Intelligent pavement disease patrol method and system based on artificial intelligence

Abstract

The application discloses a kind of based on artificial intelligence's road surface disease intelligent patrol method and system, comprising: through the pose information and relative motion physical parameter group in the same road fragment, the spatial correspondence between unmanned aerial vehicle patrol terminal and vehicle-mounted patrol terminal is established, and the mismatch of long-distance identification result and near-distance review area is reduced;Through the edge link spectrum state matrix, the signal change under unlicensed spectrum is reflected, and physical basis is provided for communication connection selection;Then, combined with the cooperative patrol geometry matrix, the beam pointing and beam switching sequence are determined, so that the suspected disease fragment index remains stable during the relative motion of the double terminals;The unmanned aerial vehicle patrol terminal first carries out road surface image recognition, and then the vehicle-mounted patrol terminal reviews the suspected disease fragment index in the near distance, thereby reducing the probability of missed detection, false detection and repeated inspection, and improving the accuracy and continuity of the road surface disease inspection result.

Description

Technical Field

[0001] This invention relates to the technical field of road surface defect inspection, and in particular to an intelligent road surface defect inspection method and system based on artificial intelligence. Background Technology

[0002] With the increasing workload of highway maintenance, pavement defect inspection is gradually evolving from manual inspection to vehicle-mounted inspection, drone inspection, and intelligent identification. Vehicle-mounted inspection terminals are usually installed on inspection vehicles and can acquire close-range pavement image data while the vehicle is in motion, making them suitable for identifying local defects such as cracks, potholes, subsidence, and repair damage. Drone inspection terminals have the advantages of a large field of view and are less restricted by road traffic conditions, making them suitable for overhead inspections of bridges, ramps, congested sections, emergency repair sections, and construction detour areas. Existing intelligent pavement defect inspection methods are mostly based on pavement image data, using defect identification models to identify defect areas in the pavement image data and combining them with location information to generate inspection results. This type of method can reduce the intensity of manual inspections and improve the inspection range and identification efficiency. However, when drone inspection terminals and vehicle-mounted inspection terminals conduct collaborative inspections, their inspection altitude, shooting direction, travel speed, and relative position are constantly changing. The suspected disease locations obtained by the drone inspection terminal need to be transmitted to the vehicle-mounted inspection terminal in a timely manner for verification at close range. Existing collaborative inspection methods usually focus on inspection path arrangement, image acquisition, and disease identification, but lack unified processing of the relative orientation, movement relationship, and wireless communication status between the drone inspection terminal and the vehicle-mounted inspection terminal. This leads to problems such as positioning offset, repeated verification, or missed verification of suspected disease locations during transmission and verification.

[0003] For example, CN117538327A discloses a road defect detection system and method for UAV-inspection vehicle collaboration. The system includes a positioning module, a data acquisition module, a docking module, a communication module, and a computing module. It can locate the UAV and the inspection vehicle's roof platform, acquire road defect images and point cloud data, and transmit the road defect features acquired by the UAV and the inspection vehicle back to the ground. Then, it combines the computing module to calculate the optimal inspection path and identify defect data. However, this method mainly focuses on the path collaboration, docking take-off and landing, image and point cloud acquisition, and ground transmission between the UAV and the inspection vehicle. Its communication module focuses on command transmission and defect feature transmission. It does not establish a side-link spectrum state matrix for the unauthorized spectrum signal state between the UAV inspection terminal and the vehicle-mounted inspection terminal, nor does it organize the pose relationship and relative motion physical parameters of the two terminals into a collaborative inspection geometric matrix. Furthermore, it does not determine a beam management parameter group containing beam pointing and beam switching order based on the above geometric matrix and spectrum state matrix. Therefore, the method disclosed in CN117538327A still has the problem that the side link communication relationship is difficult to match stably when multiple terminals move relative to each other during actual inspections, and the suspected defect segment index is difficult to transmit to the vehicle inspection terminal in a timely and accurate manner.

