Multi-Target Rapid Detection of Road Underground Based on NCG-RetinaNet

The successful detection of road underground defects such as voids, cavities and pipeline leakages plays a role in preventing road collapses. In this study, a ground-penetrating radar detection vehicle is utilized to conduct extensive, generalized detection on roads. In order to adapt the detection speed and accuracy of the deep learning network to the needs of the detecting vehicle, this paper proposes a deep learning-based rapid multi-target detection method of road underground, named NCG-RetinaNet. First, the CBAM attention mechanism is introduced into the RetinaNet feature extraction network to improve the feature expression ability. Then, the neural architecture search method NAS-FPN is used to enhance the fusion of different scale features and improve the model's generalization ability. Finally, the GIoU loss function is introduced to calculate the bounding error more accurately and improve the detection precision. The results demonstrate that compared to the baseline network, the detection performance of NCG-RetinaNet has significantly improved, with the mean average precision (mAP) @0.5 and mAP@0.5:0.95 increasing by 4.3% and 5.3%, respectively. After improvements, the FPS of the network is 20.3 frames /s, and the average precision reaches 88.7% when the IoU is 0.5. It maintains high detection accuracy while ensuring rapid detection speed, meeting the requirements of GPR detection vehicles. Therefore, the method presented in this study provides reasonable support for faster and more accurate detection of underground voids and pipelines.