Equipment end ground penetrating radar data processing method and system based on embedded GPU
By integrating an embedded GPU and FPGA control unit into the ground penetrating radar host, and combining it with a lightweight AI model, the problems of low data processing efficiency and high power consumption of traditional ground penetrating radar are solved. This enables real-time identification and autonomous controllable detection of underground targets, lowers the technical threshold, and improves detection efficiency.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- QINGDAO TANYU INTELLIGENT TECH CO LTD
- Filing Date
- 2026-03-20
- Publication Date
- 2026-05-05
AI Technical Summary
Traditional ground-penetrating radar data processing relies on manual interpretation, which is inefficient and highly subjective. High-performance servers have high power consumption and large size, making them difficult to apply to drones or small robots and unable to achieve rapid identification of underground targets.
An embedded GPU unit is used to process data inside the radar host. Combined with FPGA control and a lightweight AI model, it can realize real-time identification of radar echoes and instant positioning of targets. Data is transmitted through PCIe bus, and a two-stage target detection algorithm based on the Faster-RCNN architecture is used for model training and identification.
It has achieved real-time identification and autonomous control of ground-penetrating radar, reduced the technical threshold for operators, improved the efficiency and reliability of underground target detection, and shortened the identification delay from minutes to milliseconds.
Smart Images
Figure CN121978652A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geophysical exploration equipment technology, and in particular to a device-side ground-penetrating radar data processing method and system based on an embedded GPU. Background Technology
[0002] Ground penetrating radar (GPR) is a geophysical method that uses high-frequency electromagnetic waves to detect underground structures. With the acceleration of urbanization, the demand for underground space exploration is constantly increasing. Due to its high precision, high efficiency, and non-destructive testing characteristics, GPR is widely used in various fields such as archaeology, mineral exploration, geological disaster investigation, geotechnical engineering investigation, and engineering quality testing.
[0003] With the development of AI technology, the application of ground penetrating radar data for underground target identification is becoming more and more widespread. Traditional solutions use laptops or servers for post-processing, which has disadvantages such as low efficiency and high subjectivity due to reliance on human interpretation and experience; offline data processing is also inefficient; and high-performance servers have high power consumption and large size, making them difficult to mount on drones or small robots.
[0004] With embedded GPUs (such as domestically produced RK3588 and Jetson Orin NX, and NVIDIA RTX series) now capable of running lightweight deep learning models at the edge, there is an urgent need for a low-power, autonomously controllable ground-penetrating radar system that integrates data acquisition and local AI recognition to achieve rapid identification of underground targets. Summary of the Invention
[0005] In view of this, embodiments of the present invention provide a device-side ground-penetrating radar data processing method and system based on an embedded GPU to eliminate or improve one or more defects existing in the prior art.
[0006] One aspect of the present invention provides a device-side ground-penetrating radar data processing method based on an embedded GPU. The method includes the following steps: receiving radar echoes formed after the ground-penetrating space reflects radar waves, and transmitting them to an embedded GPU unit; the embedded GPU unit is located inside the radar host; and using a pre-trained AI model deployed in the embedded GPU unit to analyze the radar echoes and identify underground targets existing in the ground-penetrating space.
[0007] In some embodiments of the present invention, the radar echo is received by an antenna system and transmitted to an embedded GPU unit via the PCIe bus inside the radar host.
[0008] In some embodiments of the present invention, the method further includes: using the FPGA control unit inside the radar host to implement logic control and timing control, thereby realizing the reception of radar echo data.
[0009] In some embodiments of the present invention, the training set for pre-training the AI model consists of radar echoes of a preset size, and the location and category labels of underground targets in the radar echoes. The pre-trained AI model is trained based on a two-stage target detection algorithm of the Faster-RCNN architecture. The steps of pre-training the AI model include: extracting low-dimensional features to high-dimensional features from each radar echo, and using a feature pyramid network to connect low-dimensional features and high-dimensional features across layers to obtain a fused feature map; using the fused feature map as input to the candidate box generation network (RPN) training stage, constructing training targets based on the location labels of underground targets in the radar echoes included in the training set, and improving the ability of the candidate box generation network to initially locate the location of underground targets in the radar echoes through iterative training; using the output of the candidate box generation network as input to the target detection network (Fast-RCNN) training stage, constructing training targets based on the location and category labels of underground targets in the radar echoes included in the training set, and improving the ability of the target detection network to determine the location of underground targets in the radar echoes and to determine the type of underground targets in the radar echoes through multi-class regression through iterative training.
