A Smart Detection and Location Method for Underground Hard Foreign Objects Based on Dynamic Feature Fusion

Abstract

This invention discloses an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion, belonging to the field of ground-penetrating radar technology. The technical solution includes: acquiring and processing ground-penetrating radar images; using a lightweight model based on an improved YOLO11 for real-time identification, which reconstructs the backbone network through a GhostNetV3 module and introduces a dynamic weighted attention mechanism to achieve feature fusion; utilizing an adaptive feature fusion algorithm to enhance the identification of small targets; employing a Transformer architecture for spatiotemporal sequence modeling and combining multi-source sensor fusion of GPS-RTK and IMU to achieve centimeter-level positioning; and finally, automatically marking the target on the ground using an integrated spraying execution unit. This invention achieves a fully automated closed-loop process for the detection, identification, location, and marking of underground hard foreign objects.

Description

Technical Field

[0001] This invention belongs to the field of ground-penetrating radar technology, and in particular relates to an intelligent detection and positioning method for underground hard foreign objects based on dynamic feature fusion. Background Technology

[0002] In the field of ground-penetrating radar (GPR) technology, the efficient and accurate detection and location of hard underground foreign objects is a critical task, especially in practical engineering applications such as infrastructure construction, farmland reclamation, and pipeline surveys. However, existing technologies face a series of technical bottlenecks that urgently need to be addressed in practice. First, the current interpretation of GPR data relies heavily on the experience of professionals for manual interpretation, a process that is highly subjective and inefficient. In large-scale survey scenarios, such as investigating underground obstacles in tens of hectares of uncultivated wasteland, manually processing massive amounts of radar images is not only time-consuming and labor-intensive, but the interpretation results also lack consistency and standardization due to differences in personnel experience, making rapid and objective batch processing difficult and severely restricting operational efficiency.

[0003] Secondly, although existing research has attempted to introduce deep learning models based on convolutional neural networks to automatically identify targets in radar images, these models are typically designed for natural images and are inherently mismatched with the characteristics of ground-penetrating radar (GPR) data. GPR data is essentially a spatiotemporal sequence of signals continuously acquired along the survey line, containing information about the spatial and temporal relationships of targets. Existing general-purpose models fail to effectively model this spatiotemporal sequence characteristic, resulting in insufficient utilization of the features of small or slender targets such as rebar and pipelines, thus limiting identification accuracy. Furthermore, such models generally have a high number of parameters and computational complexity, making it difficult to meet the real-time processing requirements of edge computing devices mounted on mobile detection platforms, hindering the realization of on-site real-time analysis.

[0004] Finally, even if software algorithms achieve the identification and preliminary location of underground targets, the existing technological system still fails to complete the final closed loop from digital information to the physical world. In actual operations, the identified target location information still requires operators to manually mark it on the ground based on electronic drawings or data. This process not only introduces additional human error and reduces overall positioning accuracy, but is also inefficient, preventing the intelligent detection results from being efficiently, accurately, and traceably fed back to guide on-site physical construction or cleanup operations. Therefore, how to construct a closed-loop system integrating automated detection, intelligent identification, precise positioning, and on-site autonomous marking is a prominent technical challenge in this field. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes an intelligent detection and positioning method for underground hard foreign objects based on dynamic feature fusion, thereby resolving the issues present in the existing technologies.

[0006] Firstly, to achieve the above objectives, this invention provides an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion, comprising the following steps: Acquire ground-penetrating radar data and perform image processing; A lightweight model based on an improved deep learning algorithm is used to process and recognize the images in real time. An adaptive feature fusion algorithm is used to improve the recognition rate of small-sized hard foreign objects; By employing spatiotemporal sequence modeling and combining multi-source sensor fusion and localization, a closed loop of detection, identification, and localization is achieved. When a foreign object is detected, the system outputs location data and controls the actuator to spray a mark on the ground.

[0007] Optionally, the process of employing a lightweight model improved based on deep learning algorithms includes: Replace the YOLO11 backbone network with a lightweight backbone network reconstructed from the GhostNetV3 module; Design a lightweight decoupled detection head based on a moving inverted bottleneck; The original neck network is replaced by a multi-scale feature adaptive aggregation network, the core of which is a dynamic weight attention module.

[0008] Optionally, the process of real-time processing and recognition of the lightweight model also includes: Receive feature maps of different scales from the lightweight backbone network; The feature maps at various scales are analyzed by the miniature neural network in the dynamic weight attention module, and a dynamic weight map is generated. The dynamic weight map is used to perform pixel-level weighted fusion of features at each scale.

