A method and system for merging multi-target identification results from a three-dimensional ground-penetrating radar

Abstract

This invention provides a method and system for merging multi-target recognition results from three-dimensional ground-penetrating radar (GPR), belonging to the field of road engineering pavement maintenance technology. The invention first constructs a GPR dataset containing two categories of labels: "bridge" and "structural layer," and trains a YOLOv8n model to recognize targets. Then, it uses non-maximum suppression to merge redundant detection boxes from each map. Next, it determines layer stability based on multi-channel consistency: candidate layer locations are extracted according to a preset depth sampling point range, sorted, and assigned layer order categories. Layers with categories shared by at least two channels at the same location are identified as stable layers. The best channel is then selected through scoring. Finally, based on the detection results of the best channel, pavement structure detection boxes overlapping with bridges in the road number interval are modified to bridges and deleted, thus filtering bridge interference. This invention effectively fuses multi-channel recognition results, reduces false detection rate, and improves the reliability of layer recognition.

Description

Technical Field

[0001] This invention belongs to the field of road engineering pavement maintenance technology, specifically relating to a method and system for merging multi-target identification results from three-dimensional ground penetrating radar. Background Technology

[0002] The thickness of the asphalt pavement structural layer is a crucial indicator for evaluating the construction quality of asphalt pavements. If the thickness of the pavement structural layer does not meet the design requirements, it will lead to insufficient load-bearing capacity of the pavement structure. Furthermore, areas with excessive thickness variations will experience stress concentration, accelerating pavement damage. Therefore, conducting asphalt pavement structural layer thickness testing is of great significance for monitoring road construction quality.

[0003] Existing methods for detecting the thickness of pavement structural layers mainly include core sampling, nucleus-free densitometers, nuclear densitometers, and ground-penetrating radar (GPR). However, core sampling, nucleus-free densitometers, and nuclear densitometers are limited by their discontinuous nature and operational complexity, making them insufficient to meet the rapid and continuous testing requirements of modern roads. In contrast, GPR, as a rapid and continuous non-destructive testing technology, has been widely used for pavement structural thickness detection. Its thickness detection principle is based on the reflection signals of electromagnetic waves at different material interfaces, calculating the thickness of the structural layer by analyzing the two-way travel time of the reflected signals.

[0004] In recent years, image-based topometry has become a research hotspot. However, existing methods mostly rely on single two-dimensional images, resulting in poor stability and a high false detection rate. In contrast, three-dimensional ground-penetrating radar can acquire two-dimensional profile images from multiple channels, each channel reflecting different response characteristics of the same spatial location. How to utilize the complementary information between multiple channels for fusion decision-making to improve recognition reliability is currently a key research focus.

[0005] Furthermore, bridge structures are frequently present in real-world road environments. The dense steel mesh within bridges generates multiple reflections with strong amplitude and a hyperbolic shape. These signals are highly similar to the layer reflection characteristics of pavement structural layers in radar images, easily leading deep learning models to misidentify bridges as pavement structural layers, resulting in significant errors in thickness detection. Existing methods lack effective filtering mechanisms for bridge interference, and false detections directly affect the reliability of subsequent thickness calculations.

[0006] Therefore, there is an urgent need for a layer identification result merging method that can integrate multi-channel identification results and effectively eliminate bridge interference, so as to improve the accuracy and robustness of pavement structure layer thickness detection. Summary of the Invention

[0007] This invention provides a method and system for merging multi-target identification results from a three-dimensional ground-penetrating radar, aiming to solve the problems of poor stability of existing single-channel identification results and the identification of interference layers in bridge structures.

[0008] To solve the above-mentioned technical problems, the present invention is implemented as follows: A method for merging multi-target identification results from a three-dimensional ground-penetrating radar includes the following steps: Step S1: Construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". Use this dataset to train the pre-trained recognition model. Input the collected ground-penetrating radar data into the trained recognition model and output the recognition results. Step S2: The recognition results of each map are merged based on the nonmaximum suppression method, and the most representative recognition results are retained; Step S3: For the area identified as "structural layer" in the identification results, the range of depth sampling points for each layer of the road structure is preset. For multi-channel ground penetrating radar data, candidate layer points for each channel are extracted according to the preset depth sampling point range. After sorting by depth, layer order categories are assigned. Layer order categories shared by at least two channels at the same location are determined as stable layers. Step S4: Count the number of stable layers in each channel, select the channel with the same number of stable layers as the actual number of road layers as the candidate channel for the best channel, use the average confidence of all layers in the candidate channel as the score of the candidate channel, and select the channel with the highest score and the smallest channel number as the best channel. Step S5: Based on the target detection result map of the best channel, calculate the channel number interval of each target detection box. If the channel number interval of the target detection box of category "structure layer" overlaps with the channel number interval of the target detection box of category "bridge", then change the category of the target detection box to "bridge". Iterate through all target detection boxes of category "structure layer"; then delete all target detection boxes of category "bridge" and keep the target detection boxes of category "structure layer" to complete the merging of recognition results.

