| CPC G01S 13/885 (2013.01) [G01S 7/417 (2013.01)] | 6 Claims |

|
1. A ground penetrating radar (GPR) and deep learning-based underground pipeline detection method, comprising:
acquiring sample data of known underground pipelines by means of a GPR, and establishing a GPR B-scan dataset according to the sample data;
performing training according to the GPR B-scan dataset to obtain a target identification model You Only Look Once version 3 (YOLOV3)
identifying a hyperbola in a real GPR image of an underground pipeline target with the target identification model YOLOV3 to obtain a target bounding box for the underground pipeline target;
collapsing the hyperbola in the real GPR image into concentrated blobs at hyperbolic apexes by diffraction stack migration to obtain a reconstructed GPR image, and an expression of the diffraction stack migration is as follow:
![]() where E is the reconstructed GPR image, m is number of traces in a GPR B-scan, tais a two way travel-time ti, is amplitude of i-th recorded GPR signal at the two way travel-time to, which is calculated from relative coordinates of a imaging point (x,z) and antennas on a ground surface, as well as a velocity of a background medium;
applying a Hilbert transformation to each trace in the reconstructed GPR image to obtain an envelope image, and then binarising the image to distinguish target response from background;
applying a Gaussian filter to remove possible spikes in the background, thereby obtaining a round-shape blob that replaces the hyperbola in the real GPR image; and
locating position of the underground pipeline target through an apex of the round-shape blob.
|