[0004] For example, CN119723290A discloses a road defect detection method and system based on deep learning and ground-air collaboration. This scheme can identify inspection data such as road surface crack images through deep learning models and improve the defect detection coverage by utilizing collaborative data from the ground and air ends. Compared with single vehicle-mounted inspection or single UAV inspection, it has certain improvements in multi-source perception and defect identification. However, the focus of this method is still on the processing of defect images or related detection data under ground-air collaboration conditions. It does not further disclose the use of suspected defect segmentation index as the verification trigger object between UAV inspection terminal and vehicle-mounted inspection terminal, nor does it disclose the formation of sidelink communication relationship for defect verification tasks under unlicensed spectrum based on sidelink spectrum status, beam pointing and beam switching order. Therefore, the method disclosed in CN119723290A still struggles to solve the problem of accurate connection between the long-range identification results of UAVs and the close-range verification area of ​​vehicles in actual road patrols. Especially in urban elevated roads, bridge obstructions, tunnel exits, and multi-terminal simultaneous patrol scenarios, there is a lack of stable communication constraints and spatial segmentation constraints between aerial identification results and ground verification actions.

[0005] In summary, given that existing technologies for detecting road defects using drones, vehicles, and ground-air collaboration still suffer from problems such as unstructured dual-terminal pose relationships, lack of participation of unlicensed spectrum states in sidelink communication relationships, and insufficient coordination between suspected defect segment indexing and vehicle-mounted close-range verification, this invention is proposed. Therefore, the problem this invention aims to solve is how to stably transmit suspected defect segments identified by the drone at a distance to the vehicle-mounted patrol terminal under conditions of relative motion between the drone and vehicle-mounted patrol terminals and changes in unlicensed spectrum states, so that the vehicle-mounted patrol terminal can complete close-range verification within the same road segment.

[0006] To address the aforementioned issues, this invention provides an intelligent road surface defect inspection method and system based on artificial intelligence. It establishes a dual-terminal spatial relationship through road segmentation identification and a collaborative inspection geometric matrix, reflects the unauthorized spectrum signal status through a side-link spectrum state matrix, and forms a side-link communication relationship through beam management parameter groups. Furthermore, a drone inspection terminal generates a suspected defect segment index and triggers a vehicle-mounted inspection terminal for close-range verification and identification, thereby improving the continuity and accuracy of suspected defect transmission, location, and verification in intelligent road surface defect inspection. Summary of the Invention

[0007] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0008] In view of the aforementioned existing problems, the present invention is proposed.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides an intelligent road surface defect inspection method based on artificial intelligence, comprising: merging the pose information and relative motion physical parameter groups of UAV inspection terminals and vehicle-mounted inspection terminals within the same road segment to generate road segment identifiers and collaborative inspection geometric matrices; A sidelink spectrum state matrix is ​​generated based on the unlicensed spectrum signal state between the UAV patrol terminal and the vehicle-mounted patrol terminal; Based on the collaborative inspection geometry matrix and the side link spectrum state matrix, a beam management parameter group containing beam pointing and beam switching order is determined, and a side link communication relationship is formed according to the beam management parameter group. The road surface image data of the road segment is identified by the defect identification model in the UAV inspection terminal, a suspected defect segment index is generated, and the index is sent to the vehicle-mounted inspection terminal through the side link communication relationship. The vehicle-mounted inspection terminal acquires close-range road surface image data within the road segment defined by the suspected defect segment index, performs verification and identification, and generates intelligent road defect inspection results.

