GPR-OdomNet: Difference and Similarity-Driven Odometry Estimation Network for Ground Penetrating Radar-Based Localization

TL;DR

This work addresses robust GPR-based odometry for localization in challenging environments by leveraging information in consecutive B-scan images. GPR-OdomNet combines multi-scale feature extraction with difference and similarity detectors, using attention-enhanced difference features and cosine similarity to estimate inter-scan distance $O_{t-1,t}$ from $B_{t-1}$ and $B_t$. Across ablations and CMU-GPR dataset evaluations, the method outperforms state-of-the-art approaches, achieving an overall RMSE of $0.449\mathrm{\,m}$ and a $10.2\%$ improvement over the best baselines, and demonstrates additional gains when integrated into sensor fusion via a factor-graph framework. The results suggest GPR-OdomNet provides a robust, weather-resilient localization signal that can enhance autonomous navigation in environments where traditional sensors fail.

Abstract

When performing robot/vehicle localization using ground penetrating radar (GPR) to handle adverse weather and environmental conditions, existing techniques often struggle to accurately estimate distances when processing B-scan images with minor distinctions. This study introduces a new neural network-based odometry method that leverages the similarity and difference features of GPR B-scan images for precise estimation of the Euclidean distances traveled between the B-scan images. The new custom neural network extracts multi-scale features from B-scan images taken at consecutive moments and then determines the Euclidean distance traveled by analyzing the similarities and differences between these features. To evaluate our method, an ablation study and comparison experiments have been conducted using the publicly available CMU-GPR dataset. The experimental results show that our method consistently outperforms state-of-the-art counterparts in all tests. Specifically, our method achieves a root mean square error (RMSE), and achieves an overall weighted RMSE of 0.449 m across all data sets, which is a 10.2\% reduction in RMSE when compared to the best state-of-the-art method.

Figures (6)

• Figure 1: A new GPR odometry network that estimates the distance by comparing the differences and similarities between two consecutive time step B-scan images.
• Figure 2: System architecture of GPR-OdomNet. The network takes B-scan images from two consecutive time steps as input. Initially, multi-scale features are extracted from each image separately. Subsequently, the difference tensor is obtained by detecting difference between the two feature sets, and the similarity tensor is calculated through cosine similarity. Finally, the fully connected layers perform regression on the difference tensor and similarity tensor to yield the final distance. ⓒ means tensor concatenation, × means tensor multiplication, - means tensor subtraction, and $D$ means distance.
• Figure 3: Examples of intermediate features of the GPR-OdomNet. Due to the large feature dimension of $F_{t-1,d}$, $F_{t,d}$, $F_{t-1,s}$, $F_{t,s}$, $D_{t-1,t}$ and $S_{t-1,t}$, the illustration here is a randomly sampled. The first eight subgraphs in (c) belong to $F_{t-1,d}$, while the last eight belong to $F_{t,d}$
• Figure 4: Overview of the factor graph model. Square nodes represent factor nodes, and circular nodes represent variable nodes. The prior (PR) node is a prior factor node. The IMU (IM) node mainly constrains the pose and velocity between two consecutive moments. The Wheel Encoder (WE) and GPR-OdomNet (GO) nodes constrain the velocity of the next moment $t$. The Bias (BI) node constrains the biases between two consecutive moments.
• Figure 5: Comparison of odometry performance in sensor fusion using the gates_g dataset.