-
In the field of forest ecology research and resource management, accurately acquiring and analyzing high-precision three-dimensional structural information is crucial for precisely representing the morphology, structure, and dynamic changes of forests. Through the utilization of LiDAR technologies such as Terrestrial Laser Scanning (TLS), Backpack Laser Scanning (BLS), Mobile Laser Scanning (MLS), Airborne Laser Scanning (ALS), Unmanned Aerial Vehicle Laser Scanning (ULS), along with unmanned aerial vehicles and ground-based photogrammetry, researchers can capture the three-dimensional spatial structure of forests with unparalleled detail and accuracy. PCR serves as an indispensable step in the amalgamation of these technologies. Within the forest environment, PCR has progressed through three primary stages of development: initially relying on artificial markers, transitioning to the prevalence of feature-based methods in recent years, and currently moving towards the modern trend of development based on deep learning methods.
Methods based on artificial markers
-
In the early days, PCR commonly involved positioning artificial markers, such as reflector balls or artificial reflectors, within sample plots. Registration was then conducted through manual identification of these markers or by utilizing software[6−8]. Hilker et al.[9] introduced four connections within the sample plots to improve the precision and reliability of registration. Subsequently, Pueschel[10] utilized four FARO laser scanner reference spheres and one planar target for manual registration, achieving sub-millimeter accuracy. To simplify the process of using artificial markers, Zhang et al.[11] applied the back sighting orientation approach, using only one reflector as a connection point between two scans. During this stage, researchers attempted to enhance the registration precision by improving the placement and detection methods of manual markers. While offering advantages such as high accuracy and strong reliability, the process of setting up markers requires considerable time and effort, and user intervention during marker identification may be necessary[12], thus limiting their practicality and efficiency.
Methods based on feature matching
-
These methods identify and extract key features from the forest point cloud data. Subsequently, the required transformation parameters are estimated by registering the identified features, ensuring the uniform registration of individual point clouds into a common frame of reference. Due to the irregularity of forest point clouds and the frequent self-similarity in their surrounding regions, registration methods characterized by geometric elements (points, lines, and surfaces) are ineffective in forests. In forest scenes, tree stems are typically considered the most stable structures. Hence, some researchers have chosen tree location[13,14], the correlation between tree height and diameter at breast height[15], stem position or stem curvature[16−18], digital terrain models[19], and canopy height models[20] as distinctive feature points for the registration process. At the same time, researchers employed specific descriptors for a more sophisticated feature matching approach. Descriptors such as Fast Point Feature Histogram (FPFH)[21] and YOHO-Desc[22] capture local shape information in point cloud data, aiding in the registration process. Furthermore, researchers have explored the creation of registration primitives using various feature combinations, including tree locations and inter-tree distances[23], key points and stem locations extracted from tree crowns[24], or shaded regions identified in the original point cloud from a single scan[25]. Most contemporary forest PCR methods are divided into two separate steps, coarse and fine registration. Coarse registration is achieved through feature matching to obtain a better initial pose, while fine registration further refines the registration error using algorithms such as Iterative Closest Point (ICP). Feature-based methods can automatically detect and match feature points, making them widely applicable. However, they are sensitive to point cloud quality and environmental changes, often requiring data preprocessing, and parameter tuning for these algorithms may be cumbersome. Additionally, while feature-matching methods can handle some level of noise and local deformation, they may encounter issues such as mismatches and omissions.
Methods based on deep learning
-
With breakthrough advancements across various domains in recent years, deep learning technology has gradually been introduced into forest PCR research. The PointNet and PointNet++ networks provide a robust foundation for deep feature extraction and classification of point cloud data[26]. Leveraging these advanced network architectures, researchers can more accurately extract key features from forest point cloud data. Wang et al.[22] achieved a registration success rate of 99.8% by designing a learning-based sparse graph construction and a pose graph solving strategy based on historical reweighting. Although learning-based methods can automatically learn features and handle complex large-scale data, they require a large amount of annotated data for model training and have high computational complexity, potentially facing overfitting issues.
As PCR technology is increasingly applied in forest resource monitoring, the field of forest PCR is undergoing a transition from manual marker methods to automation and artificial intelligence. Each method has its advantages and limitations, as depicted in Table 1, and the choice should be based on specific circumstances. In complex forest environments and in research involving point cloud segmentation, classification, etc., manual marker methods remain indispensable. In relatively simple or moderately complex forest environments, feature-based matching methods have matured, but there are differences in feature point extraction and matching. Learning-based methods are still in their early stages, but their potential is promising. With ongoing technological advancements, it is anticipated that future forest PCR will become more efficient, accurate, and intelligent.
Table 1. Advantages and disadvantages of different PCR methods.