[0010] In some embodiments of the present invention, before the step of parsing the radar echo using a pre-trained AI model deployed in an embedded GPU unit, the method further includes: preprocessing the radar echo, including removing background and real-time gain.
[0011] In some embodiments of the present invention, after identifying an underground target body in the exploration space, the method further includes: overlaying the underground target body onto a profile map of the exploration space according to the location of the identified underground target body, and marking the category of the identified underground target body; and / or overlaying the location of the identified underground target body onto a map.
[0012] In some embodiments of the present invention, the method further includes: allowing users to configure radar parameters in a visual interactive interface, and allowing users to view underground target bodies superimposed on profile maps or maps.
[0013] Corresponding to the above method, the present invention also provides a device-side ground-penetrating radar data processing system based on an embedded GPU. The system includes: an antenna system for receiving radar echoes formed after the ground-penetrating space reflects radar waves and transmitting them to an embedded GPU unit; the embedded GPU unit is located inside the radar host; and an embedded unit for using a pre-trained AI model deployed in the embedded GPU unit to analyze the radar echoes and identify underground targets existing in the ground-penetrating space.
[0014] In some embodiments of the present invention, the system further includes: an FPGA control unit inside the radar host, used to implement logic control and timing control, and to receive radar echo data; the training set of the AI model consists of radar echoes of a preset size, and the location and category labels of underground targets in the radar echoes; the pre-trained AI model is trained based on a two-stage target detection algorithm of the Faster-RCNN architecture.
[0015] In some embodiments of the present invention, the system further includes: a visual interactive interface for receiving radar parameter configurations performed by the user, and for visually presenting the underground target body superimposed on the profile atlas or map to the user.
[0016] This invention proposes a device-side ground-penetrating radar data processing method and system based on embedded GPUs. Through the integration of "embedded GPU + lightweight AI" in the radar field, it enables device-side ground-penetrating radar to possess technical characteristics such as real-time identification and autonomous control. This results in clear and reliable radar data detection and processing, helping to lower the technical threshold for radar operators and facilitating further engineering promotion and technological advancement. Additional advantages, objects, and features of this invention will be set forth in part in the description which follows, and will in part become apparent to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of this invention can be realized and obtained by means of the structures specifically pointed out in the specification and drawings.
[0017] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description
[0018] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings: Figure 1 This is a flowchart of a ground-penetrating radar data processing method at the device end in one embodiment of the present invention.
[0019] Figure 2This is a diagram of the device-side ground-penetrating radar data processing system architecture in one embodiment of the present invention.
[0020] Figure 3 This is a flowchart illustrating the workflow of a ground-penetrating radar data processing system at the device end in one embodiment of the present invention.
[0021] Figure 4 This is a flowchart illustrating the construction process of the training set and validation set in one embodiment of the present invention.
[0022] Figure 5 This is a schematic diagram of a visual interactive interface in one embodiment of the present invention. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.
[0024] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.
[0025] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.
[0026] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.
[0027] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.
[0028] To overcome the shortcomings of existing radar systems, this invention improves the radar host by integrating an embedded GPU unit. The radar echo is transmitted to the embedded GPU unit in real time via an internal high-speed PCIe bus. The embedded GPU unit runs a lightweight AI model for real-time identification. This method enables online detection of underground targets such as cavities, pipelines, and voids, and then transmits the detection results to a host computer via a gigabit Ethernet network, achieving the beneficial technical effect of "identifying and detecting simultaneously." Radar waves refer to the electromagnetic waves emitted (radiated) by the radar system into the detection (ground-penetrating) space. Radar echoes are crucial data formed by the radar system after emitting electromagnetic waves and receiving the reflected signals from targets.
