Target-based automated matching of multiple terrestrial laser scans for complex forest scenes
Introduction
Terrestrial laser scanning (TLS) has become an increasingly practical option for forest field inventories over the past 20 years (Liang et al., 2018, Calders et al., 2020). The major advantage of applying TLS to forest inventories lies in the accurate, rapid, and automatic generation of point cloud models that offer millimeter-level detail without being destructive (Wilkes and Lau et al., 2017). TLS is often exploited to record detailed horizontal and vertical forest structures at the plot level (Bienert and Maas, 2009) for stem mapping (Kelbe and Van Aardt et al., 2016). Due to the limited field of view and measurement range of terrestrial scanners, and the natural occlusions that limit the visibility of distant tree stems, a single scan does not offer sufficient plot-wise stem mapping for TLS-based forest inventories (Liu and Liang et al., 2017). Therefore, multiple scans from different locations are required, with the point clouds acquired by those scans subsequently registered so that the multiple scans complement each other to provide a complete representation of the 3D scene. Several registration methods have been developed for matching multiple scans (Dong and Liang et al., 2020), but there has been little forest-oriented research, with most techniques focusing on a single objective (Xu and Boerner et al., 2019) or being urban-oriented (Ge and Hu, 2020).
Registering TLS point clouds in forest scenes is very challenging for two main reasons. First, there is no adequate overlapping that covers the same sides of tree stems from different standpoints without a large number of scans (Liang and Wang et al., 2015). Second, most tree stems have similar geometries and similar positional relationships (Al-Durgham and Habib, 2014), thereby reducing the distinctiveness of the feature descriptors. Additionally, most of the key points are related to the characteristics of individual stems and leaves, making the processing time-consuming and vulnerable to errors.
Despite these challenges, many studies have considered the registration of forest scans. Several methods are sensitive to the characteristics of individual trees (Liu and Liang et al., 2017), combining the geometric properties of tree objects [e.g., tree height, diameter at breast height (DBH) (Hauglin and Lien et al., 2014), and distance between trees (Polewski and Erickson et al., 2016)] with the constrained observation model (Bienert and Maas, 2009). For example, Liang and Hyyppä (2013) matched forest point clouds based on the tree stem locations. Kelbe and Van Aardt et al. (2016) developed a pairwise strategy based on matching tree position triplets using a geometric similarity metric to register TLS point clouds, and used DBH estimates to filter out false triplet pairs, thus accelerating the processing. Liu and Liang et al. (2017) improved the work of Liang and Hyyppä (2013) to recover the transformation in 3D space, although their approach also relates to the tree-trunk locations. Apart from time-consuming calculations of individual trees, one factor restricting the implementation of the above methods is that it is difficult to extract discriminative characteristics from complex forest scenes. For example, suitable characteristics are usually related to some geo-reference, e.g., the ground, but it is not trivial to extract TLS-based ground data. Fig. 1 shows a typical working environment in an extremely complex wild bamboo forest (Plot 2), in which it is difficult to generate an accurate TLS-based digital elevation model (DEM) because of the dense weeds. Dai and Yang et al. (2019) claimed to have developed a novel type of feature point for forest environments that made it possible to search for correspondences, but the similarity of the geometry introduces more challenges to the feature points in terms of their distinctiveness. Guan and Su et al. (2020) identified shaded areas from the raw point cloud of a single scan, rather than calculating other characteristics of individual trees, but this strategy is excessively reliant on the quality of the corresponding ground point cloud. It follows that the complexity of forest scenes is the bottleneck of target-less automated registration methods in handling forest scenarios.
In taking a forest inventory, investigators/surveyors often use artificial targets to register adjacent scans (Holopainen et al., 2014, Wilkes et al., 2017). A reliable and flexible solution is to use retro-reflective targets (RRTs); although cheap targets [e.g., black and white A4 paper (Ge, 2020)] are available, they are difficult to detect automatically and accurately in complex scenes. Targets can be spherical, cylindrical, or flat, and at least three should be installed in visible regions for matching purposes. The targets are usually installed on tripods and distributed in the scanning field, with the target distribution being highly dependent on the expertise and experience of the operator. Therefore, designing a specific distribution during pre-scanning tasks so as to ensure the target visibility and transformation robustness is labor-intensive and time-consuming. More troublingly, it is not trivial for investigators/surveyors to find RRTs in a complex forest scene, regardless of whether automated software-based detection or manual selection is used. The integrity of the targets and the quality of those points depend on the scanning distances, incident angles, and occlusions; therefore, it is inevitable that false positives, missing positives, and target center bias will occur in the target identification. Such mistakes result in false registration.
In this study, we present a target-based registration framework for the automated matching of multi-scans in complex forest scenes. The main contributions of the proposed method are as follows:
- (1)
It provides a target-based registration method for complex forest scenes without the need for any prior knowledge of those scans; moreover, the proposed approach is robust, accurate, and efficient;
- (2)
It is an automated approach that accurately identifies the RRTs and the corresponding centers. Moreover, the proposed strategy retrieves the relationships among the RRTs in different scans without any preset plan;
- (3)
A target-based geometric network is used to investigate the potential connectedness of scans before the matching process, with pose-graph validation implemented to prune incorrect connections so as to guarantee the geometric consistency of the scenarios.