[0009] Optionally, the feature adaptive fusion algorithm process includes: Based on the aforementioned lightweight model, an adaptive weighted fusion module guided by feature confidence is designed; The confidence level of each spatial location on feature maps at different scales is evaluated through a lightweight subnetwork, and a dynamic weight map is generated. The dynamic weight map is used to guide pixel-level adaptive fusion of multi-scale features.

[0010] Optionally, the process of using spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, includes: Simultaneously acquire ground-penetrating radar data, GPS-RTK positioning data, and IMU inertial measurement data; The radar images within a continuous time window, along with their corresponding timestamps and antenna movement speed, are treated as a spatiotemporal sequence and input into the Transformer model for identification. By employing a tightly coupled Kalman filter algorithm and fusing GPS-RTK positioning data with IMU data, the precise pose of the radar antenna at each data acquisition moment is obtained.

[0011] Optionally, the process of using spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, also includes: For the identified target, its local coordinates relative to the antenna phase center are calculated based on its pixel coordinates in the image and the radar wave velocity model. Using the precise pose, the local coordinates are transformed to the global geodetic coordinate system through three-dimensional coordinate transformation.

[0012] Optionally, after target identification and before local coordinate calculation, the process further includes: An accurate parameter extraction module based on nonmaximum suppression and linear interpolation post-processing is embedded to refine the recognition results.

[0013] Optionally, the process of controlling the actuator to spray markings on the ground is executed by a ground autonomous marking and feedback execution unit, which includes: A high-precision servo drive module is used to control the movement of the actuator in a two-dimensional plane; Programmable spraying actuator for spraying marking paint; The central controller is used to receive instructions and coordinate the driving and spraying actions.

[0014] Secondly, the present invention also provides a computer terminal device, comprising: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the intelligent detection and positioning method for underground hard foreign objects based on dynamic feature fusion in the first aspect above.

[0015] Thirdly, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, it implements the steps of the intelligent detection and positioning method for underground hard foreign objects based on dynamic feature fusion in the first aspect described above.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: This invention provides an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion. By introducing a dynamic feature fusion mechanism based on an improved lightweight model, it significantly improves the automatic identification capability and anti-interference ability of small underground hard foreign objects, while meeting the computational requirements of real-time processing by edge devices. Combining Transformer spatiotemporal sequence modeling and multi-source sensor fusion positioning technology, it achieves full automation of the entire process of underground target detection, intelligent identification, and centimeter-level precise positioning. Finally, by integrating ground autonomous marking units, a complete closed loop from digital information to physical marking is formed, completely avoiding errors introduced by manual secondary marking and greatly improving the overall efficiency and accuracy of on-site operations. Attached Figure Description

[0017] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a flowchart illustrating an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the attention-addition mechanism in an embodiment of the present invention; Figure 3 This is a diagram illustrating the real-time reading and recording of GPS positioning information from an iRTK5 device according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the structure for training a lightweight model based on an improved YOLO11 according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] like Figure 1 As shown, this embodiment provides a method for intelligent detection and location of underground hard foreign objects based on dynamic feature fusion, including: Acquire ground-penetrating radar data and perform image processing; A lightweight model based on an improved deep learning algorithm is used to process and recognize the images in real time. An adaptive feature fusion algorithm is used to improve the recognition rate of small-sized hard foreign objects; By employing spatiotemporal sequence modeling and combining multi-source sensor fusion and localization, a closed loop of detection, identification, and localization is achieved. When a foreign object is detected, the system outputs location data and controls the actuator to spray a mark on the ground.

[0021] like Figure 4 As shown, the further process of adopting a lightweight model improved based on deep learning algorithms includes: Replace the YOLO11 backbone network with a lightweight backbone network reconstructed from the GhostNetV3 module; Design a lightweight decoupled detection head based on a moving inverted bottleneck; The original neck network is replaced by a multi-scale feature adaptive aggregation network, the core of which is a dynamic weight attention module.

[0022] Specifically, the implementation process of this embodiment includes: A lightweight model based on the YOLO11-improved deep learning algorithm is proposed. The lightweight backbone network is reconstructed using GhostNetV3 modules, and a lightweight decoupled detection head based on a moving inverted bottleneck is designed, significantly reducing model parameters and computational cost. A multi-scale feature adaptive aggregation network is proposed to replace the original neck network. The core is a dynamic weight attention module, which analyzes multi-scale feature maps in real time through a micro-neural network to generate a spatial dynamic weight map, achieving pixel-level adaptive weighted fusion. This autonomously enhances effective target features and suppresses clutter. For radar signal characteristics, complex data input and local fluctuation feature enhancement modules are introduced.

