Geometrical feature extraction using 2D range scanner

San Zhang, Lihua Xie, M. Adams, Fan TangThe Fourth International Conference on Control and Automation, 2003. ICCA. Final Program and Book of Abstracts 2005 · DOI

TL;DR

This paper introduces two novel algorithms for robustly detecting geometric features like lines and circles from 2D range scanner data, crucial for mobile robot navigation in unstructured environments. The methods offer improved accuracy and robustness compared to existing techniques, enabling more reliable environmental perception.

What Problem Does This Paper Address?

Reliable navigation for mobile robots hinges on accurately understanding their environment. A common approach is to represent the environment using geometric features extracted from sensor data. However, extracting these features, especially from noisy 2D range scanner data in complex, semi-structured outdoor environments (like parks or suburban roads), remains a significant challenge. Existing methods often struggle with accuracy, robustness to noise, and the ability to handle diverse feature types simultaneously. This paper aims to bridge this gap by proposing more accurate and robust algorithms for detecting both line and circle features.

Methods

The paper presents two main approaches:

  • Off-line Circle and Edge Detection (Gaussian-Newton): This algorithm first segments the scan data to identify potential features. It then uses an optimized Gaussian-Newton method to accurately fit circle parameters to detected circular features. This method is performed after a full scan is collected.
  • On-line Multi-Model Filtering: This approach tackles feature extraction in real-time. It employs a multiple-model filtering strategy, specifically using an Extended Kalman Filter (EKF) for line detection and an Unscented Kalman Filter (UKF) for circle tracking. This allows the system to adapt and switch between line and circle models as it processes incoming data, making it suitable for dynamic environments.

Datasets

The paper describes experiments conducted using data from a Sick PLS laser scanner. The scanner has a field of view of 180 degrees and a range of up to 50 meters. To achieve a 360-degree view, the robot performs two scans at each location. The experimental data was collected from an outdoor environment featuring pillars, building walls, and low balustrades. No specific named dataset is mentioned, but the experimental setup and data collection process are detailed.

Key Insights and Ideas

  • Robustness through Optimization: The Gaussian-Newton method, when applied to circle fitting, offers a significant improvement in accuracy and robustness compared to simpler batch calculation methods, especially in the presence of noisy data.
  • Adaptability with Multi-Model Filtering: The on-line multi-model approach is crucial for real-world applications. By allowing the system to switch between line and circle models, it can effectively track features that change geometry or appear/disappear.
  • Feature Importance in Outdoor Navigation: The paper highlights the utility of natural, geometric features like tree trunks and pillars for navigation in semi-structured outdoor environments, where artificial beacons are less practical.
  • Validation Gate for Discontinuity Detection: The use of a validation gate (based on the innovation and its variance in the EKF) is an effective way to detect significant discontinuities in the range data, signaling the end of one feature and the potential start of another.

Key Results

  • The proposed Gaussian-Newton algorithm demonstrated accurate detection of circle and edge features, with experimental results showing good parameter estimation (e.g., fitting a circle with a radius of 0.4957m from noisy data with an expected radius of 0.5m).
  • The on-line multi-model filtering approach was shown to be capable of extracting most features in the experimental environment, successfully identifying pillars and other geometric structures.
  • Both methods were found to be efficient in detecting features.

Limitations and Future Work

  • False Positives: The paper acknowledges that both algorithms can sometimes output features that do not physically exist. Minimizing these false positives is identified as an area for further investigation.
  • Circle Feature Visibility: In some experimental scenarios, due to the large range of the scanner, circle features could not be seen clearly, suggesting that the algorithms' performance might be affected by the scale and visibility of the features.
  • Further Studies Needed: The authors explicitly state that further studies are needed to address the issue of minimizing physically non-existent features.

Summary updated

BibTeX
@inproceedings{zhang2005geometrical,
  author = {San Zhang and Lihua Xie and M. Adams and Fan Tang},
  title = {{Geometrical feature extraction using 2D range scanner}},
  booktitle = {The Fourth International Conference on Control and Automation, 2003. ICCA. Final Program and Book of Abstracts},
  year = {2005},
  doi = {10.1109/icca.2003.1229234},
}
Figure