A robust method for approximate visual robot localization in feature-sparse sewer pipes.

Area of Science:

• Robotics
• Computer Vision
• Geospatial Analysis

Background:

• Buried sewer pipe networks pose significant challenges for robot localization due to the absence of GPS signals and sparse visual features.
• Standard visual odometry algorithms often fail in these environments, requiring non-standard solutions.

Purpose of the Study:

• To develop a robust and accurate robot localization method for sewer pipe inspection.
• To address the limitations of existing localization techniques in feature-poor, GPS-denied environments.

Main Methods:

• A novel robot localization method based on detecting equally spaced pipe joints using a bag-of-keypoints algorithm.
• Training the visual categorization algorithm offline via unsupervised learning on sewer pipe joint images.
• Augmenting joint detection with drift correction through vision-based manhole recognition.

Main Results:

• The proposed method demonstrated superior robustness and accuracy compared to ORB-SLAM3 in feature-sparse sewer pipes.
• Achieved a mean absolute error of approximately 2%-4% localization accuracy across tested pipes.
• ORB-SLAM3 failed on one pipe and showed 12%-20% error on others, frequently losing track.

Conclusions:

• The developed pipe joint detection and manhole recognition method provides a more reliable localization solution for sewer robots.
• Highlights a trade-off between high-precision but fragile visual odometry and simpler, robust 1-DOF localization methods.