CN-121978652-A - Method and system for processing equipment-side ground penetrating radar data based on embedded GPU

CN121978652ACN 121978652 ACN121978652 ACN 121978652ACN-121978652-A

Abstract

The invention provides a device-side ground penetrating radar data processing method and system based on an embedded GPU, wherein the method comprises the steps of receiving radar echoes formed after the ground penetrating space reflects radar reflection, and transmitting the radar echoes to an embedded GPU unit; analyzing the radar echo by utilizing a pre-training AI model deployed in the embedded GPU unit, and identifying an underground target body existing in the ground penetrating space. The invention can integrate 'embedded GPU + light AI' in the radar field, so that the equipment-side ground penetrating radar has the technical characteristics of instant identification, autonomous controllability and the like.

Inventors

ZHANG AIXI
SONG JUNJUN
LU WEIFENG
LIU XU
LIN HONGYANG
CAO WENHAO
SONG JINLU

Assignees

青岛探舆智能科技有限公司
青岛坤舆智能科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260320

Claims (10)

1. An embedded GPU-based equipment-side ground penetrating radar data processing method is characterized by comprising the following steps: Receiving radar echoes formed after radar emission is reflected by a ground penetrating space and transmitting the radar echoes to an embedded GPU unit, wherein the embedded GPU unit is positioned in a radar host; And analyzing the radar echo by utilizing a pre-training AI model deployed in the embedded GPU unit, and identifying the underground target body existing in the ground penetrating space.
2. The method of claim 1, wherein the radar echo is received via an antenna system, and wherein transmitting the radar echo to the embedded GPU unit is via a PCIe bus internal to the radar host.
3. The method of claim 1, further comprising implementing logic control and timing control with an FPGA control unit internal to the radar host to enable reception of radar echo data.
4. The method of claim 1, wherein the training set for pre-training the AI model is comprised of radar returns of a preset scale, locations of subsurface targets in radar returns, and class labels; the pre-training AI model is obtained based on training a two-stage target detection algorithm of a Faster-RCNN architecture, and the step of pre-training the AI model comprises the following steps: Extracting low-dimensional features to high-dimensional features from each radar echo, and connecting the low-dimensional features and the high-dimensional features in a cross-layer manner by utilizing a feature pyramid network to obtain a fused feature map; The fused feature diagram is used as input of a candidate frame generation network RPN training stage, a training target is constructed based on a position label of an underground target body in a radar echo included in a training set, and the capability of the candidate frame generation network for initially positioning the position of the underground target body in the radar echo is improved through iterative training; And taking the output of the candidate frame generation network as the input of a Fast-RCNN training stage of the target detection network, constructing a training target based on the position and class labels of the underground target in the radar echo included in the training set, improving the position of the underground target in the radar echo by the target detection network through iterative training, and determining the capability of the underground target type in the radar echo through multi-classification regression.
5. The method of claim 1 or 4, wherein prior to the step of parsing the radar echo using a pre-trained AI model deployed in an embedded GPU unit, the method further comprises pre-processing the radar echo, including removing background and real-time gain.
6. The method of claim 1, wherein upon identifying a subsurface target present in the ground penetrating space, the method further comprises: Superposing and displaying the underground object on the profile map of the ground penetrating space according to the identified position of the underground object, and marking the identified category of the underground object, and/or The locations of the identified subsurface targets are displayed superimposed on the map.
7. The method of claim 6, further comprising allowing a user to perform radar parameter configuration at the visual interactive interface, allowing the user to view the subsurface target volume displayed superimposed on the profile map or map.
8. An embedded GPU-based device-side ground penetrating radar data processing system, comprising: the system comprises an antenna system, an embedded GPU unit, a radar host and a radar system, wherein the antenna system is used for receiving radar echoes formed by the radar after the ground penetrating space reflects radar, and transmitting the radar echoes to the embedded GPU unit; And the embedded unit is used for analyzing the radar echo by utilizing the pre-training AI model deployed in the embedded GPU unit and identifying the underground target body existing in the ground penetrating space.
9. The ground penetrating radar data processing system of claim 8, further comprising an FPGA control unit inside the radar host for implementing logic control and timing control, and for implementing radar echo data reception; The training set of the AI model consists of a radar echo of a preset scale and the position and type labels of an underground target body in the radar echo, and the pre-training AI model is obtained by training based on a two-stage target detection algorithm of a Faster-RCNN architecture.
10. The ground penetrating radar data processing system of claim 8, further comprising a visual interactive interface for receiving radar parameter configuration by a user for visual presentation to the user of the subsurface target volume displayed superimposed on the profile map or map.

