Evidential transformer for buried object detection in ground penetrating radar signals and interval-based bounding box

Three-dimensional (3D) buried object detection using ground penetrating radar (GPR) benefits from the powerful capacity of image-wise deep neural networks. However, it still faces the challenge of information loss from raw GPR signals to two- and three-dimensional images, such as the frequency-domain information loss when normalizing GPR signals into gray-scale images and spatial information loss when using stacked B- and C-scan images to replace raw GPR signals as inputs. To solve the challenge, this study has proposed an ENNreg-transformer model, directly using raw 3D GPR signals to perform buried object detection. In the proposed model, 3D GPR signals are first converted into sequential voxelization to obtain spatiotemporal features. The features are then aggregated by an intuition-guided feature aggregation layer to simulate the expert behavior to analyze 3D GPR data. Finally, an evidential detection header outputs 3D interval-based bounding boxes for buried object detection. The experiment on two 3D GPR road datasets demonstrates that the proposed model exceeds other state-of-the-art models on the tasks thanks to raw 3D signals and intuition-guided feature aggregation. In addition, the interval-based bounding box represents the spatial bounding-box uncertainty, which derives from the inherent limitations of GPR and deep networks.