[0029] Figure 1 This is a flowchart of a ground-penetrating radar data processing method at the device end according to an embodiment of the present invention. The method includes the following steps: Step S110: Receive the radar echo formed after the ground-penetrating space reflective radar wave and transmit it to the embedded GPU unit; the embedded GPU unit is located inside the radar host.
[0030] Step S120: Receive the radar echo formed after the ground-penetrating space reflective radar wave and transmit it to the embedded GPU unit; the embedded GPU unit is located inside the radar host.
[0031] The ground-penetrating radar data processing method based on embedded GPU proposed in this invention integrates "embedded GPU + lightweight AI" in the radar field, enabling the ground-penetrating radar to have technical characteristics such as real-time identification and autonomous control. This makes the radar data detection and processing results clear and reliable, helps to reduce the technical threshold for radar operators, and contributes to further engineering promotion and technological progress.
[0032] In some embodiments of the present invention, the radar echo is received by an antenna system and transmitted to an embedded GPU unit via the PCIe bus inside the radar host.
[0033] Using this embodiment of the invention, the FPGA control unit and the embedded GPU unit perform data throughput through the internal high-speed PCIe bus, which helps to reduce data transmission latency and improves the timeliness of AI intelligent recognition based on the GPU.
[0034] In some embodiments of the present invention, the method further includes: using the FPGA control unit inside the radar host to implement logic control and timing control, thereby realizing the reception of radar echo data.
[0035] The FPGA control unit controls the transmission of radar waves by controlling the transmission synchronization signal and the reception of radar echoes by controlling the reception synchronization signal, thereby controlling the manual antenna in the antenna system and receiving radar echo data.
[0036] By employing this embodiment of the invention, logic control and timing control are implemented using an FPGA control unit, thereby realizing the control of the transmitting and receiving antennas in the antenna system, thus achieving large-scale transmission of radar radio wave data and reception of radar echo data.
[0037] In some embodiments of the present invention, the training set used for pre-training the AI model consists of radar echoes of a preset size, and the location and category labels of underground targets within the radar echoes. The training set must ensure that there are a certain number of underground targets for each category; for example, at least 100 samples of each target type are selected (including radar echoes and the location and category of underground targets within the radar echoes). The radar echoes can be raw radar echo data or radar echo imaging.
[0038] Accordingly, the pre-trained AI model is trained based on the two-stage target detection algorithm of the Faster-RCNN architecture. The steps of pre-training the AI model include: (1) extracting low-dimensional features to high-dimensional features from each radar echo, and using the feature pyramid network to connect low-dimensional features and high-dimensional features across layers to obtain a fused feature map; (2) using the fused feature map as the input of the candidate box generation network RPN training stage, constructing training targets based on the location labels of underground targets in the radar echoes included in the training set, and improving the ability of the candidate box generation network to initially locate the location of underground targets in the radar echoes through iterative training; (3) using the output of the candidate box generation network as the input of the target detection network Fast-RCNN training stage, constructing training targets based on the location and category labels of underground targets in the radar echoes included in the training set, and improving the ability of the target detection network to determine the location of underground targets in the radar echoes and to determine the type of underground targets in the radar echoes through multi-class regression through iterative training.
[0039] The Region Proposal Net (RPN) is essentially a lightweight fully convolutional network. Its input is a fused feature map, which is processed using anchors to output a series of rectangular candidate boxes and the probability that each box contains the target object. In ground-penetrating radar (GPR) scenarios, considering that the echo characteristics of lesions may be very flat (e.g., interlayer voids) or very long and narrow (e.g., longitudinal cracks), the anchor size of the RPN network can be customized. If the lesion is very small in the image, the RPN needs to be mounted on a feature layer containing more low-dimensional features. The thousands of candidate boxes generated by the RPN network are finally filtered by Non-Maximum Suppression (NMS), and the remaining few hundred highest-quality boxes are sent to the second stage for classification.
[0040] The object detection network Fast-RCNN (Fast Region-based Convolutional Neural Network) takes candidate boxes of varying sizes and radar echo images as input from the previous RPN network as input. First, it transforms each candidate box into a fixed-size feature vector through the ROI Pooling layer. Then, it extracts full-image features based on the radar echo image and performs nonlinear transformation processing through a fully connected (FC) layer to output the classification result. The classification result includes the probability distribution of the target object's category and the offset of the bounding box.
