Development of Data Registration and Fusion Methods for Measurement of Ultra-Precision Freeform Surfaces

Ling Bao Kong, Mingjun Ren, Min XuSensors 2017 · DOI

TL;DR

This paper introduces a novel approach for accurately measuring complex, ultra-precision freeform surfaces by combining data from multiple sensors. It achieves this by first registering datasets using invariant surface features and then fusing overlapping data with a Gaussian process model that accounts for measurement uncertainties, leading to improved accuracy and efficiency.

What Problem Does This Paper Address?

Measuring ultra-precision freeform surfaces, which are crucial for advanced optical and mechatronic applications, presents significant challenges. These surfaces are geometrically complex, often large, and may have features at multiple scales. No single sensor can efficiently capture all necessary information with high accuracy. Therefore, multi-sensor systems are employed, but their effectiveness hinges on accurate data registration (aligning data from different sensors) and data fusion (combining overlapping data). Existing methods struggle with the complexity and noise inherent in these measurements, leading to accumulated errors and reduced precision, especially for sub-micrometer accuracy requirements.

Methods

The paper proposes a two-stage approach:

  1. Data Registration using Intrinsic Surface Features (ISFs):
    • ISF Representation: Instead of directly using 3D coordinates, the method represents surface geometry using ISFs like Gaussian and mean curvature. These features are invariant to coordinate transformations and surface parameterization, making them robust.
  2. 2D Feature Pattern: The 3D ISF data is converted into a 2D "intrinsic feature pattern." This is achieved by re-sampling the data onto a grid and arranging the ISF values, effectively creating a texture-like representation. A woven mesh model is used to optimize point distribution and minimize distortion when fitting this 2D pattern to the freeform surface.
  3. 2D Registration: The registration problem is then transformed from a 3D alignment to a 2D alignment of these feature patterns. This is accomplished using a Fourier-Mellin transform and phase correlation method, which efficiently recovers translation and rotation in 2D. This significantly simplifies the registration process compared to 3D methods.
  4. Coordinate Transformation: Once correspondences are established in 2D, they are used to estimate a 3D coordinate transformation matrix (including translation and rotation) using the Levenberg-Marquardt algorithm.
  5. Data Fusion using Gaussian Process (GP) Model:
    • Addressing Uncertainty: After registration, overlapping areas from different datasets are fused. The key innovation here is the use of a Gaussian Process (GP) model, which explicitly accounts for the varying uncertainties associated with data from different sensors or measurement conditions.
  6. Bayesian Approach: GP is a Bayesian regression model that provides not only a fused surface prediction but also an estimate of its uncertainty. This is crucial for understanding the reliability of the fused data.
  7. Model Formulation: A zero-offset mean function is used, and a squared exponential covariance function is employed. The model is extended to handle multiple datasets with potentially different noise variances.

Datasets

  • Simulated Datasets: The method was first validated using simulated datasets. A defined "normal surface" was used, from which two sub-surfaces were extracted at different locations, orientations, and with added Gaussian noise of varying standard deviations (0.2 µm and 0.5 µm) to mimic real-world measurement errors and different sensor resolutions.
  • Actual Measurement Datasets:
    • A machined micro-structured surface was measured using a 3D optical profiler with two different objectives (5.5× and 20×), resulting in datasets with different resolutions and coordinate frames.
  • A general freeform surface was measured using a combination of a high-precision Coordinate Measuring Machine (CMM) and a laser scanner. The CMM provided sparse, high-accuracy data, while the laser scanner provided dense, lower-accuracy data.

Key Insights and Ideas

  • ISFs as Robust Features: Using invariant surface features (like curvature) as the basis for registration is a powerful idea. It decouples the registration process from the specific coordinate system or parameterization of the surface, making it more robust to noise and variations.
  • 2D Registration Simplification: Transforming the 3D registration problem into a 2D feature pattern registration significantly simplifies the computational complexity and improves efficiency.
  • GP for Uncertainty-Aware Fusion: Explicitly incorporating measurement uncertainty into the data fusion process using Gaussian Processes is critical for achieving high-accuracy results, especially when combining data from sensors with different error characteristics. This allows for a more principled way to weigh information from different sources.
  • Synergy of Multi-Sensor Systems: The paper demonstrates that by effectively registering and fusing data from sensors with complementary strengths (e.g., high accuracy vs. high speed/range), the overall measurement performance can surpass that of individual sensors.

Key Results

  • Registration Accuracy: The proposed ISF-based registration method showed comparable or slightly better performance (lower variance) than the Iterative Closest Point (ICP) method in simulated experiments, particularly in recovering six spatial parameters (rotations and translations).
  • Fusion Accuracy: In simulation, the fused GP model achieved significantly lower form error (RMS error of 2 nm, PV of 24 nm) compared to an existing ICP + Weighted Mean (WM) fusion method (RMS 7 nm, PV 78 nm).
  • Actual Measurement Performance:
    • For the micro-structured surface, the RMS error of the fused dataset (0.087 µm) closely matched the errors from individual high-resolution measurements (0.011 µm and 0.097 µm), indicating accurate registration and fusion.
  • For the CMM and laser scanner combined measurement, the hybrid approach achieved a good balance of accuracy (PV 33.2 µm, RMS 4.8 µm) and efficiency (time ~0.5 h), outperforming individual sensor measurements in terms of accuracy while maintaining reasonable speed.

Limitations and Future Work

  • Computational Cost of GP: While GP provides excellent results, it can be computationally intensive, especially for very large datasets. The paper doesn't explicitly detail the computational burden for the actual measurements.
  • Noise Sensitivity of ISF Calculation: The paper mentions that direct calculation of ISFs can be sensitive to noise, which is why a B-spline fitting is used. However, extreme noise levels might still pose challenges.
  • Future Work: The paper suggests that the proposed method is beneficial for multi-scale and multi-sensor measurement of ultra-precision freeform surfaces. Further applications and optimizations for even more complex scenarios could be explored. The authors also highlight the need to address key issues in multi-sensor measurement, implying ongoing research in this area.

Summary updated

Abstract

The measurement of ultra-precision freeform surfaces commonly requires several datasets from different sensors to realize holistic measurements with high efficiency. The effectiveness of the technology heavily depends on the quality of the data registration and fusion in the measurement process. This paper presents methods and algorithms to address these issues. An intrinsic feature pattern is proposed to represent the geometry of the measured datasets so that the registration of the datasets in 3D space is casted as a feature pattern registration problem in a 2D plane. The accuracy of the overlapping area is further improved by developing a Gaussian process based data fusion method with full consideration of the associated uncertainties in the measured datasets. Experimental studies are undertaken to examine the effectiveness of the proposed method. The study should contribute to the high precision and efficient measurement of ultra-precision freeform surfaces on multi-sensor systems.
BibTeX
@inproceedings{kong2017development,
  author = {Ling Bao Kong and Mingjun Ren and Min Xu},
  title = {{Development of Data Registration and Fusion Methods for Measurement of Ultra-Precision Freeform Surfaces}},
  booktitle = {Sensors},
  year = {2017},
  doi = {10.3390/s17051110},
}
Figure