TL;DR
This paper introduces an improved Progressive TIN Densification (PTD) filtering algorithm designed to better handle point clouds with high density and variance, common in Dense Image Matching (DIM) data. By dividing the filtering process into two stages with adaptive densification thresholds, the method significantly reduces type I errors (ground points misclassified as non-ground) while maintaining acceptable type II errors, offering a more robust solution for digital terrain model generation.
What Problem Does This Paper Address?
Accurately separating ground points from non-ground points in 3D point clouds is crucial for applications like digital terrain model (DTM) generation. The Progressive TIN Densification (PTD) algorithm is a widely used and effective method for this task. However, the standard PTD algorithm struggles with point clouds exhibiting high density and high standard variance, such as those derived from Dense Image Matching (DIM). In these challenging datasets, the PTD algorithm tends to produce a high rate of Type I errors (misclassifying ground points as non-ground), which can lead to inaccuracies in the generated DTM. Existing approaches to mitigate this, like downsampling, can lead to loss of topographic detail and difficulty in parameter selection. This paper aims to address these limitations by proposing an enhanced PTD algorithm that explicitly accounts for point cloud density and variance.
Methods
The core of the proposed method is a two-stage iterative densification process for the TIN (Triangulated Irregular Network).
- Stage 1: Classical PTD Densification: This initial stage employs the standard PTD algorithm. The TIN is progressively densified using decreasing iteration angle and distance thresholds. This stage continues until the density of points within the TIN reaches a predefined "standard variance intervention density" (DENthr).
- Stage 2: Multi-Scale Densification with Adaptive Thresholds: Once the DENthr is reached, the algorithm transitions to a new strategy. This stage uses a multi-scale approach where the survey region is divided into grids. A new TIN is constructed using resampled points. Crucially, the densification thresholds (angle and distance) are recalculated based on the points not included in the current TIN. In this stage, the angle threshold is increased, while the distance threshold is decreased. This adaptive strategy aims to better accommodate variations in point cloud density and variance, allowing more ground points to be correctly classified.
The DENthr itself is calculated based on the standard variance of the point clouds and a density coefficient, aiming to dynamically adjust the transition point between the two stages.
Datasets
The proposed method was evaluated using the Vaihingen dataset, which includes both airborne LiDAR (ALS) data and Dense Image Matching (DIM) point clouds. Two specific test areas from this dataset were used:
- Test Area 11: A slope area.
- Test Area 32: A flat area.
These areas provided both high-density, high-variance DIM data and lower-density, lower-variance ALS data, allowing for a comprehensive comparison of the algorithm's performance under different conditions. The paper does not introduce new benchmark datasets.
Key Insights and Ideas
- Density and Variance are Critical: The paper strongly emphasizes that high point cloud density and standard variance are significant factors that degrade the performance of traditional PTD filters, particularly increasing Type I errors.
- Two-Stage Filtering is Beneficial: Dividing the filtering process into two distinct stages, with adaptive threshold adjustments, is key to handling complex point clouds. The initial stage captures general terrain, while the second stage refines it by accounting for local variations.
- Adaptive Thresholds for Robustness: The novel approach of dynamically adjusting angle and distance thresholds in the second stage, based on the density and variance of the points, is crucial for improving accuracy, especially in challenging datasets.
- Type I vs. Type II Error Trade-off: The improved method effectively reduces Type I errors but may slightly increase Type II errors. However, the paper argues that these Type II errors have minimal impact on the quality of the generated Digital Surface Model (DSM).
Key Results
The experimental results demonstrate that the improved PTD algorithm significantly outperforms the classical PTD algorithm, especially on DIM point clouds (high density and variance):
- Type I Errors: Reduced by 7.53% on average for DIM point clouds.
- Total Errors: Reduced by 4.09% on average for DIM point clouds.
- Qualitative Improvement: Visual inspection of filtering results shows a cleaner separation of ground and non-ground points, with fewer misclassified ground points in the improved method.
- ALS Data Performance: The algorithm also performs well on ALS data, though the improvement is less pronounced due to the inherently lower density and variance of ALS data, causing the algorithm to primarily operate in its first stage.
Limitations and Future Work
- Density Coefficient Estimation: The paper acknowledges that estimating the density coefficient (k) in the DENthr calculation can be challenging, as it relies on assumptions about facet symmetry. This could lead to inaccuracies if not properly estimated.
- Type II Error Increase: While Type II errors are noted to increase slightly, the authors argue their impact on DSM generation is minimal. However, further work is suggested to reduce these errors.
- Future Work: The authors suggest future research will focus on improving the filter to further reduce Type II errors and developing a more accurate method for calculating DENthr.
Summary updated