The remainder of this paper is organized as follows. After introducing the study areas and materials in Sec. 2, we describe the proposed method in Sec. 3. We report the results of experiments and associated analysis in Sec. 4, and present our conclusions and outlook in Sec. 5.
Section snippets
Scan areas
The two scan areas considered in this study are in Jiangxi Province, China (see Fig. 1). The first is a 250 m × 135 m managed camphor forest (MCF) plot (Plot 1 in Fig. 1), and the second is a 100 m × 200 m wild bamboo forest (WBF) plot (Plot 2 in Fig. 1). For each plot, we further show the locations of each scan in the corresponding scanning scopes from a birds-eye view. The MCF plot is considered to be an easy target, and the WBF plot is considered as a hard target. All trees in the MCF plot
Automated target-based registration approach
The proposed framework is illustrated in Fig. 3.
Point cloud registration results
As a qualitative evaluation of the proposed approach, we demonstrate the overall and local enlarged registration results of two datasets in Fig. 9; for the enlarged demonstrations, we only use four details of the whole scene for brevity. First, remarkably, all scans converge satisfactorily without any human intervention based on the identified RRTs. Apart from having the correct geometric consistency, the matched results in both datasets also exhibit outstanding registration accuracy, i.e., the
Conclusion
We have proposed an automatic target-based approach for unordered multi-view registration in complex forest scenes. Although target-less approaches have received considerable attention, target-based methods provide the best solution for realistic projects in complex scenes because of their stability and efficiency. The proposed approach first identifies RRTs from their intensity values and the model constraints. All scans are then associated based on the regular geometric constraints, and an
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgement
This study was jointly supported by the National Natural Science Foundation of China, under Projects 42071437 and 62006199, and the Fundamental Research Funds for the Central Universities, under Project 2682021CX070.
References (34)
- et al.
Practical optimal registration of terrestrial LiDAR scan pairs
ISPRS J. Photogramm. Remote Sens.
(2019) - et al.
Terrestrial laser scanning in forest ecology: Expanding the horizon
Remote Sens. Environ.
(2020) - et al.
Automated fusion of forest airborne and terrestrial point clouds through canopy density analysis
ISPRS J. Photogramm. Remote Sens.
(2019) - et al.
Registration of large-scale terrestrial laser scanner point clouds: A review and benchmark
ISPRS J. Photogramm. Remote Sens.
(2020) Automatic markerless registration of point clouds with semantic-keypoint-based 4-points congruent sets
ISPRS J. Photogramm. Remote Sens.
(2017)- et al.
Object-based incremental registration of terrestrial point clouds in an urban environment
ISPRS J. Photogramm. Remote Sens.
(2020) - et al.
A marker-free method for registering multi-scan terrestrial laser scanning data in forest environments
ISPRS J. Photogramm. Remote Sens.
(2020) - et al.
International benchmarking of terrestrial laser scanning approaches for forest inventories
ISPRS J. Photogramm. Remote Sens.
(2018) - et al.
Automated matching of multiple terrestrial laser scans for stem mapping without the use of artificial references
Int. J. Appl. Earth Obs. Geoinf.
(2017) - et al.
Keypoint-based 4-Points Congruent Sets-Automated marker-less registration of laser scans
ISPRS J. Photogramm. Remote Sens.
(2014)
Data acquisition considerations for terrestrial laser scanning of forest plots
Remote Sens. Environ.
Pairwise coarse registration of point clouds in urban scenes using voxel-based 4-planes congruent sets
ISPRS J. Photogramm. Remote Sens.
Association-Matrix-Based Sample Consensus Approach for Automated Registration of Terrestrial Laser Scans Using Linear Features
Photogramm. Eng. Remote Sens.
Methods for the automatic geometric registration of terrestrial laser scanner point clouds in forest stands
ISPRS Int. Arch. Photogramm. Rem. Sens. Spat. Inf. Sci.
Evaluation of the Range Accuracy and the Radiometric Calibration of Multiple Terrestrial Laser Scanning Instruments for Data Interoperability
IEEE Trans. Geosci. Remote Sens.
Configuration Requirements for Panoramic Terrestrial Laser Scanner Calibration Within a Point Field
IEEE Geosci. Remote Sens. Lett.
Cited by (8)
Semantic maps for cross-view relocalization of terrestrial to UAV point clouds
2022, International Journal of Applied Earth Observation and GeoinformationCitation Excerpt :As it is known that the resolution decreases with the elevation of scanning distance, we only used the point clouds within 20 m in our tasks related to the property of our scanner. Based on a cylinder fitting algorithm, the program automatically finds the tree trunks in the environment (Ge and Zhu, 2021). Fig. 9 shows the results of trees (the red points) detected and instantiation from the point clouds.
Meta-learning an intermediate representation for few-shot prediction of landslide susceptibility in large areas
2022, International Journal of Applied Earth Observation and GeoinformationLearning Neural Radiance Fields of Forest Structure for Scalable and Fine Monitoring
2024, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)