A Depth-Based Weighted Point Cloud Registration for Indoor Scene

Shuntao Liu, Dedong Gao, Peng Wang, Xifeng Guo, Jing Xu, Du-Xin Liu2018 · DOI

TL;DR

This paper introduces an improved Iterative Closest Point (ICP) algorithm for 3D point cloud registration that enhances accuracy and speed by prioritizing point pairs with smaller Euclidean distances and incorporating depth measurement errors into a weighted approach. This method is particularly beneficial for indoor scene reconstruction where depth accuracy can vary.

What Problem Does This Paper Address?

Creating a complete 3D model of a scene often requires capturing multiple point clouds from different viewpoints. These clouds are initially in separate coordinate systems, necessitating a process called "point cloud registration" to align them into a single, unified coordinate system. This registration is crucial for the overall quality of the 3D reconstruction. While the Iterative Closest Point (ICP) algorithm is a standard for this task, it has limitations: it can be sensitive to initial alignment, prone to getting stuck in local optima, and doesn't explicitly account for measurement errors inherent in depth sensors. This paper aims to address these issues by developing a more robust and accurate ICP variant that considers depth measurement errors and strategically selects point pairs for registration.

Methods

The core of this paper is an enhanced ICP algorithm. Here's a breakdown of the key modifications:

  • Weighted Point Pair Selection: Instead of using all point pairs for registration, the proposed method prioritizes point pairs with smaller Euclidean distances. This is based on the intuition that closer points are more likely to be true correspondences.
  • Depth Measurement Error Model: The authors analyze and model the depth measurement error of depth cameras, specifically using the Intel RealSense D415 as an example. They derive a formula showing that depth error is proportional to the square of the depth value.
  • Weighted Registration: This depth error model is then used to assign weights to point pairs. Pairs with smaller depth errors (typically at closer distances) are given higher weights, influencing the optimization process to favor these more reliable correspondences.
  • Iterative Refinement: The algorithm iteratively refines the rotation and translation matrices by minimizing a weighted error function, aiming to improve both accuracy and convergence speed.

Datasets

The experiments were conducted using point cloud data acquired by an Intel RealSense D415 depth camera in four different indoor scenes. The paper does not introduce new benchmark datasets but rather uses real-world indoor environments to evaluate the proposed method.

Key Insights and Ideas

  • Depth Error Matters: The paper strongly emphasizes that measurement errors from depth cameras are not negligible and significantly impact registration accuracy. Ignoring them can lead to accumulated errors in the final 3D model.
  • Closer Points are Better: Prioritizing point pairs with smaller Euclidean distances for registration is a simple yet effective heuristic for improving accuracy, as these pairs are more likely to represent true correspondences.
  • Weighting for Reliability: Assigning weights based on depth error allows the algorithm to "trust" more reliable measurements more, leading to a more robust registration process, especially in scenes with varying depth.
  • Improved Convergence: The combination of weighted selection and error consideration not only improves accuracy but also leads to faster convergence of the registration error.

Key Results

The experimental results demonstrate that the proposed depth-based weighted ICP method (Method D) outperforms standard ICP (Method A) and weighted ICP without distance prioritization (Method B), as well as ICP with distance prioritization but without weighting (Method C). Specifically:

  • Improved Accuracy: Method D consistently achieved lower registration RMS errors across all tested indoor scenes compared to the other methods.
  • Faster Convergence: The registration error converged more rapidly in Method D, indicating a quicker and more efficient registration process.
  • Qualitative Improvement: Visual comparisons (Figure 5) show that Method D produces more accurate and visually coherent alignments of the point clouds.

Limitations and Future Work

  • Assumptions: The method assumes that the source and target point clouds are obtained from two different static postures and that there is a large overlap area between them.
  • Future Work: The authors suggest extending this work to 4D modeling, which would involve incorporating temporal information and environmental features to further enhance registration accuracy and speed.

Summary updated

BibTeX
@misc{liu2018depth,
  author = {Shuntao Liu and Dedong Gao and Peng Wang and Xifeng Guo and Jing Xu and Du-Xin Liu},
  title = {{A Depth-Based Weighted Point Cloud Registration for Indoor Scene}},
  year = {2018},
  doi = {10.3390/s18113608},
}
Figure