[0041] Using this embodiment of the invention, an end-to-end training method is adopted to directly obtain task-related outputs (i.e., target position and category information). Through a two-stage target detection algorithm based on the Faster-RCNN architecture, compared with traditional intelligent recognition models based on CPU or offline GPU, it can automatically learn the latent feature description of the target with less sample data. It can achieve transfer applications between tasks through model transfer. In addition, due to the lightweight nature of the AI model, it can run in real time on an embedded GPU platform.
[0042] In some embodiments of the present invention, before the step of parsing the radar echo using a pre-trained AI model deployed in an embedded GPU unit, the method further includes: preprocessing the radar echo, including background removal and real-time gain. Optionally, the radar echo also needs to be converted into an image form. The image form of the radar echo (radar image) arranges the delay time, Doppler frequency, and amplitude / phase according to spatial coordinates, and uses pixel brightness to represent reflection intensity, thereby visually presenting the situation of the ground-penetrating space.
[0043] By employing this embodiment of the invention, the quality of radar echo data can be improved through preprocessing of radar echoes.
[0044] In some embodiments of the present invention, after identifying an underground target body existing in the exploration space, the method further includes: displaying the underground target body on a profile map of the exploration space according to the location of the identified underground target body, and marking the category of the identified underground target body; and / or displaying the location of the identified underground target body on a map.
[0045] The profile map in the exploration space shows the longitudinal (section) position of the underground target, which tells us how deep the underground target is underground and how large its influence range is, etc. The map or topographic map shows the horizontal position of the underground target, that is, which position on the ground corresponds to the underground target and where on the map this position corresponds.
[0046] Optionally, the above visualization process is implemented on the radar host or its host computer, and the location and type of the identified underground target are transmitted to the host computer via wired or wireless network.
[0047] By adopting this embodiment of the invention, users can quickly delineate underground cavities, voids, and other defective areas (i.e., target bodies) during the detection process without much ground-penetrating radar map interpretation ability, thanks to the user-friendly interface and intuitive operation design. The operation is simple, which helps to lower the technical threshold for radar users and expand the application scope of ground-penetrating radar.
[0048] In some embodiments of the present invention, the method further includes: allowing users to configure radar parameters in a visual interactive interface, and allowing users to view underground target bodies superimposed on profile maps or maps.
[0049] Using this embodiment of the invention, radar parameter configurations from users can be received in real time through a user-friendly interface, and the detection results of the ground-penetrating radar on the device can be displayed to the user in real time, with the location of the detected target marked on the map.
[0050] Corresponding to the above method, the present invention also provides a device-side ground-penetrating radar data processing system based on an embedded GPU. Figure 2 This is a diagram of the device-side ground-penetrating radar data processing system architecture in one embodiment of the present invention, as shown below. Figure 2 As shown, the system includes: The radar host 100 has an embedded GPU unit 110 deployed inside. The embedded unit 110 is used to analyze the radar echo and identify underground targets in the ground-penetrating space using a pre-trained AI model deployed in the embedded GPU unit.
[0051] Antenna system 200 is used to receive radar echoes formed after the ground-penetrating space radar reflects the radar waves and transmit them to embedded GPU unit 110.
[0052] By adopting this embodiment of the invention, the integration of "embedded GPU + lightweight AI" in the field of radar can be realized. Real-time identification of underground targets can be achieved through embedded GPU, which can reduce the identification latency from the traditional "minute level" to "millisecond level". For on-site detection of underground targets (diseases), alarm information can be generated in real time.
[0053] In some embodiments of the present invention, the system further includes: an FPGA control unit 120 inside the radar host, used to implement logic control and timing control, and to receive radar echo data.
[0054] In some embodiments of the present invention, the system further includes: a host computer 300, which receives disease information uploaded by the embedded GPU unit 110 and processes it in real time. This method helps to improve the timeliness of early warning.
[0055] Optionally, the training set of the AI model consists of radar echoes of a preset size, and the location and category labels of underground targets in the radar echoes; the pre-trained AI model is trained based on a two-stage target detection algorithm of the Faster-RCNN architecture.
