Multi-Perspective Semantic Segmentation of Ground Penetrating Radar Images for Pavement Subsurface Objects

Effective infrastructure health monitoring is crucial within transportation cyber-physical systems, where accurate road health detection is vital for ensuring road safety and the stability of intelligent transportation systems. To address the challenges of identifying pavement subsurface objects using 3D ground penetrating radar (GPR) data, we propose a multi-perspective cascading recognition method that integrates B-scan and C-scan images. This method is built on a lightweight dual-stream semantic segmentation model called AttnGPRNet, developed in this work to improve feature extraction through attention mechanisms and enhance subsurface object recognition. Initially, the model segments B-scan images to identify potential target regions, followed by more precise segmentation of 3L-C-scan images based on preliminary results. Additionally, we constructed a multi-view dataset using 3D GPR scans from over 100 kilometers of urban roads and evaluated the effectiveness of the proposed method through experiments. Experimental results show that our model outperforms existing advanced methods, achieving mIoU of 78.80% and 83.96% on B-scan and 3L-C-scan, and F1 scores of 87.65% and 91.07%, respectively. Moreover, the method has been deployed in Xiaoning Road GPR image intelligent recognition system and verified through on-site drilling, demonstrating its practical potential for road health monitoring.