motion.

II. PROBLEM FORMULATION

The paper formulates forest DBH estimation as semantic LOAM: jointly estimate sensor motion, object instances, and landmark models from lidar sweeps. SLOAM uses semantic tree and ground features to register scans and support the specialized DBH task.

II. PROBLEM FORMULATION: Forest DBH estimation requires estimating the number, positions, and diameters of trees.
II. PROBLEM FORMULATION: Mobile lidar requires registering sequential sweeps in a common frame by estimating the sensor trajectory.
II. PROBLEM FORMULATION: Semantic LOAM jointly estimates the sensor trajectory, object count, object classes, and model parameters from lidar sweeps.
II. PROBLEM FORMULATION: Under assumptions about object parameterization and classes, solving semantic LOAM yields a solution to forest DBH estimation.
III. SEMANTIC LIDAR ODOMETRY AND MAPPING: Lidar odometry estimates 6-DOF motion within a sweep, while lidar mapping estimates world pose and registers the point cloud.
III. SEMANTIC LIDAR ODOMETRY AND MAPPING: SLOAM models forest trees as cylinders and ground as planes, using segmented and clustered tree instances as semantic landmarks.
III. SEMANTIC LIDAR ODOMETRY AND MAPPING: The formulation represents tree landmarks, tree features, ground features, projected models, and pose transforms across successive lidar frames.
III. SEMANTIC LIDAR ODOMETRY AND MAPPING: Odometry and mapping associate current tree and ground features with models from a previous sweep or the map to estimate pose transforms.

A. Model Parameterization and Distance Functions

The method represents ground as a local plane and trees as parameterized cylinders, then uses distances from lidar features to these models for semantic registration and estimation.

Ground model: The ground is modeled locally as a plane defined by its normal and offset, with point-to-plane distance used for fitting.
Ground model: Ground features constrain the vertical axis but not horizontal motion, so tree models and features are also required.
Tree model: Tree landmarks use cylinder parameters ρ, φ, ν, α, and κ to represent position, orientation, and radius.
Tree model: The cylinder normal and axis are computed from the parameterized model and its angular derivatives.
Distance functions: Point-to-cylinder distance provides the geometric residual for comparing tree feature points with cylinder models.

B. Estimating Tree and Ground Models

The method models trees as cylinders and estimates ground planes from semantic feature points. It uses a robust point-to-cylinder distance and SVD-based ground estimation to address forest-specific geometric difficulties.

Tree models: Cylinder parameters are estimated by stacking point-to-cylinder distances for tree feature points and solving a geometric least-squares problem.The optimization yields a cylinder landmark model for each tree instance.
Tree models: The approximated point-to-cylinder distance preserves the true distance’s zero set and derivatives while remaining well behaved as curvature decreases.This approximation avoids singularities that complicate numerical optimization.
Tree models: Optimizing over √κ and √ρ enforces nonnegative curvature and radius parameters during cylinder fitting.The resulting solution provides a cylinder landmark model from the tree features.
Ground models: Forest underbrush can corrupt standard ground-plane estimation because irregular shrubs are difficult to filter completely.Methods using only three points are particularly vulnerable to this noise.
Ground models: The robust ground estimator uses SVD on mean-shifted ground feature points, selecting the plane normal from the singular vector associated with the least singular value.This estimates a local ground-plane model from the available ground features.

C. Motion Estimation and Data Association

Motion estimation associates current tree and ground features with prior models, then optimizes pose using point-to-cylinder and point-to-plane constraints. Two tree-association strategies trade scalability against retained feature detail.

Pose estimation: The pose transform is estimated from current tree and ground features associated with models from the previous sweep or map.The optimization combines tree and ground correspondences for motion estimation.
Tree association: Tree features are projected using an initial pose estimate before landmark correspondences are identified.This projection places features in the frame used for association.
Tree association: One tree-association method matches each feature to the cylinder with the smallest orthogonal distance, while another uses the nearest feature point and its associated cylinder.The two methods differ in whether association is performed directly against models or through neighboring features.
Tree association: The direct model-matching method scales better because the landmark-model set does not grow with the number of observed sweeps.The nearest-feature method can be more accurate but requires a growing feature map.
Generalization: The proposed data-association methods extend beyond forest DBH estimation because they require either convex landmark representations or direct feature-to-feature comparison.Their stated scope includes more general Semantic LOAM applications.
Ground association: Ground features are associated with nearby prior ground points, whose local plane is computed using SVD rather than maintained as an explicit global plane model.This treatment computes a ground plane for each feature.

D. Semantic Lidar Odometry and Mapping

Semantic lidar odometry updates tree models and pose within each sweep, while semantic mapping integrates these outputs into persistent world-frame landmarks. Explicit tree models support automatic tree counts and radius estimates.

Semantic lidar odometry: Semantic lidar odometry takes current tree and ground features plus an initial pose, updates tree models, and solves for a new pose transform.The algorithm follows the LOAM framework while incorporating semantic models and features.
Semantic lidar odometry: Odometry computes cylinder and plane distances for associated tree and ground points before updating the pose through nonlinear least-squares optimization.Tree correspondences use cylinder distances, while ground correspondences use plane distances.
Semantic lidar mapping: The mapping algorithm receives odometry outputs and compares new features with persistent world-frame map parameters rather than only the previous sweep.It initializes an empty landmark map and updates it across subsequent sweeps.
Semantic lidar mapping: Unassigned features are grouped by tree in the odometry output, and sufficiently large groups are added as new world-frame tree cylinders.New tree landmarks are projected into the world frame when inserted.
Integrated system: The combined odometry and mapping process accumulates global poses, feature points, and tree landmark models.This preserves both trajectory information and explicit semantic landmarks.
Integrated system: Explicit tree landmark models allow SLOAM to retrieve the number of trees and estimate their radii automatically at the end of processing.These outputs provide a solution to the paper’s DBH estimation problem.

