Abstract: As a pivotal nondestructive testing technique for urban road subsurface defect detection, Ground Penetrating Radar (GPR) has significantly surpassed traditional manual interpretation in field data acquisition efficiency. Meanwhile, manual interpretation suffers from several limitations, including high missed detection rates, inconsistent diagnostic criteria, and low timeliness. With advancements in computer hardware and the continuous evolution of object detection algorithms, object detection technology has achieved remarkable results across various sectors, including autonomous driving, smart manufacturing, and medical image analysis. Consequently, applying object detection to the analysis of GPR profiles is both theoretically grounded and practically meaningful. In this study, a defect sample database was constructed by systematically collating and categorizing typical GPR profiles from cities including Beijing, Hangzhou, and Shenzhen. By integrating the K-means clustering method into the YOLOV11 framework, the dataset was re-clustered to generate anchor box dimensions tailored for the current detection task, thereby improving the bounding box prediction accuracy and Intersection over Union (IOU). Through validation, the model achieved a mean Average Precision at 50% IOU (mAP50) of approximately 0.6 and mAP50-95 of approximately 0.4. The confidence levels for identifying typical defects such as cavities and voids consistently exceeded 0.9, enabling efficient and accurate automated detection of road hazards in GPR imagery. This work provides a significant reference for advancing the interpretation of GPR data from manual analysis toward intelligent, real-time automated processing.