[0056] By employing this embodiment of the invention, a two-stage object detection algorithm based on the Faster-RCNN architecture can automatically learn the latent feature description of the target with less sample data compared to traditional intelligent recognition models based on CPU or offline GPUs. This enables cross-task application through model transfer.
[0057] In some embodiments of the present invention, the system further includes: a visual interactive interface for receiving radar parameter configurations performed by the user, and for visually presenting the underground target body superimposed on the profile atlas or map to the user.
[0058] Optionally, in order to enhance the cross-platform access capability of the damage information, after the radar equipment itself can intelligently identify the data and generate damage information (information such as underground cavities, voids, etc. that endanger road safety), the damage information can be directly reported to the back-end server platform, which will then respond and process it in real time. This method helps to improve the timeliness of early warning.
[0059] By adopting this embodiment of the invention, users can quickly locate underground targets during the detection process without much ground-penetrating radar map interpretation ability through a user-friendly interface and intuitive operation design. The operation is simple, which helps to lower the technical threshold for radar users and expand the application scope of ground-penetrating radar.
[0060] Figure 3 This is a flowchart illustrating the workflow of a ground-penetrating radar data processing system at the device end, according to one embodiment of the present invention. Figure 3 As shown, the radar host 100 includes an embedded GPU unit 110 and an FPGA control unit 120. Data throughput is achieved through an internal high-speed PCIe bus. The raw radar echo data is transmitted back to the embedded GPU unit 110 via the PCIe bus. The embedded GPU unit 110 is used to realize intelligent identification of underground targets and send the intelligent identification results to the host computer 300 in real time.
[0061] Optionally, the embedded GPU unit 110 can also be used to pre-process the raw radar echo data, such as removing background and real-time gain, and then run AI algorithms to intelligently identify underground targets in the radar echo.
[0062] Figure 4 This is a flowchart illustrating the construction of the training and validation sets in one embodiment of the present invention. Optionally, a sample dataset of no less than 100 samples can be provided for each target type to jointly create the training and validation sets, and based on, for example... Figure 4 The process shown is used for logical judgment.
[0063] In a specific embodiment of the present invention, a method for constructing and training a lightweight AI model is provided. Optionally, a two-stage algorithm based on the Faster-RCNN architecture can be used to achieve intelligent target recognition. The two-stage algorithm is a target detection algorithm based on a two-stage model, which outperforms one-stage target detection algorithms in terms of detection rate and target location regression accuracy, but is weaker in real-time performance.
[0064] The two-stage algorithm consists of two stages: RPN (Region Proposal Net, used to extract candidate bounding boxes) and Fast-RCNN (based on the candidate bounding boxes, performing position correction and class prediction). The algorithm is implemented and object detection is achieved by training the RPN candidate bounding box generation network and the Fast-RCNN object detection network.
[0065] On the other hand, the deep learning-based object detection and recognition framework consists of two parts in the design and optimization phase: a backbone feature extraction module and a task regression module.
[0066] For the backbone feature extraction module, given the current limited number of data samples, this invention improves the deep network through network feature enhancement technology to further enhance detection capabilities. Since the target signal-to-noise ratio is low and features are weak, the network's feature representation capabilities need to be enhanced to improve the target detection rate. However, as the network framework increases in depth, while its semantic representation capabilities gradually improve, the target location features weaken. To resolve this contradiction, this invention introduces pyramid features into the deep network framework. By connecting low-dimensional and high-dimensional features in a pyramid structure, high-precision location information and high-precision semantic information are effectively fused, thereby improving the target detection rate.
[0067] Furthermore, this invention selects an efficiency evaluation parameter exposure method to achieve module encapsulation. Specifically, the intelligent recognition algorithm (AI model) is engineered to output a dynamic link library, enabling detection efficiency evaluation on large-scale profile data. By generating a B / S architecture, the algorithm calls the dynamic link library file and corresponding header files and calling examples using a C / S architecture software, thereby achieving large-scale profile data detection and evaluating detection efficiency in a GPU environment.