IV. ROBUST TREE AND GROUND DETECTION

The feature-extraction module is necessary because semantic odometry and mapping depend strongly on reliable tree and ground features. Its three tree-feature modules are intended to generalize to other semantic features.

Feature detection: Semantic odometry and mapping performance depends strongly on the quality and reliability of extracted tree and ground features.Obtaining these features is itself a challenging part of the system.
Feature detection: The proposed front end uses three tree-feature extraction modules together with a specialized ground-segmentation method.These components are designed to provide the inputs required by the downstream semantic SLAM system.
Feature detection: The tree-feature extraction modules can be generalized to other semantic features and combined with broader semantic SLAM applications.The stated generalization applies to the extraction modules rather than only the forestry setting.

A. Virtual Reality Point Cloud Labeling

The authors create a custom Oculus Rift and ROS virtual-reality tool for labeling 3D lidar point clouds, addressing the limitations of conventional 2D-oriented labeling interfaces. A cylinder primitive and in-environment storage support efficient tree-label acquisition.

Conventional mouse-and-keyboard labeling tools designed for 2D images are ill-suited to labeling 3D lidar point clouds.
The custom VR tool interfaces with ROS to obtain and label point clouds from standard mobile lidar platforms.
A labeling cylinder primitive lets users place labels over tree points while rotating, translating, and scaling the point cloud with handsets.
Users can store labeled point clouds and cylinders in a database without leaving the labeling environment, facilitating large label acquisitions.

B. Deep Learning Tree Segmentation

The pipeline segments forest lidar efficiently by processing range images with an FCN, extracting ground heuristically, and organizing tree points into trellis graphs for instance detection and feature initialization.

Natural forests create occlusion and noise that make large-scale lidar tree-point segmentation difficult for clustering-based methods.
The segmentation network processes h × w lidar range images with a simplified ERFNet-inspired FCN to enable faster inference than direct point-cloud architectures.
Sampled range-image predictions are summed into a semantic mask and projected back onto the corresponding 3D lidar points.
Ground extraction divides the lidar sweep into a circular distance-and-angle grid and retains the lowest z points in each cell.
A trellis graph represents beam-wise lidar clusters as vertices and cross-beam connections as edges for grouping segmented points into individual trees.
Greedy shortest routes become tree instances when they satisfy minimum path-length and maximum path-cost thresholds.
The trellis also initializes cylinder parameters through vertex focus points and orthogonal distance regression, while exposing fork structures.

V. EXPERIMENTS

Experiments compare SLOAM with lidar, ICP, and visual-inertial baselines on handheld and aggressive UAV forest data. SLOAM achieves the lowest UAV drift, preserves crisp tree maps, and supports DBH estimation despite the precision demands of forest inventory.

Experimental setup: The evaluation benchmarks SLOAM against A-LOAM, GICP, and the Intel Realsense T265 on medium handheld and hard UAV forest scenarios.The hard UAV experiment includes significant rotation motions over a 65m loop trajectory.
Trajectory evaluation: SLOAM achieves the lowest drift on the hard UAV trajectory, while A-LOAM drifts and the T265 fails outright.GICP tracks closely with SLOAM and achieves a similar magnitude of drift according to trajectory metrics.
Mapping evaluation: A-LOAM and GICP produce blurry maps with frequent tree ghosting, whereas SLOAM preserves fine details in tree shapes.GICP’s comparable positional trajectory metrics nevertheless mask difficulty with rotation motions such as yaw.
Feature comparison: SLOAM’s semantic features improve data association and provide landmark-shape costs, while texture-based features are vulnerable to misassociation during extreme rotations.A-LOAM features appear nondistinctive, and GICP does not distinguish semantic points.
Feature comparison: Point-to-plane and point-to-line costs impose false planar models on cylindrical trees, creating errors that appear as wider, blurry trees.These errors may not cause outright state-estimation failure but are unacceptable for high-precision diameter measurement.
DBH estimation: SLOAM detects 29 of 35 manually measured trees in the hard UAV dataset and estimates DBH from its semantic tree models.Taking the median of radius estimates across beams and scans was more effective than using the cylinder diameter parameter directly; no additional post-processing was required.
Evaluation rationale: Landmark shape ground truth provides a quantitative mapping benchmark because GPS trajectory ground truth can be unavailable under dense forest canopy.Canopy-induced GPS errors can reach tens of meters, whereas landmark shapes are directly measurable.

VI. CONCLUSIONS

The paper frames DBH estimation as semantic lidar odometry and mapping, combining geometric modeling, semantic extraction, and landmark-based optimization. It reports that semantic shape models are critical for accurate and scalable operation in challenging natural environments.

Conclusion: The paper poses DBH estimation as a special case of semantic lidar odometry and mapping with a formulation extensible to other scenarios.It develops an end-to-end system spanning 3D lidar annotation, segmentation, instance detection, odometry, and mapping.
Conclusion: Semantic shape models are critical for achieving accuracy and scalability in challenging natural environments.The conclusion connects these models to the paper’s demonstrated forest DBH-estimation problem.