[0023] The lightweight backbone network replaces the native YOLO11 backbone network with a GhostNetV3 module. During training, it enhances representation capabilities through reparameterization techniques, and during inference, it merges into a minimalist structure, significantly reducing the number of parameters and computational cost, laying the foundation for edge deployment. The lightweight detection head employs a lightweight, decoupled detection head based on the moving inversion bottleneck and channel attention. While reducing computational complexity, it improves the ability to filter key features through an attention mechanism.

[0024] Furthermore, the real-time processing and recognition process of the lightweight model also includes: Receive feature maps of different scales from the lightweight backbone network; SENet is introduced as a front-channel attention mechanism; The feature maps at various scales are analyzed by the miniature neural network in the dynamic weight attention module, and a dynamic weight map is generated. The dynamic weight map is used to perform pixel-level weighted fusion of features at each scale.

[0025] Specifically, the implementation process of this embodiment includes: like Figure 2 As shown, the SENet module is used to recalibrate the channel dimensions of the feature maps at each level, suppressing clutter channel responses and enhancing target-related channel responses, thus providing cleaner input features for the subsequent dynamic weight attention module.

[0026] A lightweight model based on a YOLO11-improved deep learning algorithm, integrating multi-scale radar signal features with a machine learning dynamic weight allocation mechanism, includes: The system receives feature maps at three different scales from a lightweight backbone: low-level high-resolution and high-level high-semantic features. A lightweight dynamic weight attention module is introduced, which uses a micro neural network to analyze the spatial and channel information of the feature maps at each scale in real time and generate corresponding dynamic weight maps. The generated dynamic weight maps are used to perform pixel-level weighted fusion of features at each scale, enabling the network to autonomously enhance the effective features of targets of interest in ground-penetrating radar images and suppress invalid noise.

[0027] Special optimizations were made for ground-penetrating radar data, using the real and imaginary parts of the radar signal as dual-channel inputs to preserve complete electromagnetic field information. In the early stage of feature extraction, a lightweight hollow convolutional spatial pyramid module was embedded to capture the unique multi-scale local fluctuation patterns of underground target reflections.

[0028] Furthermore, the process of the feature adaptive fusion algorithm includes: Based on the aforementioned lightweight model, an adaptive weighted fusion module guided by feature confidence is designed; The confidence level of each spatial location on feature maps at different scales is evaluated through a lightweight subnetwork, and a dynamic weight map is generated. The dynamic weight map is used to guide pixel-level adaptive fusion of multi-scale features.

[0029] Specifically, the implementation process of this embodiment includes: The feature adaptive fusion algorithm is based on an improved lightweight YOLO framework, and a feature confidence-guided adaptive weighted fusion module is designed. This module uses a lightweight sub-network to evaluate the confidence of each spatial location on feature maps of different scales in relation to the current detection task in real time, and generates a dynamic weight map to guide pixel-level adaptive fusion of multi-scale features. This significantly enhances the weak feature response of small targets, improves the recognition rate, and suppresses interference features from complex clutter backgrounds. It achieves a balance between accuracy and speed on embedded edge computing devices, with a positioning error ≤5cm.

[0030] Furthermore, the process of employing spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, includes: Simultaneously acquire ground-penetrating radar data, GPS-RTK positioning data, and IMU inertial measurement data; The radar images within a continuous time window, along with their corresponding timestamps and antenna movement speed, are treated as a spatiotemporal sequence and input into the Transformer model for identification. By employing a tightly coupled Kalman filter algorithm and fusing GPS-RTK positioning data with IMU data, the precise pose of the radar antenna at each data acquisition moment is obtained.

[0031] Furthermore, the process of employing spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, also includes: For the identified target, its local coordinates relative to the antenna phase center are calculated based on its pixel coordinates in the image and the radar wave velocity model. Using the precise pose, the local coordinates are transformed to the global geodetic coordinate system through three-dimensional coordinate transformation.