Method Advantages Disadvantages Artificial markers 1. High accuracy 1. Placement of markers is limited, requiring uniform distribution 2. Strong reliability 2. Setting up markers is time-consuming and labor-intensive 3. Strong controllability 3. User intervention might be needed during marker identification Feature matching 1. Automatically detects feature points 1. Sensitive to point cloud quality and environmental changes 2. Broad applicability 2. Typically requires preprocessing of point cloud data 3. Handle noise and local distortion 3. Parameters of algorithms need experimentation and adjustment 4. Integrate with other technologies 4. May encounter mismatches and omissions Deep learning 1. Automatically learns features 1. Requires a large amount of annotated data for model training 2. handle complex, large-scale data 2. High computational complexity, may face overfitting issues -
Forest PCR is mostly based on closed datasets curated by various research institutions and data providers. These datasets are typically sourced from forest point cloud data obtained through different platforms. To facilitate a reliable assessment and comparison of forest PCR methods, there is a pressing need for open benchmark datasets featuring a substantial data volume and diversity. Unfortunately, the currently available datasets are limited in size.
Evaluation criteria
-
To impartially and effectively assess the performance of forest PCR methods, a unified evaluation framework must be established, incorporating both quantitative and qualitative analyses.
Qualitative metrics
-
Qualitative metrics play a crucial role in assessing the performance of forest PCR methods by providing intuitive detailed close-up diagrams of the registration outcomes[25,31,43]. In practical applications, comparing cross-sectional plots, as well as the smoothness of curves before and after registration, effectively demonstrates the continuity and consistency achieved through registration.
Quantitative metrics
-
Quantitative metrics are pivotal for evaluating the performance of forest PCR methods. These metrics can be categorized into three main types: feature extraction and matching, rigid transform estimation, and runtime. Feature extraction-based metrics include Feature-Match Recall (FMR) and Registration Recall (RR)[32]. Metrics based on rigid transform estimation typically include Mean Absolute Error (MAE)[36], Mean Square Error (MSE), and Root Mean Square Error (RMSE)[31,48,55], focusing on the rotation matrix and translation vectors in the rigid transform. These metrics evaluate the precision and accuracy of estimating the rigid transform post-registration. Finally, runtime reflects the computational efficiency and speed of the algorithm[35,43].
(1) FMR
The FMR is the ratio of correct point pairs extracted from features to all the point pair data in the point cloud.
(2) RR
The RR is calculated by assessing how well a registration method recovers point clouds with overlapping regions from two sets of point clouds, both having rigid transformations and overlapping parts.
(3) MAE
The MAE is the average of the absolute errors between the true and predicted values in the rigid transformation. It can be expressed using formula (1).
$ MAE=\dfrac{1}{n}\sum _{i=1}^{n}\left|{y}_{i}-\overline{{y}_{i}}\right| $ (1) where
is the predicted value of the registration method, and$ \overline{{y}_{i}} $ is the true value of the rigid transformation.$ {y}_{i} $ (4) MSE
The MSE is the average of the squared differences between the true and predicted values in the rigid transformation. It can be expressed using formula (2).
$ MS E=\dfrac{1}{n}\sum _{i=1}^{n}{\left({y}_{i}-\overline{{y}_{i}}\right)}^{2} $ (2) where
is the predicted value of the registration method, and$ \overline{{y}_{i}} $ is the true value of the rigid transformation.$ {y}_{i} $ (5) RMSE
The RMSE is the average square root error between the true and predicted values in the rigid transformation. It can be expressed using formula (3).
$ RMS E=\sqrt{\dfrac{1}{n}\sum _{i=1}^{n}{\left({y}_{i}-\overline{{y}_{i}}\right)}^{2}} $ (3) where
is the predicted value of the registration method, and$ \overline{{y}_{i}} $ is the true value of the rigid transformation.$ {y}_{i} $ -
All data generated or analyzed during this study are included in this published article.
-
About this article
Cite this article
Liu J, Guo Y, Yang J, Zhu N, Dai W, et al. 2024. Forest point cloud registration: a review. Forestry Research 4: e018 doi: 10.48130/forres-0024-0015
Forest point cloud registration: a review
- Received: 21 December 2023
- Revised: 01 April 2024
- Accepted: 15 April 2024
- Published online: 08 May 2024
Abstract: Point cloud registration is a necessary prerequisite for conducting precise, large-scale forest surveys and management. This paper focuses on providing a systematic overview and summary of the work on forest point cloud registration over the past 20 years. The developmental process of forest point cloud registration methods, spanning from the early reliance on manual markers to the subsequent evolution towards automatic registration based on feature matching, and then to the advanced technology based on deep learning were reviewed. Furthermore, the paper offered detailed discussions on the registration between different point cloud platforms: ground platforms, between ground platforms and aerial platforms, and between aerial platforms. Additionally, the paper delved into mainstream datasets and evaluation metrics in the domain of forest point cloud registration. Finally, the paper summarized the current state of research in this area, highlighted challenges, and provided future research outlooks. This review aims to provide researchers with a comprehensive understanding of forest point cloud registration, and to promote the advancement of point cloud technology, hopefully inspiring further applications in the field.
-
Key words:
- Forestry /
- Point cloud /
- Registration /
- Feature matching /
- Multi-source platforms