[0010] Secondly, this invention provides an intelligent road surface defect inspection system based on artificial intelligence, comprising a drone inspection terminal, a vehicle-mounted inspection terminal, and a sidelink communication module; wherein: The UAV inspection terminal includes a first pose processing module, a spectrum status processing module, a beam management module, and a defect identification module; the vehicle-mounted inspection terminal includes a second pose processing module, a close-range verification module, and an inspection result generation module. The first pose processing module is communicatively connected to the second pose processing module. The two modules merge the pose information and relative motion physical parameter groups within the same road segment to generate road segment identifiers and collaborative inspection geometric matrices. The spectrum status processing module is connected to the side link communication module. The spectrum status processing module generates a side link spectrum status matrix based on the unlicensed spectrum signal status between the UAV patrol terminal and the vehicle-mounted patrol terminal. The beam management module is connected to the first pose processing module, the second pose processing module, the spectrum state processing module, and the side link communication module respectively. Based on the cooperative inspection geometry matrix and the side link spectrum state matrix, the beam management module determines a beam management parameter group that includes beam pointing and beam switching order, and controls the side link communication module to form a side link communication relationship according to the beam management parameter group. The defect identification module is connected to the side link communication module. The defect identification module identifies the road surface image data of the road segment, generates a suspected defect segment index, and sends it to the vehicle-mounted inspection terminal via the side link communication module. The close-range verification module is connected to the inspection result generation module. The close-range verification module obtains close-range road surface image data within the road segment defined by the suspected defect segment index and performs verification and identification. The inspection result generation module generates intelligent inspection results for road surface defects based on the verification and identification results.

[0011] The beneficial effects of this invention are as follows: This invention establishes a spatial correspondence between UAV patrol terminals and vehicle-mounted patrol terminals by merging pose information and relative motion physical parameter groups within the same road segment, reducing mismatches between long-distance identification results and close-range verification areas; it reflects signal changes under unlicensed spectrum through the sidelink spectrum state matrix, providing a physical basis for communication connection selection; and it combines the collaborative patrol geometry matrix to determine beam pointing and beam switching order, ensuring stable transmission of suspected defect segment indexes during the relative motion of the two terminals; the UAV patrol terminal first performs road image recognition, and the vehicle-mounted patrol terminal then performs close-range verification according to the suspected defect segment index, thereby reducing the probability of missed detections, false detections, and duplicate patrols, and improving the accuracy and continuity of road defect patrol results. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of the present invention, the 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 This is a flowchart illustrating the intelligent road surface defect inspection method based on artificial intelligence as shown in this invention. Figure 2 This is a schematic diagram of the module distribution of the AI-based intelligent road surface defect inspection system shown in this invention. Detailed Implementation

[0013] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0014] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0015] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0016] In a preferred embodiment, the AI-based intelligent road surface defect inspection method is applied to a UAV-vehicle-mounted terminal collaborative road surface defect inspection scenario for urban overpasses and interchange ramps. The UAV inspection terminal includes a multi-rotor UAV, an airborne visible light camera, an airborne positioning module, an airborne inertial measurement unit, and an airborne unlicensed spectrum communication module. The vehicle-mounted inspection terminal includes an inspection vehicle, a vehicle-mounted close-range camera, a vehicle-mounted positioning module, a vehicle-mounted inertial measurement unit, a vehicle-mounted odometer, and a vehicle-mounted unlicensed spectrum communication module. The UAV inspection terminal acquires overhead road surface image data from outside the overpass guardrail or above the interchange ramp, while the vehicle-mounted inspection terminal acquires close-range road surface image data along the road travel direction and verifies the suspected defect segment index sent by the UAV inspection terminal.

[0017] According to an embodiment of the present invention, in combination Figure 1 The flowchart shown illustrates an intelligent road surface defect inspection method based on artificial intelligence, which specifically includes the following steps: S1. Merge the pose information and relative motion physical parameter sets of the UAV patrol terminal and the vehicle-mounted patrol terminal within the same road segment to generate road segment identifiers and a collaborative patrol geometric matrix. Note that the following should be noted in this step: S1.1. Based on road mileage, lane direction, and road boundary location, assign drone patrol terminals and vehicle-mounted patrol terminals to the same road segment and generate road segment identifiers.

[0018] Specifically, the road mileage location is obtained jointly by the vehicle positioning module of the vehicle-mounted patrol terminal, the vehicle odometer, and the vector data of the centerline of the urban elevated road. Among them, the vehicle positioning module obtains the latitude and longitude coordinates and altitude of the vehicle-mounted patrol terminal, the vehicle odometer obtains the cumulative travel distance of the vehicle-mounted patrol terminal along the driving direction, and the vector data of the centerline of the urban elevated road includes the centerline of the elevated bridge main line, the centerline of the interchange ramps, the starting mileage of the road, the ending mileage of the road, and the ramp number.

