Automatic Detection And Localization Of Underground Target By Deep Learning Using Ground Penetrating Radar

With the development of construction industry,the workload of construction quality inspection is increasing.However,the current detection technology of rebar is inefficient and the detection data processing cycle is long,so it is difficult to meet the increasing engineering demand.Therefore,the rebar detection method with high efficiency and intelligent auxiliary data processing function has become the demand of the industry development.In our cities,there are many types and numbers of urban underground pipelines.The lack of pipeline location information will have a great impact on urban running and construction.Hence,it is necessary to detect underground pipelines and establish underground pipeline databases.For this huge workload,a high efficiency,high precision and non-destructive detection technology is required to carry out intelligent non-destructive detection of pipelines.The paper has done the research for a method of intelligent detection and localization,which is utilized to detect and localize the rebar in concrete and underground pipeline based on deep learning and ground penetrating radar(GPR).Rebars and underground pipelines are cylindrical targets,which show hyperbolas in B-scan images of GPR.According to this hyperbola’s feature,the Single Shot Multibox Detector(SSD)model is used to recognize the hyperbola in GPR image.Firstly,the real data of rebar in concrete and underground pipelines on-site were collected by GPR,and then the data augmentation is utilized to make dataset.Secondly,the SSD model was trained by dataset.After training,the SSD-R models and SSD-P model were obtained separately.It is found that both the SSD-R model and the SSD-P model have a high precision of target recognition.Specifically,the average precision of the SSD-R model for rebar detection is 90.9%,and the detection time is 0.47 s(2.1FPS),which can meet the requirements of real-time detection.The average precision of the SSD-P model for underground pipeline detection is 89.1%,and the detection time is 3.1 s(0.32 FPS),which can assist the detection work and shorten the cycle in engineering application.In order to localize the underground target,this paper proposes an automatic localization method,which consists of diffraction-summation migration,image binarization and target positioning estimation method.In the migration,the value of electromagnetic wave velocity is obtained by the high-order autofocus method,and the threshold of binarization is obtained by iterative algorithm.After the hyperbola recognition by the trained SSD model,the target area with the bounding box is extracted.The apex of hyperbola is determined by the proposed localization method,and then the buried depth or horizontal distance of the target is calculated.Verification experiments show that the proposed localization method has a high accuracy.The estimated error of the cover thickness of rebar is less than 1.2 mm,and the lateral position error is less than0.6 cm.The estimated error of the depth of underground pipeline is less than 0.05 m.In a word,the proposed method of intelligent detection and localization can improve the detection accuracy and work efficiency of field operations,which have high engineering application prospect.