[0032] Specifically, the implementation process of this embodiment includes: By employing a Transformer architecture for spatiotemporal sequence modeling, combined with GPS+IMU multi-source sensor fusion and GPS-RTK centimeter-level positioning, a closed-loop "detection-identification-positioning" process is achieved, including: By constructing a multi-source sensor array integrating ground-penetrating radar, GPS-RTK centimeter-level positioning module, and IMU inertial measurement unit, the Transformer deep learning architecture is applied to the direct modeling and real-time analysis of ground-penetrating radar spatiotemporal sequence signals, enabling synchronous detection and intelligent identification of underground targets. Furthermore, through innovative spatiotemporal coding and sensor fusion algorithms, the identified target features are bound to high-precision spatiotemporal pose information in real time, automatically generating a 3D distribution map of the target with absolute geographic coordinates. This achieves a fully automated closed-loop process from data acquisition and target recognition to centimeter-level precise positioning without the need for manual post-processing intervention.

[0033] By employing a Transformer architecture for spatiotemporal sequence modeling, combined with GPS+IMU multi-source sensor fusion and GPS-RTK centimeter-level positioning, a closed-loop "detection-identification-positioning" process is achieved, including: Synchronous data acquisition and preprocessing: The control unit synchronously triggers the ground-penetrating radar to emit electromagnetic waves and receives the echoes, generating a ground-penetrating radar result map of A-scan signal. Simultaneously, it reads the latitude, longitude, and elevation coordinates output by the GPS-RTK module, with an accuracy error controlled within 1 cm, as well as the triaxial acceleration and angular velocity output by the IMU. The raw radar signal undergoes preprocessing such as gain adjustment and filtering, and is organized into a continuous B-scan image frame sequence along the survey line.

[0034] like Figure 3 As shown, the iRTK5 device's GPS positioning information is read and recorded in real time according to this process.

[0035] Unlike processing a single image, this embodiment of spatiotemporal sequence modeling and real-time recognition based on Transformer takes multiple frames of B-scan images within a continuous time window, their corresponding timestamps, and the antenna movement speed from the IMU as a whole spatiotemporal sequence input into the improved Transformer model.

[0036] This system employs GPS-RTK+IMU multi-source fusion and centimeter-level pose estimation to address the potential loss of lock or drift of GPS-RTK signals in obstructed environments. It utilizes a tightly coupled Kalman filter algorithm to fuse absolute GPS-RTK position, high-frequency inertial data from the IMU, and radar antenna velocity constraints in real time. During short-term GPS signal interruptions, the IMU is used for high-precision dead reckoning, continuously outputting a stable and continuous six-DOF pose of the antenna phase center at the moment of each radar data acquisition frame.

[0037] The system generates a closed-loop "detection-identification-localization" process, automatically mapping pixel coordinates to geographic coordinates. Coordinate Mapping: For each target identified in the Transformer-based spatiotemporal sequence modeling and real-time recognition, the local polar coordinates of the target relative to the antenna phase center are calculated based on the pixel coordinates in the image and the radar wave velocity model. Spatial Transformation: Utilizing the precise position and attitude of the antenna when acquiring the data frame, obtained through GPS-RTK+IMU multi-source fusion and centimeter-level pose estimation, the local coordinates of the target are transformed to the global geodetic coordinate system via 3D coordinate transformation. Result Generation: The system outputs all identified and located targets in real-time in a standardized data structure with timestamps, target categories, confidence levels, and 3D geographic coordinates. This data can be visualized in real-time on tablets or AR glasses, and can also automatically generate 3D point cloud maps or profiles of underground targets with geographic information.

[0038] Furthermore, after target identification and before calculating local coordinates, the process also includes: An accurate parameter extraction module based on nonmaximum suppression and linear interpolation post-processing is embedded to refine the recognition results.

[0039] Specifically, the implementation process of this embodiment includes: By employing a Transformer architecture for spatiotemporal sequence modeling, combined with GPS+IMU multi-source sensor fusion and GPS-RTK centimeter-level positioning, a closed-loop "detection-identification-positioning" process is achieved, including: Between the identification and coordinate mapping steps, a precise parameter extraction module based on "non-maximum suppression-linear interpolation" collaborative post-processing is embedded. This improves the accuracy of quantitative description of underground foreign objects by using the preliminary detection results output by the front-end deep learning model.

[0040] Before outputting the positioning coordinates, a linear regression equation is established for the GPS coordinates and time of the last 10 frames based on the driving speed of the ground penetrating radar to achieve positioning correction and ensure positioning accuracy.

[0041] Furthermore, the process of controlling the actuator to spray markings on the ground is executed by a ground autonomous marking and feedback execution unit, which includes: A high-precision servo drive module is used to control the movement of the actuator in a two-dimensional plane; Programmable spraying actuator for spraying marking paint; The central controller is used to receive instructions and coordinate the driving and spraying actions.