[0019] The road mileage location of the drone patrol terminal is obtained by projecting the latitude and longitude coordinates obtained by the airborne positioning module onto the vector data of the center line of the urban overpass. The road mileage location of the vehicle-mounted patrol terminal is obtained by jointly verifying the latitude and longitude coordinates obtained by the vehicle-mounted positioning module and the cumulative driving distance obtained by the vehicle-mounted odometer.

[0020] Preferably, the road mileage location includes the route number, the mileage value from the road start point to the current location, the ramp number, and the bridge span number, such as GJ03, K12+360, R2, P18-P19.

[0021] In a preferred embodiment, the lane direction is obtained jointly from the driving direction attribute in the urban elevated road centerline vector data, the vehicle heading angle of the vehicle-mounted patrol terminal, and the shooting direction of the drone patrol terminal; wherein, the lane direction includes the uphill direction, downhill direction, ramp entry direction, and ramp exit direction; the road boundary position is obtained jointly from the urban elevated road boundary vector data, lane line recognition results, and guardrail boundary recognition results; the road boundary position includes the left guardrail boundary, right guardrail boundary, central median boundary, shoulder boundary, and ramp divergence nose boundary; the lane line recognition results and guardrail boundary recognition results are obtained from the top-view road surface image data obtained by the drone patrol terminal and the close-range road surface image data obtained by the vehicle-mounted patrol terminal, respectively, after edge detection and semantic segmentation processing.

[0022] Furthermore, the method for classifying drone patrol terminals and vehicle-mounted patrol terminals into the same road segment is as follows: First, using road mileage location as the primary index, urban elevated bridges and interchange ramps are divided into continuous road segments along the road centerline, with each road segment having a mileage length of 20m; then, using lane direction as the second index, the uphill, downhill, ramp entry, and ramp exit directions within the same mileage length range are divided separately; then, using road boundary location as the third index, the left lane area, central lane area, right lane area, and shoulder area within the same lane direction are divided separately; when the projection position of the drone patrol terminal and the projection position of the vehicle-mounted patrol terminal fall within the same mileage length range, the same lane direction, and the same road boundary area, the drone patrol terminal and the vehicle-mounted patrol terminal are classified into the same road segment.

[0023] Preferably, the road segmentation identifier consists of a route number, mileage interval, lane direction, road boundary area, and segmentation sequence number; for example, if the route number is GJ03, the mileage interval is from K12+360 to K12+380, the lane direction is the exit direction of the ramp, the road boundary area is the right lane area, and the segmentation sequence number is 018, the road segmentation identifier is recorded as GJ03-K12+360-K12+380-OUT-R-018.

[0024] S1.2. Merge the spatial position, flight attitude, and shooting direction of the UAV patrol terminal with the spatial position, vehicle attitude, and shooting direction of the vehicle-mounted patrol terminal to generate a terminal pose combination.

[0025] It should be noted that the spatial position of the drone inspection terminal includes the drone's longitude, latitude, altitude, and relative height to the ground; the flight attitude of the drone inspection terminal includes the drone's heading angle, pitch angle, and roll angle; the shooting direction of the drone inspection terminal includes the azimuth angle of the onboard camera's optical axis, the pitch angle of the onboard camera's optical axis, and the field of view coverage boundary of the onboard camera; among these, the spatial position of the drone inspection terminal is obtained by the onboard positioning module, the flight attitude is obtained by the onboard inertial measurement unit, and the shooting direction is obtained by the gimbal angle, the onboard camera mounting angle, and the onboard camera's intrinsic parameters.

[0026] Furthermore, the spatial location of the vehicle-mounted patrol terminal includes the vehicle's longitude, latitude, altitude, and relative height to the road surface; the vehicle's attitude includes its heading angle, pitch angle, and roll angle; and its shooting direction includes the azimuth angle of the vehicle's optical axis, pitch angle, and field of view coverage boundary. The spatial location of the vehicle-mounted patrol terminal is obtained jointly by the vehicle positioning module and the vehicle's odometer; the vehicle's attitude is obtained by the vehicle's inertial measurement unit; and the shooting direction is obtained jointly by the vehicle's camera mounting angle, external parameters of the vehicle's camera, and the vehicle's attitude.