[0009] As a preferred improvement, each image in the dataset has a resolution of 320×320, and the images are named according to the following rules: road segment name_file name_left lane number of the image_right lane number of the image_channel number.

[0010] As a preferred improvement, the recognition model is a pre-trained YOLOv8n model.

[0011] As a preferred improvement, the non-maximum suppression method in step S2 is specifically as follows: set the IoU threshold to 0.5, and when the IoU of two target detection boxes exceeds this threshold, delete the target detection box with lower confidence and retain the target detection box with higher confidence.

[0012] As a preferred improvement, the total number of depth sampling points in step S3 is 256, and the theoretical sampling point number range corresponding to each layer of the road structure is: 30~60 for the first layer, 80~120 for the second layer, and 150~180 for the third layer.

[0013] As a preferred improvement, the channel score in step S4 is the average confidence level of all layers in the channel.

[0014] As a preferred improvement, in step S5, the channel number range of the target detection box is determined in the following way: In the formula, , These are the start and end channel numbers corresponding to the target detection box, respectively; Left lane number for the image; , These represent the left and right pixel boundaries of the target detection bounding box, respectively. The channel number corresponding to the pixel is calculated using the following formula: In the formula, The right sidebar number is 'W', and 'W' is the image width.

[0015] A system for performing the above-described method for merging multi-target identification results from a three-dimensional ground-penetrating radar includes: The identification module is used to construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". This dataset is used to train the pre-trained identification model. The collected ground-penetrating radar data is input into the trained identification model, and the identification results are output. The recognition result merging module is used to merge the recognition results of each map based on the nonmaximum suppression method, retaining the most representative recognition results; The layer stability determination module is used to determine the depth sampling point range of each layer of the road structure for areas identified as "structural layer" in the identification results. For multi-channel ground penetrating radar data, it extracts candidate layer points of each channel according to the preset depth sampling point range, sorts them by depth and assigns them to layer order categories. Layer order categories shared by at least two channels at the same location are determined as stable layers. The optimal channel selection module is used to count the number of stable layers in each channel, select the channel whose number of stable layers is equal to the number of actual road surface layers as the candidate channel for the optimal channel, use the average confidence of all layers in the candidate channel as the score of the candidate channel, and select the channel with the highest score and the smallest channel number as the optimal channel. The result filtering module is used to calculate the channel number interval of each target detection box based on the target detection result map of the best channel. If the channel number interval of the target detection box of category "structure layer" overlaps with the channel number interval of the target detection box of category "bridge", the category of the target detection box is changed to "bridge". All target detection boxes of category "structure layer" are traversed. Then, all target detection boxes of category "bridge" are deleted, and the target detection boxes of category "structure layer" are kept to complete the merging of recognition results.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention trains the YOLOv8n model by constructing a dataset containing two types of labels: “bridge” and “strument layer”, which can simultaneously identify bridge and road surface structure layers, providing a basis for subsequent merging; (2) This invention utilizes multi-channel ground penetrating radar data to extract candidate layer locations through a preset depth sampling point range, and determines stable layers based on multi-channel consistency, effectively eliminating single-channel random noise and false detections, and improving the reliability of layer identification; (3) The present invention automatically selects the best channel through the channel scoring mechanism and filters out the erroneous layer detection in the bridge area based on the channel number interval overlap detection, thereby avoiding the interference of the bridge structure on the road surface layer identification and finally outputting a clean road surface structure layer identification result, which is convenient to import into the three-dimensional ground penetrating radar data processing software for visualization analysis. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein: Figure 1 A schematic diagram of multi-target ground-penetrating radar data map annotation; Figure 2 One of the illustrations of image naming rules; Figure 3 Image naming rules diagram two; Figure 4 This is a schematic diagram illustrating the YOLOv8n model's identification results for "structural layers". Figure 5 This is a schematic diagram of the YOLOv8n model's identification of the "bridge" structure. Figure 6 This is a schematic diagram of the redundant detection box before nonmaximum suppression. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figures 1-6 As shown, this embodiment provides a method for merging multi-target identification results from a three-dimensional ground-penetrating radar, including the following steps: Step S1: Construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". Use this dataset to train the pre-trained recognition model. After training, input the collected ground-penetrating radar data into the recognition model and output the recognition results.