[0042] Specifically, the implementation process of this embodiment includes: When a foreign object is detected, the system outputs location data and sprays a mark on the ground, including: An additional ground-based autonomous marking and feedback execution unit is added to achieve the ultimate closed loop from "digital identification and positioning" to "physical on-site marking." The execution unit is electrically connected to the edge computing and control unit and mechanically integrated onto the mobile detection platform. It mainly includes a high-precision servo drive module to control the movement of the actuator in a two-dimensional plane; a programmable spraying actuator using a micro-orifice piezoelectric nozzle or a micro pneumatic valve, supporting various fast-drying, environmentally friendly, and high-visibility marking coatings; and a central controller that receives instructions from the main edge computing unit and coordinates the driving and spraying actions.

[0043] In this embodiment, a computer terminal device is provided, including: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-described intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion.

[0044] In this embodiment, a computer-readable storage medium is also provided, on which a computer program is stored. When the computer program is executed by a processor, it implements the steps of the above-described intelligent detection and positioning method for underground hard foreign objects based on dynamic feature fusion.

[0045] like Figure 1 As shown, this embodiment includes a large-scale farmland fine-grained underground foreign object survey system consisting of ground-penetrating radar image reading, ground-penetrating radar image detection, and soil foreign object location.

[0046] The execution subject of this embodiment is an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion. It can also be an underground foreign object survey system before large-scale farmland fine reclamation or other methods, and is not limited here. This embodiment uses an underground foreign object survey system before large-scale farmland fine reclamation as an example for illustration.

[0047] The detection platform employs a tractor-mounted integrated detection sled. The sled houses a radar system equipped with one low-to-medium frequency (200MHz) and one high-to-medium frequency (500MHz) antenna array, used for surveying larger obstacles in deep soil (1.5-3 meters) and detecting fine foreign objects in shallow soil (0-1.5 meters). The positioning and attitude determination unit, integrated into the agricultural machinery's automatic driving system, utilizes a consumer-grade IMU and a GNSS-RTK receiver, providing horizontal accuracy of ±2 cm and vertical accuracy of ±3 cm. Vehicle speed information is acquired via the tractor's own CAN bus as an auxiliary observation value. The edge computing and storage unit is equipped with a rugged industrial computer with a built-in high-performance GPU card.

[0048] The automatic marking unit uses a pneumatic dot spray marking system that can spray environmentally friendly biodegradable lime powder or washable paint to form temporary but clear markings on cultivated soil.

[0049] In this example, the multi-frequency data fusion acquisition system simultaneously acquires dual-frequency radar data. Through time-depth calibration, the deep structural information obtained from the 200MHz data and the shallow, fine features obtained from the 500MHz data are fused at the feature level and input together into the recognition network. For model training and recognition of agricultural foreign objects, a large amount of labeled ground-penetrating radar data, including categories such as stones, bricks, metal, and cavities, is used to specifically train the Transformer spatiotemporal sequence model described in this invention. The model specifically learns to distinguish subtle feature differences between easily confused targets such as natural gravel and artificial concrete blocks, and wet clay clumps and plastic pipes.

[0050] In this example, since farmland operations typically involve regular reciprocating plowing paths, the system incorporates path-constrained Kalman filtering in the GPS-RTK+IMU multi-source fusion and centimeter-level pose estimation steps. In addition to sliding window regression smoothing, prior information about the straight lines of the work rows is utilized to further correct for sporadic jumps that may still exist in open farmland, outputting an extremely smooth trajectory line that conforms to agronomical practices.

[0051] In this example, during parameter refinement, the system not only calculates the precise depth but also automatically categorizes targets by burial depth into an emergency layer (0-30 cm, directly affecting plowing safety), a tillage layer (30-60 cm), and a lower layer (below 60 cm). Based on the analysis and judgment of target categories such as "large rocks" and "fine metal," the system automatically assigns high, medium, and low risk levels.

[0052] In this example, the marking unit sprays markers of different colors or sizes based on the risk level. Simultaneously, the system generates and updates a "heat map of underground foreign object distribution" for the entire field in real time, transmitting it wirelessly to the farm management cloud platform. This map can be directly imported into the tillage control system, providing a precise navigation data layer for subsequent selective deep plowing or targeted clearing.