[0027] In this embodiment, "same-segment merging" refers to merging the spatial position, flight attitude, and shooting direction of UAV patrol terminals at the same sampling time within the same road segment based on the road segment identifier, and arranging them in the same terminal pose combination with the spatial position, vehicle attitude, and shooting direction of vehicle-mounted patrol terminals; wherein the terminal pose combination includes UAV spatial position combination, UAV attitude combination, UAV shooting direction combination, vehicle spatial position combination, vehicle attitude combination, and vehicle shooting direction combination.

[0028] Preferably, the time difference between the same sampling time is no more than 100ms. When the time difference between the sampling time of the UAV patrol terminal and the sampling time of the vehicle-mounted patrol terminal is greater than 100ms, a set of terminal pose data with a smaller time difference is selected to participate in the same-slice merging.

[0029] S1.3. Based on the terminal pose combination, determine the relative distance, relative height, relative direction and relative speed between the UAV patrol terminal and the vehicle-mounted patrol terminal, and generate a set of relative motion physical parameters.

[0030] It should be noted that, in order to reduce the curvature error caused by direct calculation of latitude and longitude, the spatial positions of the UAV patrol terminal and the vehicle-mounted patrol terminal are first converted to the same local northeast-sky coordinate system of the road segment. The local northeast-sky coordinate system of the road segment takes the center of the starting point of the road segment as the coordinate origin, the road driving direction as the north reference, the road horizontal direction as the east reference, and the vertical upward direction as the sky reference.

[0031] For example, the relative distance, relative height, relative direction, and relative speed in this embodiment can be obtained according to the following formulas:

[0032]

[0033]

[0034]

[0035] Where D is the relative distance between the drone patrol terminal and the vehicle-mounted patrol terminal; The eastward coordinates of the drone patrol terminal in the local northeast-sky coordinate system of the road segment; The coordinates of the vehicle-mounted patrol terminal in the local northeast-sky coordinate system of the road segment; The north coordinates of the drone patrol terminal in the local northeast-sky coordinate system of the road segment; The north-facing coordinates of the vehicle-mounted patrol terminal in the local northeast-south coordinate system of the road segment; The celestial coordinates of the drone patrol terminal in the local northeast celestial coordinate system of the road segment; The celestial coordinates of the vehicle-mounted patrol terminal in the local northeast-central celestial coordinate system of the road segment; The relative height between the drone patrol terminal and the vehicle-mounted patrol terminal; V represents the relative direction of the drone patrol terminal to the vehicle-mounted patrol terminal; V represents the relative speed between the drone patrol terminal and the vehicle-mounted patrol terminal. The eastward speed of the drone patrol terminal; The eastward speed of the vehicle-mounted patrol terminal; The northbound speed of the drone patrol terminal; The northbound speed of the vehicle-mounted patrol terminal; The upward speed of the drone inspection terminal; The aerial speed of the vehicle-mounted patrol terminal.

[0036] Preferably, the relative motion physical parameter set only selects relative distance, relative height, relative direction, and relative speed because these four types of parameters correspond to the propagation path length, pitch angle compensation, horizontal beam azimuth, and terminal relative motion changes in sidelink communication, respectively. Relative distance directly affects the received signal strength of unlicensed spectrum signals, relative height directly affects the pitch pairing of the transmit and receive beams, relative direction directly affects beam pointing selection, and relative speed directly affects the beam switching sequence and channel switching timing. Compared with simply using latitude and longitude coordinates, the above-mentioned relative motion physical parameter set can reduce road segmentation errors and communication pairing deviations in environments such as urban overpass curves, ramp merging, and overlapping upper and lower bridge decks.

[0037] S1.4. According to the road segmentation identifier, arrange the terminal pose combination and the relative motion physical parameter group in columns to generate a collaborative inspection geometric matrix.