[0020] The dataset includes several ground-penetrating radar (GPR) images, each with a resolution of 320×320. Due to significant differences in the dielectric constants of materials between road structural layers (especially given the high standards of construction and material control in highways), electromagnetic waves emitted by GPR will generate strong reflection signals when they encounter these interlayer interfaces. These reflection signals are characterized by smooth, continuous, and high-brightness phase axes on the images. The greater the difference in dielectric constants between layers (i.e., the reflection coefficient), the more pronounced this characteristic becomes. For bridges, due to the presence of strong reflective targets such as reinforcing bars, typical characteristics include strong amplitude, a distinct hyperbolic shape, phase reversal, and multiple reflections, often arranged at equal intervals. Based on these differences, the Labelimg annotation tool can be used to label road surface structures and bridge structures separately on the images. The label categories include "Bridge" and "Structural Layer," where "Structural Layer" corresponds to road surface structures and "Bridge" corresponds to bridge structures. A schematic diagram of multi-target GPR data image annotation is shown below. Figure 1 As shown.

[0021] In the dataset, each image is named according to the following rule: Road Segment Name_File Name_Left Lane Number_Right Lane Number_Channel Number. For example, the image name is: 1_01_37801_38120_CH01, where 1 is the road segment name; 01 is the file name; 37801 is the left lane number; 38120 is the right lane number; and CH01 is the channel number. This unified naming rule facilitates grouping images with the same lane number and provides location information for subsequent result merging. A diagram illustrating the image names is shown below. Figure 2 as well as Figure 3 As shown.

[0022] Specifically, the recognition model is a pre-trained YOLOv8n model with a map 50 metric of 97.4% and a map 50-90 metric of 64.6%, achieving a good balance between recognition accuracy and speed. The architecture and pre-training process of the YOLOv8n model can be implemented using conventional techniques in the field, and will not be elaborated upon in this embodiment.

[0023] The pre-trained YOLOv8n model was retrained using the dataset described above. After training, it can be well applied to the multi-object detection scenario in this invention, outputting recognition results for road surface structures and bridge structures. A schematic diagram of the recognition results is shown below. Figure 4 as well as Figure 5 As shown, Figure 4 In the text, “loyer thickness 0.87”, “loyer thickness 0.84”, and “loyer thickness 0.80” indicate that the recognition results of the three target detection boxes are all “structural layer”, with corresponding confidence levels of 0.87, 0.84, and 0.80, respectively. Figure 5 In the text, "Bridge 0.69" indicates that the object detection box was identified as "bridge", with a confidence level of 0.69.

[0024] Step S2: The recognition results of each map are merged based on the nonmaximum suppression method, and the most representative recognition results are retained.

[0025] The recognition result for each map includes object detection boxes, category labels for the object detection boxes, and corresponding confidence scores. For each map recognition process, the YOLOv8n model may output multiple overlapping or neighboring object detection boxes, such as... Figure 6 As shown. To address this issue, the present invention uses non-maximum suppression to remove redundant target detection boxes. Specifically, the method employed is as follows: an IoU (Intersection over Union) threshold of 0.5 is set. When the IoU between two target detection boxes exceeds this threshold, the target detection box with lower confidence is deleted, while the target detection box with higher confidence is retained. This method can retain the most representative detection results.

[0026] Step S3: For the area identified as "structural layer" in the identification results, the range of depth sampling points for each layer of the road structure is preset. For multi-channel ground penetrating radar data, candidate layer points for each channel are extracted according to the preset depth sampling point range. After sorting by depth, layer order categories are assigned. Layer order categories shared by at least two channels at the same location are determined as stable layers.

[0027] Depth sampling points represent discrete signal points acquired by ground-penetrating radar (GPR) at fixed time intervals along the vertical depth direction. This parameter determines the digital resolution of the radar image in the longitudinal direction. The depth positions corresponding to different layers can be identified and tracked using their sampling point numbers. The theoretical sampling point number range for each layer of the road structure is determined based on empirical values ​​or known data. In this embodiment, the GPR includes a total of 256 sampling points, of which: The theoretical sampling point number range for the first layer of the road structure is 30-60. The theoretical sampling point number range for the second layer of the pavement structure is 80-120. The theoretical sampling point number range for the third layer of the pavement structure is 150-180.

