Abstract
Registration of unmanned aerial vehicle laser scanning (ULS) and ground light detection and ranging (LiDAR) point clouds in forests is critical to create a detailed representation of forest structure. However, forest occlusion and the computational cost of tree segmentation pose significant challenges. We propose an automated "divide-and-conquer" strategy that splits registration into vertical and horizontal alignment. Vertical alignment is achieved through ground normal vector fitting, while horizontal alignment uses canopy projection image matching based on shape context features. Experimental results demonstrate that our method achieves horizontal alignment errors of less than 0.02 m with an average runtime of less than 1 second, proving its efficiency and accuracy for large-scale forest monitoring.
Introduction
Integrating data from multiple LiDAR platforms is essential for a holistic view of forest ecosystems. Unmanned Aerial Vehicle Laser Scanning (ULS) captures the upper canopy, while Terrestrial (TLS) or Backpack Laser Scanning (BLS) provides high-resolution details of the understory and tree trunks. Combining these datasets—a process known as registration—is often the bottleneck in forest data processing.
Traditional registration methods typically rely on artificial markers or GNSS data. In dense forests, markers are difficult to place and often occluded, while GNSS signals are frequently degraded by the canopy. Marker-free methods that rely on individual tree segmentation are often slow and sensitive to segmentation errors. This paper presents a more robust, segmentation-free approach that leverages the global geometry of the forest canopy as a natural fingerprint.
Methodology: A Divide-and-Conquer Strategy
The proposed method breaks the complex 3D registration problem into two simpler stages: vertical alignment and horizontal alignment.
1. Vertical Alignment
The first stage ensures both point clouds share a common vertical orientation. Using the Cloth Simulation Filtering (CSF) algorithm, ground points are extracted from both the ULS and ground-based datasets. The algorithm then fits a plane to these ground points and calculates their normal vectors. By rotating the point clouds so these normal vectors align with the Z-axis, the datasets are vertically normalized and aligned on the Z-axis.
2. Horizontal Alignment via Canopy Shape Context
With the vertical axis fixed, the problem is reduced to a 2D image matching task. Vegetation points above a certain height (typically 3 meters) are projected onto a horizontal plane to create binary images.
- Feature Extraction: Keypoints are identified along the contours of the canopy shapes using a corner detection response.
- Matching: A "two-point congruent set" strategy finds potential matches between the ULS and ground-based images.
- Verification: The final match is determined by the maximum pixel-wise overlap between the two projected canopy maps, filtering out false positives.
3. Fine Registration
Coarse alignment provides a near-perfect starting point for the standard Iterative Closest Point (ICP) algorithm, which performs the final fine-tuning to achieve sub-centimeter precision.
Results and Discussion
The method was rigorously evaluated across various forest types in China, including coniferous (larch, pine) and broadleaf (birch, poplar) stands.
- Efficiency: The coarse alignment process is exceptionally fast, averaging just 0.71 seconds. This is significantly more efficient than object-level registration methods that can take minutes.
- Accuracy: The horizontal alignment error was consistently low, with best-case scenarios under 0.02 m. Even in complex plots, the coarse alignment provided a robust foundation for final registration.
- Robustness: The use of the entire canopy as a "natural marker" makes the method resilient to the occlusion of individual tree trunks and effective across different tree species.
Conclusion
The divide-and-conquer approach successfully addresses the efficiency and robustness issues of forest LiDAR registration. By reducing dimensionality and utilizing canopy shape context, the method enables rapid, automated integration of multi-source point clouds. Future work will focus on adaptive height thresholds for canopy projection to improve performance in extremely dense, multi-layered forest structures.
Summary updated