Monday 31 March 2025
In recent years, researchers have been working on developing more effective and efficient ways to localize robots and vehicles in environments where traditional sensors like cameras and LiDAR may struggle. One such approach involves using ground-penetrating radar (GPR) to detect stable subsurface features and create detailed maps of the underground environment.
A new neural network architecture, dubbed EDENet, has been designed specifically for this task. The network uses a combination of learnable Gabor filters and direction-aware attention mechanisms to extract relevant information from GPR sequence data. This allows it to accurately identify corresponding locations in different environments, even when faced with changing weather conditions or varying underground dielectric constants.
The authors of the study tested EDENet on two public datasets: GROUNDED, which features sequences of GPR data collected under various weather conditions, and CMU-GPR, which includes single-channel GPR data. The results show that EDENet outperforms existing methods in both cases, with a particularly significant advantage when dealing with the complexities of GROUNDED.
One key innovation behind EDENet is its use of learnable Gabor filters to capture directional features in the GPR data. These filters are designed to respond to specific orientations and scales, allowing the network to effectively extract relevant information from the data. The direction-aware attention mechanism then focuses on these features, ensuring that the network pays the most attention to the most important parts of the sequence.
The authors also experimented with different scale combinations for the EDEBlocks, which are the building blocks of EDENet’s architecture. They found that a combination of larger and smaller kernels (35×35, 11×11, and 5×5) resulted in the best performance, as it allowed the network to capture both broad and fine-scale details.
The study also highlights the importance of carefully balancing the configuration of EDEBlocks and the learnable Gabor filters. The authors found that a smaller number of parameters in the filter bank led to better results, as it reduced redundancy and improved the overall quality of the features extracted.
In addition to its superior performance, EDENet is also notable for its efficiency and compactness. It requires significantly fewer parameters than other methods tested in the study, making it more suitable for real-time deployment on robots and vehicles. The authors also report that EDENet can process GPR sequences at a rate of 188 Hz, which is faster than the typical data acquisition rate.
Cite this article: “Ground-Penetrating Radar-Based Localization with EDENet”, The Science Archive, 2025.
Gpr, Edenet, Neural Network, Ground-Penetrating Radar, Gabor Filters, Direction-Aware Attention, Underground Mapping, Robotics, Vehicle Localization, Sensor Fusion.