Description

Method and system for processing equipment-side ground penetrating radar data based on embedded GPU Technical Field The invention relates to the technical field of geophysical prospecting equipment, in particular to an equipment-end ground penetrating radar data processing method and system based on an embedded GPU. Background Ground penetrating radar (Ground PENETRATING RADAR, GPR) is a geophysical method for detecting subsurface structures using high frequency electromagnetic waves. With the acceleration of the urban process, the detection demand of the underground space is continuously increased, and the GPR is widely applied to multiple fields such as archaeology, mineral exploration, disaster geological investigation, geotechnical engineering investigation, engineering quality detection and the like due to the high precision, high efficiency and nondestructive detection characteristics of the GPR. Along with the development of the AI technology, the application of underground target identification on the ground penetrating radar data is wider and wider, and the traditional scheme adopts a notebook or a server for post-processing, so that the defects of dependence on experience, low efficiency, strong subjectivity, off-line data processing, low efficiency, high power consumption and large volume of a high-performance server, difficulty in carrying on an unmanned plane or a small robot and the like exist in the traditional scheme. With embedded GPUs (such as the domestic RKs 3588, jetson Orin NX, the inflight NVIDIA RTX series, etc.) have been conditioned to run lightweight deep learning models at the edge. Therefore, a ground penetrating radar system integrating acquisition and local AI identification, with low power consumption and being independently controllable is needed to realize quick identification of an underground target. Disclosure of Invention In view of this, embodiments of the present invention provide a method and a system for processing device-side ground penetrating radar data based on an embedded GPU, so as to eliminate or improve one or more drawbacks existing in the prior art. The invention provides a device-side ground penetrating radar data processing method based on an embedded GPU, which comprises the following steps of receiving radar echoes formed after radar emission is reflected by a ground penetrating space and transmitting the radar echoes to an embedded GPU unit; analyzing the radar echo by utilizing a pre-training AI model deployed in the embedded GPU unit, and identifying an underground target body existing in the ground penetrating space. In some embodiments of the invention, the radar returns are received through an antenna system, and the radar returns are transmitted to the embedded GPU unit through a PCIe bus internal to the radar host. In some embodiments of the invention, the method further comprises the step of utilizing an FPGA control unit inside the radar host to realize logic control and timing control and receiving radar echo data. In some embodiments of the invention, a training set for pre-training the AI model is composed of radar echoes of a preset scale and positions and class labels of underground targets in the radar echoes, the pre-training AI model is obtained by training a two-stage target detection algorithm based on a Fast-RCNN architecture, the pre-training AI model comprises the steps of extracting low-dimensional features to high-dimensional features from each radar echo, connecting the low-dimensional features and the high-dimensional features by utilizing a feature pyramid network in a cross-layer manner to obtain a fused feature map, generating inputs of a network RPN training stage by taking the fused feature map as a candidate frame, constructing a training target based on position labels of the underground targets in the radar echoes included in the training set, generating the capability of the network to initially position the underground targets in the radar echoes by iterating training lifting candidate frames, taking the output of the candidate frame generation network as the input of the target detection network Fast-RCNN training stage, constructing the training target based on the positions and class labels of the underground targets in the radar echoes included in the training set, determining the position of the underground targets in the positioning radar by the training lifting target detection network, and determining the capability of the underground targets in the multiple classes by returning the underground targets. In some embodiments of the present invention, prior to the step of resolving the radar echo using a pre-trained AI model deployed in the embedded GPU unit, the method further includes pre-processing the radar echo, including removing background and real-time gain. In some embodiments of the invention, after identifying the underground object present in the ground penetrating space, the method further comprise