[0038] In a preferred embodiment, the collaborative inspection geometric matrix is ​​a 2x2 matrix; wherein: The first row of the collaborative patrol geometry matrix is ​​the UAV patrol terminal geometry parameter row. The UAV patrol terminal geometry parameter row is arranged in the following order: road segment identifier, terminal category identifier, spatial location, terminal altitude, heading angle, pitch angle, shooting direction, road boundary offset, relative distance to vehicle-mounted patrol terminal, relative altitude, relative direction, and relative speed. The second row of the collaborative patrol geometry matrix is ​​the vehicle-mounted patrol terminal geometry parameter row. The vehicle-mounted patrol terminal geometry parameter row is arranged in the following order: road segmentation identifier, terminal category identifier, spatial location, terminal height, vehicle heading angle, vehicle pitch angle, shooting direction, road boundary offset, relative distance, relative height, relative direction, and relative speed with the drone patrol terminal.

[0039] It should be noted that step S1 unifies the UAV overhead patrol view and the vehicle-mounted close-range patrol view under the same road segment identifier in urban elevated bridges and interchange ramps, and transforms the relative spatial relationship between the two types of patrol terminals into a collaborative patrol geometric matrix. This solves the problem that UAV images and vehicle-mounted images are difficult to classify into the same road segment under conditions such as elevated bridge curves, ramp divergence, and overlapping upper and lower road levels. This makes the road segment classification clear, the relative orientation of the terminals clear, and the selection of subsequent communication beams have a geometric basis.

[0040] S2. Generate a sidelink spectrum state matrix based on the unlicensed spectrum signal state between the UAV patrol terminal and the vehicle-mounted patrol terminal. Note that the following points should be noted in this step: S2.1 According to the road segmentation identifier, the unlicensed spectrum channel between the UAV patrol terminal and the vehicle-mounted patrol terminal is divided into a candidate side link channel set.

[0041] In this embodiment, the unlicensed spectrum channels are obtained by the airborne unlicensed spectrum communication module of the UAV patrol terminal and the vehicle-mounted unlicensed spectrum communication module of the vehicle patrol terminal; wherein, the unlicensed spectrum channels include channels in the 2.4GHz band, channels in the 5.8GHz band, and short-range communication channels that can be used under the condition of adjacent unlicensed spectrum in 5.9GHz.

[0042] Preferably, in the environment of urban elevated bridges and interchange ramps, 5735MHz, 5755MHz, 5775MHz, 5795MHz, 5815MHz and 5835MHz in the 5.8GHz band are selected as candidate frequency points, and each candidate frequency point is combined with a 20MHz channel bandwidth to form a candidate sidelink channel.

[0043] In a preferred embodiment, the method for dividing the candidate sidelink channel set includes: first, determining that the UAV patrol terminal and the vehicle-mounted patrol terminal are located in the same road segment according to the road segment identifier; then, the UAV patrol terminal sends probe frames at each candidate frequency point, and the vehicle-mounted patrol terminal returns confirmation frames; then, unlicensed spectrum channels that can obtain confirmation frames and have received signal strength not lower than -82dBm are included in the candidate sidelink channel set; wherein, each road segment generates a separate candidate sidelink channel set, and different road segments do not share the same candidate sidelink channel set, so as to reduce spectrum state confusion between patrol terminals of parallel roads on ramps, auxiliary roads under bridges, and adjacent lanes.

[0044] S2.2 Merge the received signal strength, signal-to-noise ratio, channel occupancy rate and retransmission status of each candidate sidelink channel in the candidate sidelink channel set to generate a sidelink signal parameter group.

[0045] Specifically, the received signal strength is obtained by measuring the probe frames sent by the UAV patrol terminal using the vehicle-mounted unlicensed spectrum communication module; the signal-to-noise ratio (SNR) is calculated by combining the received signal power and noise power; the channel occupancy rate is obtained by statistically analyzing the carrier sensing results of the airborne and vehicle-mounted unlicensed spectrum communication modules on the same candidate sidelink channel; and the retransmission status is obtained by statistically analyzing the number of retransmissions of probe frames, acknowledgment frames, and index transmission frames on the same candidate sidelink channel. Among these, the received signal strength includes the received signal strength of the probe frames and the received signal strength of the acknowledgment frames; the SNR includes the SNR of the probe frames and the SNR of the acknowledgment frames; the channel occupancy rate includes the air interface occupancy ratio and the number of consecutive occupancy segments; and the retransmission status includes the number of retransmissions, the number of consecutive retransmissions, and the retransmission ratio.

