Comprehensive Analysis of System Calibration between Optical Camera and Range Finder

Kang Hyeok Choi, Changjae Kim, Yongil KimISPRS Int. J. Geo Inf 2018 · DOI

TL;DR

This paper presents a comprehensive analysis of system calibration between optical cameras and range finders, crucial for accurate 3D data fusion. They propose guidelines for efficient calibration by comparing one-step and two-step methods, and evaluating the impact of dataset configurations, finding that the one-step method with multiple datasets yields superior results.

What Problem Does This Paper Address?

Creating accurate 3D models is essential for a wide range of applications, from virtual environments to structural health monitoring. While single-sensor systems (like multiple cameras) are common, multi-modal systems combining different sensor types, such as optical cameras and range finders, offer advantages in data quality, especially in areas with low texture or depth discontinuities. However, effectively fusing data from these disparate sensors requires precise calibration to align their spatial and temporal information. This paper addresses the challenge of achieving accurate and efficient system calibration for optical camera and range finder pairs, aiming to provide guidelines for optimal data fusion.

Methods

The study employs a two-phase calibration process:

  • Self-Calibration: This initial step determines the intrinsic parameters (Interior Orientation Parameters or IOPs) of each individual sensor. For the optical camera, standard least-squares estimation (bundle block adjustment) was used. For the range finder, a one-step integrated method, found to be more accurate in prior research, was adopted. This ensures each sensor's unique characteristics are accounted for.
  • System Calibration: This phase establishes the geometric relationship (Relative Orientation Parameters or ROPs) between the optical camera and the range finder. Two primary approaches were investigated:
  • One-Step System Calibration: This method simultaneously estimates the exterior orientation parameters (EOPs) of the reference sensor (optical camera) and the ROPs between sensors.
  • Two-Step System Calibration: This approach first determines the EOPs for each sensor independently and then sequentially calculates the ROPs between them, often by averaging multiple ROP sets.

The researchers also systematically varied the data acquisition configuration, specifically the number and geometric arrangement (single vs. multiple columns) of datasets used for calibration, to assess their impact on accuracy.

Datasets

The experiments utilized data acquired from a custom-built optical-range-finder system.

  • Self-Calibration Datasets: Fifteen datasets were collected, including optical camera images, range-finder intensity images, and range data. Three of these datasets were specifically acquired with the system vertically erected to minimize correlations between IOPs and EOPs.
  • System Calibration Datasets: Eleven datasets were acquired for system calibration. These were then combined in various ways to create 46 unique combinations, exploring different numbers of datasets (from 2 to 11) and locations (single vs. multiple columns).
  • Data Fusion Evaluation Datasets: Three datasets were used for evaluating the accuracy of the fused RGB-D data. These included scenarios with various objects (boxes, cylinders) and a corridor environment to test performance on indoor elements like doors and windows.

Key Insights and Ideas

  • One-Step Wins: The one-step system calibration method consistently outperformed the two-step method in terms of both matching rate (data fusion accuracy) and parameter precision (standard deviations of ROPs). This suggests that a simultaneous estimation approach is more robust for this sensor combination.
  • Convergence is Key: Using datasets acquired from a "converging configuration" (multiple columns) significantly improved calibration accuracy and stability compared to using datasets from a single column. This highlights the importance of diverse viewpoints for robust calibration.
  • More Data, Better Precision: For the one-step method, increasing the number of datasets used for calibration directly led to improved precision (lower standard deviations of ROPs). This effect became more pronounced with four or more datasets.
  • Four is the Magic Number (for stability): Using at least four datasets, especially from converging configurations, was found to be crucial for achieving stable and accurate results, both in terms of matching rates and ROP precision, for both one-step and two-step approaches.
  • Best Performance: The highest matching rate (99.08%) was achieved using the one-step system calibration method with six datasets acquired from multiple columns.

Key Results

  • Self-Calibration Accuracy: The RMS values of residual errors for the optical camera and range finder were less than 0.8 pixels, indicating high precision in the estimated IOPs. The range finder's RMS error of 6.84 mm was also better than its nominal specification and previous studies.
  • System Calibration Performance:
    • The one-step approach generally yielded higher matching rates and more precise ROPs than the two-step approach.
  • The best result (99.08% matching rate) was achieved with the one-step method using six datasets from multiple columns.
  • Using multiple columns for data acquisition led to higher mean matching rates and lower standard deviations compared to single-column configurations.
  • Increasing the number of datasets (especially beyond four) generally reduced the standard deviations of ROPs for the one-step method.
  • Data Fusion Quality: Visual inspection of data fusion results showed fewer mismatching points (indicated by white regions) for cases with higher matching rates, confirming the quantitative findings.

Limitations and Future Work

  • Range Data Challenges: The authors acknowledge that future work needs to address inherent limitations of range data, such as low resolution, noise, multi-path effects, ambiguity range, and scattering problems.
  • Dataset Configuration Impact: While the study analyzed the impact of dataset configurations, further exploration into optimal dataset selection strategies could be beneficial.
  • Generalizability: The findings are based on a specific optical-range-finder system and testbed. Further validation with different sensor combinations and environments would enhance generalizability.
  • Application-Specific Optimization: While the methodology is presented as applicable to various fields, future work could focus on optimizing calibration strategies for specific application requirements.

Summary updated

BibTeX
@article{choi2018comprehensive,
  author = {Kang Hyeok Choi and Changjae Kim and Yongil Kim},
  title = {{Comprehensive Analysis of System Calibration between Optical Camera and Range Finder}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2018},
  doi = {10.3390/ijgi7050188},
}

Loading related content...

Figure