[0053] In this example, a survey was conducted on a 20-hectare plot of uncultivated land. A tractor operated at a constant speed of 12 km / h, completing the survey in a single operation. This embodiment effectively identifies stones, bricks, and metal objects larger than 5 cm in diameter, achieving a comprehensive identification rate of over 90% for shallow, high-risk foreign objects. The generated distribution map clearly guides subsequent engineering machinery for precise and efficient targeted obstacle removal, avoiding the high cost and soil structure damage of traditional deep tillage, thus achieving precision agriculture-style "targeted" soil management.

[0054] This invention provides an intelligent detection and location method for underground hard foreign objects based on dynamic feature fusion. By introducing a dynamic feature fusion mechanism based on an improved lightweight model, it significantly improves the automatic identification capability and anti-interference ability of small underground hard foreign objects, while meeting the computational requirements of real-time processing by edge devices. Combining Transformer spatiotemporal sequence modeling and multi-source sensor fusion positioning technology, it achieves full automation of the entire process of underground target detection, intelligent identification, and centimeter-level precise positioning. Finally, by integrating ground autonomous marking units, a complete closed loop from digital information to physical marking is formed, completely avoiding errors introduced by manual secondary marking and greatly improving the overall efficiency and accuracy of on-site operations.

[0055] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent detection and location of underground hard foreign objects based on dynamic feature fusion, characterized in that, Includes the following steps: Acquire ground-penetrating radar data and perform image processing; A lightweight model based on an improved deep learning algorithm is used to process and recognize the images in real time. An adaptive feature fusion algorithm is used to improve the recognition rate of small-sized hard foreign objects; By employing spatiotemporal sequence modeling and combining multi-source sensor fusion and localization, a closed loop of detection, identification, and localization is achieved. When a foreign object is detected, the system outputs location data and controls the actuator to spray a mark on the ground.

2. The method according to claim 1, characterized in that, The process of adopting a lightweight model improved based on deep learning algorithms includes: Replace the YOLO11 backbone network with a lightweight backbone network reconstructed from the GhostNetV3 module; Design a lightweight decoupled detection head based on a moving inverted bottleneck; The original neck network is replaced by a multi-scale feature adaptive aggregation network, the core of which is a dynamic weight attention module.

3. The method according to claim 2, characterized in that, The process of real-time processing and recognition of lightweight models also includes: Receive feature maps of different scales from the lightweight backbone network; The feature maps at various scales are analyzed by the miniature neural network in the dynamic weight attention module, and a dynamic weight map is generated. The dynamic weight map is used to perform pixel-level weighted fusion of features at each scale.

4. The method according to claim 1, characterized in that, The process of the feature adaptive fusion algorithm includes: Based on the aforementioned lightweight model, an adaptive weighted fusion module guided by feature confidence is designed; The confidence level of each spatial location on feature maps at different scales is evaluated through a lightweight subnetwork, and a dynamic weight map is generated. The dynamic weight map is used to guide pixel-level adaptive fusion of multi-scale features.

5. The method according to claim 1, characterized in that, The process of using spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, includes: Simultaneously acquire ground-penetrating radar data, GPS-RTK positioning data, and IMU inertial measurement data; The radar images within a continuous time window, along with their corresponding timestamps and antenna movement speed, are treated as a spatiotemporal sequence and input into the Transformer model for identification. By employing a tightly coupled Kalman filter algorithm and fusing GPS-RTK positioning data with IMU data, the precise pose of the radar antenna at each data acquisition moment is obtained.

6. The method according to claim 5, characterized in that, The process of using spatiotemporal sequence modeling, combined with multi-source sensor fusion and localization, also includes: For the identified target, its local coordinates relative to the antenna phase center are calculated based on its pixel coordinates in the image and the radar wave velocity model. Using the precise pose, the local coordinates are transformed to the global geodetic coordinate system through three-dimensional coordinate transformation.

7. The method according to claim 5, characterized in that, After target identification and before local coordinate calculation, the process further includes: An accurate parameter extraction module based on nonmaximum suppression and linear interpolation post-processing is embedded to refine the recognition results.

8. The method according to claim 1, characterized in that, The process of controlling the actuator to spray markings on the ground is performed by a ground autonomous marking and feedback execution unit, which includes: A high-precision servo drive module is used to control the movement of the actuator in a two-dimensional plane; Programmable spraying actuator for spraying marking paint; The central controller receives instructions and coordinates the driving and spraying actions.

9. A computer terminal device, characterized in that, include: One or more processors; A memory, coupled to the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors perform the steps of the method as described in any one of claims 1-8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method as described in any one of claims 1-8.

Images