3D point cloud registration with shape constraint

Swapna Agarwal, Brojeshwar BhowmickICIP 2017 · DOI · OpenReview

TL;DR

This paper introduces a novel 3D point cloud registration algorithm that enhances the traditional "gravitational approach" by incorporating a shape-similarity constraint. This allows for more robust and accurate alignment of point clouds, even in the presence of significant noise, missing data, and large rotations, making it particularly useful for applications like change detection.

What Problem Does This Paper Address?

Registering 3D point clouds is a fundamental step in many applications, including scene understanding, industrial quality control, and robotics. This process involves aligning a "template" point cloud to a "reference" point cloud, often by estimating translation, rotation, and scale. However, existing methods like Iterative Closest Point (ICP) and even more advanced techniques like the Gravitational Approach (GA) struggle with real-world data that often contains noise, outliers, and missing parts. The GA method, while an improvement over some, can still lead to misalignment when dealing with large rotations or significant structural changes because it primarily focuses on aligning the center of mass rather than the local shapes. This paper aims to overcome these limitations by developing a registration algorithm that prioritizes shape matching.

Methods

The core of this paper is a modified Gravitational Approach (GA). The original GA algorithm models point cloud registration as a gravitational system where points in the reference cloud attract points in the template cloud. The force of attraction is inversely proportional to the squared distance between points.

The authors propose two key modifications:

  1. Shape-Based Attraction: Instead of just distance, the force of attraction is now influenced by a measure of local shape similarity between corresponding points in the reference and template clouds. This means that points with similar local geometry will attract each other more strongly, leading to better alignment of the actual shapes. They use a radial basis function (RBF) to quantify this similarity, incorporating curvature as a key feature.
  2. Distance-Dependent Force: To improve convergence, especially near local minima, the authors modify the force function to be proportional to a monotonically increasing function of distance, rather than inversely proportional. This ensures that the template takes smaller, finer steps when it's close to alignment, preventing overshooting and improving stability. They also introduce a modified gravitational constant G that increases near convergence.

Additionally, the paper addresses scale estimation by using an orientation and translation invariant model based on the spatial point distribution of the two point clouds, derived from the eigenvalues of their covariance matrices. The overall algorithm is designed to be fully automatic, requiring fewer manual parameters than previous GA methods.

Datasets

The authors evaluate their method on both synthetic and real-world data.

  • Synthetic Data: They use the "bunny" point cloud from the Stanford 3D Scanning Repository, introducing varying percentages of Gaussian and uniform noise, as well as structured outliers.
  • Real-World Data: They capture 3D point cloud scans of scenes using a Kinect sensor, with two captures per scene to simulate changes.

Key Insights and Ideas

  • Shape Matters More Than Mass: The most significant insight is that prioritizing local shape similarity over simple mass attraction (as in the original GA) is crucial for robust registration, especially when dealing with complex geometries or significant changes.
  • Gentle Convergence for Precision: Modifying the force to be distance-dependent, rather than distance-squared dependent, helps the algorithm converge more smoothly and accurately, avoiding oscillations around the optimal alignment.
  • Automatic Scale Estimation: Developing an invariant scale estimation method based on point distribution eigenvalues reduces the need for manual parameter tuning and improves robustness.
  • Change Detection Utility: The algorithm's ability to accurately align point clouds even with significant changes makes it a promising tool for change detection applications.

Key Results

The proposed method demonstrates superior performance compared to state-of-the-art methods like ICP, CPD, and the original GA, particularly in challenging scenarios:

  • Structured Outliers: The method consistently aligns the human shape with the reference, even when the outlier's mass is significantly larger than the reference, unlike GA which can be misled.
  • Missing Parts: The algorithm performs well even when 18% of points are missing from the template.
  • Rotation: It handles large rotations (over 50°) more effectively than GA.
  • Noise: It shows better performance with both Gaussian and uniform noise, outperforming other methods in terms of Root Mean Square Error (RMSE) across various noise percentages.
  • Real-World Scenarios: Qualitative results on real Kinect data show successful alignment despite object repositioning, demonstrating its practical applicability for change detection.

Limitations and Future Work

  • Limitations: The paper notes that the original GA method's performance degrades with increasing rotation (more than 45°) and change (extra or missing parts), which their modified approach aims to address. While their method shows improvement, the exact performance limits with extreme rotations or missing data are not exhaustively detailed.
  • Future Work: The authors plan to employ their proposed registration approach for efficient change detection. They also mention that for noisy real scenarios, a pyramidal approach (decreasing local neighborhood size and σ with iteration) can be beneficial.

Summary updated

Abstract

In this paper, a shape-constrained iterative algorithm is proposed to register a rigid template point-cloud to a given reference point-cloud. The algorithm embeds a shape-based similarity constraint into the principle of gravitation. The shape-constrained gravitation, as induced by the reference, controls the movement of the template such that at each iteration, the template better aligns with the reference in terms of shape. This constraint enables the alignment in difficult conditions indtroduced by change (presence of outliers and/or missing parts), translation, rotation and scaling. We discuss efficient implementation techniques with least manual intervention. The registration is shown to be useful for change detection in the 3D point-cloud. The algorithm is compared with three state-of-the-art registration approaches. The experiments are done on both synthetic and real-world data. The proposed algorithm is shown to perform better in the presence of big rotation, structured and unstructured outliers and missing data.
BibTeX
@inproceedings{agarwal2017d,
  author = {Swapna Agarwal and Brojeshwar Bhowmick},
  title = {{3D point cloud registration with shape constraint}},
  booktitle = {ICIP},
  year = {2017},
  doi = {10.1109/icip.2017.8296672},
}
Figure