TL;DR
This paper introduces a novel, fully automatic method for registering terrestrial laser scanning (TLS) point clouds using panoramic reflectance images. By leveraging inherent pixel-to-point correspondence in these images and employing an iterative refinement process for pair-wise registration, followed by a circular self-closure constraint for global registration, the approach achieves millimetre-level accuracy with significantly reduced processing time compared to traditional methods.
What Problem Does This Paper Address?
Terrestrial Laser Scanning (TLS) is a powerful tool for creating detailed 3D models of environments, but a critical challenge lies in "registration" – aligning multiple scans taken from different viewpoints into a single, coherent point cloud. Traditional methods often rely on artificial markers, manual intervention, or computationally intensive algorithms like Iterative Closest Point (ICP). These methods can be time-consuming, prone to errors, and may struggle with complex scenes. The paper addresses the need for a fully automatic, efficient, and accurate registration process that can handle the large datasets generated by modern TLS scanners.
Methods
The proposed approach is a two-stage process:
- Pair-wise Registration:
- Reflectance Image Generation: Panoramic reflectance images are generated directly from TLS scans, utilizing the angular coordinates and reflectance values of each 3D point. This inherently provides pixel-to-point correspondence.
- Image Matching (Pixel-to-Pixel Correspondence): The Scale-Invariant Feature Transform (SIFT) algorithm is used to extract distinctive features from these panoramic images and find correspondences between adjacent scans. SIFT is chosen for its robustness to changes in illumination, scale, and viewpoint.
- Point Cloud Registration (Point-to-Point Correspondence): Using the established pixel-to-point correspondence, 3D corresponding points are identified.
- Outlier Detection: A rigid geometric invariance check, based on the Euclidean distance between corresponding 3D points, is employed to prune false matches. This check is made self-adaptive by considering the scanner's accuracy.
- Iterative Refinement: The pixel-to-point correspondence and the computation of Rigid Transformation Parameters (RTPs) are integrated into an iterative process. This loop refines matches, removes outliers, corrects false correspondences, and ensures a better distribution of points, leading to optimized RTPs.
- Global Registration:
- Circular Self-Closure Constraint: For registering multiple scans (forming a "ring" scheme), a bundle adjustment approach is used. Instead of a simple last-to-first closure, a circular self-closure constraint is implemented. This means the entire ring of scans is adjusted simultaneously to minimize errors, leading to a more robust global optimization.
Datasets
The proposed method was tested on four datasets acquired using a FARO LS 880 laser scanner:
- Dataset 1: 24 million TLS points, office environment.
- Dataset 2: 33 million TLS points, Aula Conference Centre.
- Dataset 3: 20 million TLS points, Aula Conference Centre.
- Dataset 4: 26 million TLS points, TU Delft campus building.
These datasets represent both indoor and outdoor scenes with varying resolutions. The authors also used Leica HDS Cyclone 5.5 software for comparison.
Key Insights and Ideas
- Leveraging Inherent Correspondence: The core innovation is using panoramic reflectance images, which naturally provide pixel-to-point correspondence. This bypasses the need for separate camera calibration and camera-to-scanner registration, which are common sources of error in other image-based methods.
- Iterative Optimization for Pair-wise Registration: Integrating image matching and 3D point cloud registration into an iterative loop significantly improves the quality of initial matches, leading to more accurate RTPs. This iterative refinement is key to handling challenging correspondences.
- Circular Self-Closure for Global Registration: Moving beyond simple last-to-first closure constraints in bundle adjustment for global registration, the circular self-closure constraint allows for simultaneous optimization across all scans in a ring, leading to a more globally optimized and accurate result.
- Automation and Efficiency: The method is designed to be fully automatic, eliminating manual intervention and significantly reducing processing time compared to traditional ICP-based methods.
Key Results
- Accuracy: Pair-wise registration achieved accuracies on the order of millimetres (e.g., RMS error of 0.0037 m for Dataset 1). Global registration also resulted in centimetre-scale accuracy.
- Speed: The proposed method processed scans much faster than commercial ICP software. For example, Dataset 1 registration took 5 minutes, while Cyclone 5.5 took 21 minutes.
- Robustness: The SIFT algorithm and the iterative refinement process demonstrated robustness in handling repetitive patterns and challenging geometric configurations.
- Automation: The entire process, from feature extraction to global registration, is fully automatic.
Limitations and Future Work
- Limitations:
- The paper notes that while the SIFT threshold of 0.8 was chosen to eliminate 90% of false matches, it was found to be "loose" in some cases, accepting incorrect matches due to repetitive patterns in building facades. This was addressed by the subsequent rigid geometric invariance check.
- The performance of the ICP comparison was affected by the sensitivity to initial corresponding point pairs, leading to failures in some cases.
- Future Work:
- Investigate improving registration accuracy and the distance invariance error by considering laser scanner accuracy factors like incident angle and edge effects.
- Extend the proposed method to an integrated change detection/registration approach for emergency response applications.
Summary updated