GPR-based Model Reconstruction System for Underground Utilities Using GPRNet

Yang Liang, Jinglun Feng, Liang Yang, Ejup Hoxha, Diar Sanakov, Stanislav Sotnikov, Jizhong Xiao

TL;DR

This paper introduces a novel robotic system for reconstructing detailed 3D models of underground utilities from sparse Ground Penetrating Radar (GPR) data. By combining an omnidirectional robot for data collection with a deep learning approach (GPRNet) for interpretation and reconstruction, they achieve dense point cloud models, significantly improving visualization and robustness against noise.

What Problem Does This Paper Address?

Detecting and locating underground utilities like pipes and cables is crucial for infrastructure maintenance and safety. Ground Penetrating Radar (GPR) is a powerful non-destructive tool for this, but existing methods face several challenges. Traditional GPR data interpretation often relies on manual analysis of B-scan images, which can be time-consuming and requires expert knowledge. Furthermore, reconstructing detailed 3D models from sparse GPR measurements has been difficult, often resulting in noisy or incomplete models. This paper aims to bridge this gap by developing an automated system that can collect GPR data efficiently and reconstruct high-fidelity 3D point cloud models of underground utilities, even from limited and noisy input.

Methods

The proposed system comprises three key modules:

  • Visual-Inertial-based GPR Data Collection: An omnidirectional robot equipped with Mecanum wheels allows for flexible movement (forward, backward, sideways) without rotation. This robot carries a GPR antenna and an Intel NUC with an RGB-D camera and IMU. The visual-inertial odometry system provides accurate pose information for each GPR measurement, crucial for later 3D reconstruction.
  • GPR Data Interpretation (MigrationNet): This module uses a deep neural network, MigrationNet, to interpret raw GPR B-scan images. Unlike traditional methods that focus on hyperbolic features, MigrationNet directly infers a migration result by interpreting the B-scan, aiming to produce a cross-section image of underground objects. It employs a Multiple Resolution Encoder (MRE) to capture features at different scales, which are then concatenated to form a global feature map. A decoder then predicts a binary image where white pixels represent utilities.
  • 3D Reconstruction (GPRNet): This is the core novel contribution. GPRNet takes the interpreted cross-section data (registered into 3D space using the collected pose information) as a sparse point cloud input. It then employs an encoder-decoder architecture to complete this sparse input into a dense and smooth 3D point cloud model. The encoder, inspired by PointNet, extracts multi-resolution features. The decoder is a hierarchical structure combining fully-connected and folding operations to generate a detailed point cloud, aiming for both global shape and local feature accuracy.

Datasets

  • Synthetic GPR Dataset: The authors utilize a synthetic GPR dataset described in [2]. This dataset simulates underground environments with buried rebars, utilities, and PVC pipes, providing B-scan data and corresponding ground truth cross-section images. For their experiments, they used 1628 B-scan data samples, with 100 for validation and 150 for testing.
  • Field Test Data: The system was also evaluated on field test data collected from a concrete slab at River Edge, NJ, USA. This real-world data was used to demonstrate the system's effectiveness in practical scenarios.

Key Insights and Ideas

  • Robotics for Data Acquisition: The use of an omnidirectional robot significantly improves the efficiency and flexibility of GPR data collection, overcoming limitations of traditional grid-based scanning.
  • End-to-End Deep Learning for Interpretation and Reconstruction: The paper demonstrates the power of deep learning for both interpreting raw GPR B-scans and completing sparse 3D point clouds, moving beyond traditional signal processing techniques.
  • GPRNet's Coarse-to-Dense Reconstruction: The GPRNet architecture is specifically designed to take a sparse, potentially incomplete point cloud (derived from GPR interpretation) and generate a dense, detailed representation, crucial for visualization.
  • Robustness to Noise: The proposed method shows promising robustness to noise in the input data, a common issue in GPR measurements.

Key Results

  • Quantitative Performance: In comparisons against baseline methods like PCN and PF-Net on the GPR dataset, GPRNet achieved lower Chamfer Distance (CD) and L1 distance metrics, indicating more accurate reconstruction. For example, GPRNet achieved a CD of 6.214 compared to PCN's 6.965 and PF-Net's 7.005.
  • Qualitative Performance: Visual comparisons show that GPRNet reconstructs more spatially continuous and accurate 3D models of underground utilities compared to other methods.
  • Noise Robustness: When tested with varying levels of Gaussian white noise (standard deviations of 0.01 to 0.2), GPRNet consistently outperformed baselines, demonstrating its reliability in real-world noisy conditions.

Limitations and Future Work

  • Limited to Pipe-Shaped Utilities: The current system is primarily designed for pipe-shaped utilities with round cross-sections. The paper does not explicitly discuss its performance on objects with more complex or irregular shapes.
  • Generalization to Diverse Materials: While the synthetic dataset includes various materials, the paper could explore further generalization to a wider range of soil types and utility materials.
  • Real-time Performance: While the system achieves good results, further optimization for real-time reconstruction during data collection could be a future direction.
  • Integration of Other Sensors: Future work could explore integrating other sensing modalities alongside GPR to further enhance reconstruction accuracy and robustness.

Summary updated

Abstract

Ground Penetrating Radar (GPR) is one of the most important non-destructive evaluation (NDE) instruments to detect and locate underground objects (i.e. rebars, utility pipes). Many of the previous researches focus on GPR image-based feature detection only, and none can process sparse GPR measurements to successfully reconstruct a very fine and detailed 3D model of underground objects for better visualization. To address this problem, this paper presents a novel robotic system to collect GPR data, localize the underground utilities, and reconstruct the underground objects’ dense point cloud model. This system is composed of three modules: 1) visual-inertial-based GPR data collection module which tags the GPR measurements with positioning information provided by an omnidirectional robot; 2) a deep neural network (DNN) migration module to interpret the raw GPR B-scan image into a cross-section of object model; 3) a DNN-based 3D reconstruction module, i.e., GPRNet, to generate underground utility model with the fine 3D point cloud. In this paper, both the quantitative and qualitative experiment results verify our method that can generate a dense and complete point cloud model of pipe-shaped utilities based on a sparse input, i.e., GPR raw data, with incompleteness and various noise. The experiment results on synthetic data as well as field test data further support the effectiveness of our approach.
BibTeX
@inproceedings{liang2021gpr,
  author = {Yang Liang and Jinglun Feng and Liang Yang and Ejup Hoxha and Diar Sanakov and Stanislav Sotnikov and Jizhong Xiao},
  title = {{GPR-based Model Reconstruction System for Underground Utilities Using GPRNet}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2021},
  doi = {10.1109/icra48506.2021.9561355},
}
Figure