EDENet: Echo Direction Encoding Network for Place Recognition Based on Ground Penetrating Radar

Pengyu Zhang, Xieyuanli Chen, Yuwei Chen, Beizhen Bi, Zhuo Xu, Tian Jin, Xiaotao Huang, Liang ShenAAAI 2025 · DOI

TL;DR

This paper introduces EDENet, a novel deep learning architecture for place recognition using Ground Penetrating Radar (GPR). EDENet effectively tackles the challenges of sparse underground features and varying dielectric constants by employing learnable Gabor filters and a direction-aware attention mechanism to extract robust directional features, achieving state-of-the-art performance with improved efficiency.

What Problem Does This Paper Address?

Robots and autonomous vehicles often rely on sensors like LiDAR and cameras for localization, which involves matching current sensor readings to a pre-existing map. However, these sensors struggle in environments with poor visibility or lack of distinct features. Ground Penetrating Radar (GPR) offers a promising alternative as it can detect stable subsurface features, unaffected by surface conditions.

Despite GPR's potential, applying it to place recognition, especially in large-scale environments, presents significant hurdles. Underground features are inherently sparse and less distinct compared to surface features. Furthermore, the dielectric properties of the subsurface can change due to environmental factors (like moisture), altering the GPR echo characteristics and complicating accurate localization. Existing methods often fail to robustly handle these variations, leading to high mismatch rates or requiring computationally expensive models. This paper aims to bridge this gap by developing a more robust and efficient GPR-based place recognition system.

Methods

The core of EDENet lies in its novel Echo Direction Encoding (EDE) Block. This block is designed to extract and encode directional features from GPR echoes, addressing the variability caused by changing dielectric constants.

Key components of the EDE Block include:

  • Learnable Gabor Filters (LGF): Instead of fixed filters, EDENet uses learnable Gabor filters. These filters are inspired by the hyperbolic nature of GPR echoes and can adaptively capture directional responses. The learnability allows them to represent various filter types, from Gaussian to edge-fitting filters, making them versatile for extracting valuable directional information and filtering out noise.
  • Direction-Aware Attention (DAA) Module: This module dynamically reweights the features extracted by the LGF layer. It models the interdependence between different directional kernels, allowing the network to focus on the most informative features and suppress less useful ones. This adaptive reweighting is crucial for handling interference and noise.
  • Shift Invariant Unit: This component is incorporated to improve robustness against slight shifts in target positions caused by variations in underground dielectric constants. It uses a cascaded max-pooling scheme to enhance feature specificity while maintaining a degree of shift invariance.
  • Multi-Scale Architecture: EDENet employs multiple EDE Blocks of different scales (e.g., 35x35, 23x23, 11x11, 5x5). This multi-scale approach allows the network to capture both broad, global underground structures and fine-grained local details, leading to a more comprehensive representation.

Finally, these multi-scale features are aggregated using a Multi-Layer Perceptron (MLP) to generate a compact global descriptor. For place recognition, a vector retrieval scheme is used to find the best matching descriptor in a pre-built database.

Datasets

The paper evaluates EDENet on two datasets:

  • GROUNDED: A 3D multi-channel GPR dataset.
  • CMU-GPR: A 2D single-channel GPR dataset.

These datasets are used for both training and testing, with specific splits for validation and testing outlined in the paper. No new benchmark datasets were introduced.

Key Insights and Ideas

  • Directional Features are Key: The paper highlights that despite variations in dielectric constants, the directional geometric features within GPR echoes remain relatively consistent. This insight is fundamental to the design of EDENet, focusing on extracting these directional cues.
  • Learnable Filters for Adaptability: Moving from fixed filters to learnable Gabor filters is a critical step. It allows the network to adapt its feature extraction to the specific characteristics of GPR data and the challenges posed by underground environments, rather than relying on generic, pre-defined filters.
  • Attention for Robustness: The Direction-Aware Attention mechanism is not just about selecting the "best" feature but about adaptively reweighting and combining directional information. This is vital for filtering out noise and interference, which are significant problems in GPR data.
  • Multi-Scale for Comprehensive Understanding: The use of multi-scale EDE Blocks is essential. It allows EDENet to build a rich representation by considering features at different levels of detail, from large structures to fine textures, which is crucial for robust place recognition.

Key Results

EDENet demonstrates state-of-the-art performance on both the GROUNDED and CMU-GPR datasets, outperforming existing GPR-specific methods and even adapted camera/LiDAR methods.

  • Performance: EDENet achieves the highest recall@5 rates across various scenarios, notably achieving 95% recall@5 on the GROUNDED dataset and 76% recall@5 on the CMU-GPR dataset (with k=64).
  • Efficiency: EDENet is significantly more compact in model size (5.0*10^8 parameters) and offers a competitive runtime (0.005s per frame) compared to other methods, making it suitable for real-time deployment. It achieves a runtime of 188 Hz with a 100-frame GPR sequence.
  • Ablation Studies: Experiments show that both the Learnable Gabor Filter (LGF) and the Direction-Aware Attention (DAA) module are crucial. Removing LGF leads to a 6% performance drop, while removing DAA results in a 4% drop, validating the effectiveness of these novel components.

Limitations and Future Work

  • Parameter Optimization: The paper acknowledges that the configuration of EDEBlocks (scale) and the parameter 'k' (angular resolution of Gabor filters) require careful optimization to balance performance and avoid redundancy or overfitting.
  • Feature Aggregation Layer: While the current feature aggregation layer is effective, the authors suggest future work could focus on designing a more efficient and lightweight aggregation layer specifically for positioning based on GPR sequences.
  • Real-time Deployment: Although EDENet shows promising efficiency, further research into optimizing the feature aggregation layer could lead to even more efficient and lightweight solutions for real-time deployment.

Summary updated

BibTeX
@inproceedings{zhang2025edenet,
  author = {Pengyu Zhang and Xieyuanli Chen and Yuwei Chen and Beizhen Bi and Zhuo Xu and Tian Jin and Xiaotao Huang and Liang Shen},
  title = {{EDENet: Echo Direction Encoding Network for Place Recognition Based on Ground Penetrating Radar}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i10.33092},
}
Figure