[0028] It should be noted that the above-mentioned number and range of layers are merely examples. In this embodiment, the three-layer structure has already reached the roadbed depth, therefore no further layers are involved. In other embodiments, the number of layers in the pavement structure can be adjusted according to the specific road structure and the detection parameters of the detection radar.

[0029] Specifically: Based on the pre-given range of sampling point numbers corresponding to each layer of the pavement structure, local amplitude maxima points within the corresponding range are extracted from each data point in each channel as candidate layer points; all candidate layer points in the same channel and channel are sorted in ascending order of sampling point number and assigned category labels of first layer, second layer, third layer, etc.; for different channels with the same measurement point location (same channel number), the number of channels with the same category label is counted. If the number of channels is ≥2, the layer is determined to be a stable layer; otherwise, it is determined to be an unstable layer and is removed.

[0030] Step S4: Count the number of stable layers in each channel, select the channel with the same number of stable layers as the actual number of road layers as the candidate channel for the best channel, use the average confidence level of all layers in the candidate channel as the score of the candidate channel, and select the channel with the highest score and the smallest channel number as the best channel.

[0031] It should be noted that the actual number of road surface layers is determined in advance based on the design and construction data of the road to be tested or the results of on-site investigation.

[0032] During the selection of candidate channels, there may be situations where the number of stable layers in multiple channels meets the actual number of road surface layers. In this case, the score of each channel is calculated, and the channel with the highest score is selected as the best channel. If multiple channels have the same score, the channel with the smallest channel number is selected as the best channel. The channel score is the average confidence level of all layers in the channel (the average confidence level of all identified "structural layer" annotation boxes is taken).

[0033] If the number of stable layers in all channels is not equal to the actual number of road surface layers, it indicates that the identification result at that measuring point is unreliable. It can be manually verified or other methods can be used for processing as needed.

[0034] Step S5: Based on the target detection result map of the best channel, calculate the channel number interval of each target detection box. If the channel number interval of the target detection box of category "structure layer" overlaps with the channel number interval of the target detection box of category "bridge", then change the category of the target detection box to "bridge". Iterate through all target detection boxes of category "structure layer"; then delete the target detection boxes of category "bridge" and keep the target detection boxes of category "structure layer" to complete the merging of recognition results.

[0035] The channel number range of the target detection box is determined in the following way: In the formula, , These are the start and end channel numbers corresponding to the target detection box, respectively; Left lane number for the image; , These represent the left and right pixel boundaries of the target detection bounding box, respectively. The channel number corresponding to the pixel is calculated using the following formula: In the formula, The right sidebar number is W, and W is the image width (320 pixels).

[0036] Through the above steps, this invention can effectively merge the multi-channel identification results of 3D ground-penetrating radar, eliminate unstable layers and bridge interference, and finally output reliable pavement structure layer identification results. The retained records (txt format files) are written to the "Filtered" directory according to the original file path to filter the bridge results for subsequent import into 3D ground-penetrating radar data processing software.

[0037] This embodiment also provides a system for performing the above-described method for merging multi-target identification results from a three-dimensional ground-penetrating radar, comprising: The identification module is used to construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". This dataset is used to train the pre-trained identification model. The collected ground-penetrating radar data is input into the trained identification model, and the identification results are output. The recognition result merging module is used to merge the recognition results of each map based on the nonmaximum suppression method, retaining the most representative recognition results; The layer stability determination module is used to determine the depth sampling point range of each layer of the road structure for areas identified as "structural layer" in the identification results. For multi-channel ground penetrating radar data, it extracts candidate layer points of each channel according to the preset depth sampling point range, sorts them by depth and assigns them to layer order categories. Layer order categories shared by at least two channels at the same location are determined as stable layers. The optimal channel selection module is used to count the number of stable layers for each channel, determine whether the number of stable layers is equal to the number of actual road surface layers, and if so, calculate the score for each channel and select the channel with the highest score as the optimal channel; otherwise, mark the measurement point as an unreliable identification area. The result filtering module is used to calculate the channel number interval of each target detection box based on the target detection result map of the best channel. If the channel number interval of the target detection box of category "structure layer" overlaps with the channel number interval of the target detection box of category "bridge", the category of the target detection box is changed to "bridge". All target detection boxes of category "structure layer" are traversed. Then, all target detection boxes of category "bridge" are deleted, and the target detection boxes of category "structure layer" are kept to complete the merging of recognition results.

[0038] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of the present invention.