[0068] Figure 5 This is a schematic diagram of a visual interactive interface in one embodiment of the present invention. Figure 5 The image shows a visualization of the real-time intelligent detection results of the radar host. The data can be transmitted to the host computer via gigabit Ethernet, overlaid on the profile map, and the detected target locations can be displayed in the map display area.
[0069] Corresponding to the above method, the present invention also provides an embedded DPU unit, which includes a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the embedded DPU unit implements the steps of the method described above.
[0070] Corresponding to the methods described above, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implements the steps of the method as described in any of the above embodiments. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.
[0071] Corresponding to the above methods, the present invention also provides a computer program product, including a computer program / instructions that, when executed by a processor, implement the steps of the method as described in any of the above embodiments.
[0072] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.
[0073] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.
[0074] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.
[0075] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A device-side ground-penetrating radar data processing method based on an embedded GPU, characterized in that, The method includes: The radar echo formed after the ground-penetrating space radar wave is received is transmitted to the embedded GPU unit; the embedded GPU unit is located inside the radar host. The radar echoes are analyzed using a pre-trained AI model deployed in an embedded GPU unit to identify underground targets in the probe space.
2. The method according to claim 1, characterized in that, The radar echo is received by the antenna system and transmitted to the embedded GPU unit via the PCIe bus inside the radar host.
3. The method according to claim 1, characterized in that, The method further includes: using the FPGA control unit inside the radar host to implement logic control and timing control, thereby realizing the reception of radar echo data.
4. The method according to claim 1, characterized in that, The training set used to pre-train the AI model consists of radar echoes of a preset size, and the location and category labels of underground targets in the radar echoes. The pre-trained AI model is trained based on a two-stage object detection algorithm using the Faster-RCNN architecture. The steps for pre-training the AI model include: Extract low-dimensional features to high-dimensional features from each radar echo, and use a feature pyramid network to connect low-dimensional features and high-dimensional features across layers to obtain a fused feature map. The fused feature map is used as the input to the candidate box generation network (RPN) training stage. Training targets are constructed based on the location labels of underground targets in radar echoes included in the training set. The ability of the candidate box generation network to initially locate the location of underground targets in radar echoes is improved through iterative training. The output of the candidate box generation network is used as the input to the training stage of the target detection network Fast-RCNN. Training targets are constructed based on the location and category labels of underground targets in radar echoes included in the training set. Through iterative training, the ability of the target detection network to determine the location of underground targets in radar echoes and to determine the type of underground targets in radar echoes is improved.
5. The method according to claim 1 or 4, characterized in that, Before the step of parsing the radar echo using a pre-trained AI model deployed in an embedded GPU unit, the method further includes: preprocessing the radar echo, including background removal and real-time gain.
6. The method according to claim 1, characterized in that, After identifying the underground target body present in the exploration space, the method further includes: Based on the location of the identified underground targets, overlay the targets onto the profile map of the exploration space, and label the categories of the identified underground targets; and / or The location of the identified underground targets is overlaid on the map.
7. The method according to claim 6, characterized in that, The method further includes: allowing users to configure radar parameters in a visual interactive interface, and allowing users to view underground target bodies superimposed on profile maps or maps.
8. A device-side ground-penetrating radar data processing system based on an embedded GPU, characterized in that, The system includes: The antenna system is used to receive the radar echo formed after the ground-penetrating space radar reflects the radar wave and transmit it to the embedded GPU unit; the embedded GPU unit is located inside the radar host. An embedded unit is used to analyze the radar echo and identify underground targets in the probe space by utilizing a pre-trained AI model deployed in the embedded GPU unit.
9. The ground-penetrating radar data processing system according to claim 8, characterized in that, The system also includes: an FPGA control unit inside the radar host, used to implement logic control and timing control, and to receive radar echo data; The training set of the AI model consists of radar echoes of a preset size, the location and category labels of underground targets in the radar echoes; the pre-trained AI model is trained based on a two-stage target detection algorithm of the Faster-RCNN architecture.
10. The ground-penetrating radar data processing system according to claim 8, characterized in that, The system also includes: a visual interactive interface for receiving radar parameter configurations from the user and for visually presenting the underground target body overlaid on the profile map or map to the user.