Claims

1. A method for merging multi-target recognition results of a three-dimensional ground penetrating radar, characterized in that, Includes the following steps: Step S1: Construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". Use this dataset to train the pre-trained recognition model. Input the collected ground-penetrating radar data into the trained recognition model and output the recognition results. Step S2: The recognition results of each map are merged based on the nonmaximum suppression method, and the most representative recognition results are retained; Step S3: For the area identified as "structural layer" in the identification results, a preset range of depth sampling points for each layer of the road structure is established. For multi-channel ground-penetrating radar data, candidate layer points for each channel are extracted according to the preset depth sampling point range. After sorting by depth, the layer order category is assigned. For different channels with the same channel number, the number of channels with the same category identifier is counted. If the number of channels is not less than a preset threshold, the layer is determined to be a stable layer; otherwise, it is determined to be an unstable layer and is removed. Step S4: Count the number of stable layers in each channel, select the channel with the same number of stable layers as the actual number of road layers as the candidate channel for the best channel, use the average confidence of all layers in the candidate channel as the score of the candidate channel, and select the channel with the highest score and the smallest channel number as the best channel. Step S5: Based on the target detection result map of the best channel, calculate the channel number interval of each target detection box. If the channel number interval of the target detection box with the category "structure layer" overlaps with the channel number interval of the target detection box with the category "bridge", then change the category of the target detection box to "bridge" and traverse all target detection boxes with the category "structure layer". Then delete all object detection boxes in the category of "bridge" and keep the object detection boxes in the category of "structure layer" to complete the merging of recognition results.

2. The method of claim 1, wherein Each image in the dataset has a resolution of 320×320, and the images are named according to the following rules: road segment name_file name_left lane number of the image_right lane number of the image_channel number.

3. The method of claim 1, wherein the method further comprises: The recognition model is a pre-trained YOLOv8n model.

4. The method of claim 1, wherein, The non-maximum suppression method described in step S2 is as follows: set the IoU threshold to 0.

5. When the IoU of two target detection boxes exceeds this threshold, delete the target detection box with lower confidence and retain the target detection box with higher confidence.

5. The method of claim 1, wherein, The number of depth sampling points mentioned in step S3 is 256. The theoretical sampling point number range corresponding to each layer of the road structure is: 30~60 for the first layer, 80~120 for the second layer, and 150~180 for the third layer.

6. The method of claim 1, wherein, The channel score mentioned in step S4 is the average confidence level of all layers in the channel.

7. The method of claim 1, wherein, In step S5, the channel number range of the target detection box is determined in the following way: In the formula, , are the starting lane number and the ending lane number corresponding to the target detection frame, respectively; is the left lane number of the picture; , are the left pixel boundary and the right pixel boundary of the target detection frame, respectively; is the lane number corresponding to the pixel, and is calculated by the following formula: wherein is the picture right channel number and W is the picture width.

8. A system for performing the multi-target identification result merging method of three-dimensional ground-penetrating radar according to any one of claims 1-7, characterized in that, include: The identification module is used to construct a ground-penetrating radar dataset containing two types of labels: "bridge" and "structural layer". This dataset is used to train the pre-trained identification model. The collected ground-penetrating radar data is input into the trained identification model, and the identification results are output. The recognition result merging module is used to merge the recognition results of each map based on the nonmaximum suppression method, retaining the most representative recognition results; The layer stability determination module is used to determine the depth sampling point range for each layer of the road structure in the identification results of the "structural layer". For multi-channel ground penetrating radar data, it extracts the candidate layer points of each channel under each channel according to the preset depth sampling point range, sorts them by depth and assigns them to layer order categories. For different channels with the same channel number, it counts the number of channels with the same category identifier. If the number of channels is not less than the preset threshold, the layer is determined to be a stable layer; otherwise, it is determined to be an unstable layer and is removed. The optimal channel selection module is used to count the number of stable layers in each channel, select the channel whose number of stable layers is equal to the number of actual road surface layers as the candidate channel for the optimal channel, use the average confidence of all layers in the candidate channel as the score of the candidate channel, and select the channel with the highest score and the smallest channel number as the optimal channel. The result filtering module is used to calculate the channel number interval of each target detection box based on the target detection result map of the best channel. If the channel number interval of the target detection box with the category of "structure layer" overlaps with the channel number interval of the target detection box with the category of "bridge", then the category of the target detection box is changed to "bridge" and all target detection boxes with the category of "structure layer" are traversed. Then delete all object detection boxes in the category of "bridge" and keep the object detection boxes in the category of "structure layer" to complete the merging of recognition results.

Images