[0046] For example, the signal-to-noise ratio, channel occupancy rate, and retransmission status in this embodiment can be obtained using the following formulas:

[0047]

[0048]

[0049] Wherein, SNR is the signal-to-noise ratio of the candidate side link channel; The received signal power on the candidate side link channel; denoted as , where is the noise power on the candidate side link channel; and C is the channel occupancy rate of the candidate side link channel. The cumulative duration during which the candidate side link channel is occupied; R represents the total measurement duration of the candidate side link channel; R represents the retransmission status of the candidate side link channel. The number of retransmitted frames on the candidate side link channel; This represents the total number of frames transmitted on the candidate side link channel.

[0050] It should be noted that the sidelink signal parameter group in this embodiment only selects received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status. This is because received signal strength reflects the propagation attenuation between the UAV patrol terminal and the vehicle-mounted patrol terminal, signal-to-noise ratio reflects the demodulation quality of the candidate sidelink channel, channel occupancy rate reflects the degree of occupation of unlicensed spectrum by co-frequency devices, and retransmission status reflects the transmission stability of the suspected defect fragment index on the candidate sidelink channel. The above four types of parameters cover the four aspects of propagation, noise, contention occupancy, and transmission failure, respectively, and can form a spectrum status description suitable for the transmission of the suspected defect fragment index.

[0051] S2.3. Based on the channel arrangement order of the candidate side link channel set, arrange the side link signal parameter group by column to generate the initial spectrum state matrix.

[0052] Specifically, the channel arrangement order of the candidate sidelink channel set is arranged from low to high according to the candidate frequency point; when two candidate sidelink channels have the same candidate frequency point, they are arranged from small to large according to the channel bandwidth; when the candidate frequency point and channel bandwidth are the same, they are arranged from first to last according to the acquisition time of the probe frame; the number of rows of the initial spectrum state matrix is ​​consistent with the number of candidate sidelink channels in the candidate sidelink channel set, and each row is the sidelink signal parameter group of one candidate sidelink channel; the columns of the initial spectrum state matrix are, in order, road segmentation identifier, channel arrangement identifier, received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status.

[0053] S2.4. Remove candidate sidelink channel entries from the initial spectrum state matrix that are simultaneously in the high-order sorting of channel occupancy and the high-order sorting of retransmission status, and generate the sidelink spectrum state matrix.

[0054] In a preferred embodiment, the sidelink spectrum state matrix is ​​a multi-row, six-column matrix; wherein: The number of rows in the sidelink spectrum state matrix is ​​the same as the number of candidate sidelink channels retained in the sidelink spectrum state matrix, and each row is the spectrum state parameter row of a retained candidate sidelink channel. The six columns of the sidelink spectrum status matrix are arranged in the order of road segmentation identifier, channel arrangement identifier, received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status.

[0055] In this embodiment, the high-order ranking of channel occupancy refers to the candidate sidelink channels that are in the top 30% of the channel occupancy ranking from high to low under the same road segment identifier, and whose channel occupancy is greater than 0.60; the high-order ranking of retransmission status refers to the candidate sidelink channels that are in the top 30% of the retransmission status ranking from high to low under the same road segment identifier, and whose retransmission status is greater than 0.15; when a candidate sidelink channel belongs to both the high-order ranking of channel occupancy and the high-order ranking of retransmission status, the entire row corresponding to the candidate sidelink channel is removed from the initial spectrum state matrix to obtain the sidelink spectrum state matrix.

[0056] For example, the initial spectrum state matrix includes 6 candidate sidelink channels, where CH05 has a channel occupancy rate of 0.74 and a retransmission status of 0.21, and CH06 has a channel occupancy rate of 0.69 and a retransmission status of 0.18. Both are in the high-order order of channel occupancy rate and retransmission status. Remove the candidate sidelink channel columns corresponding to CH05 and CH06, and keep CH01, CH02, CH03 and CH04 to generate a 4-row, 6-column sidelink spectrum state matrix.

[0057] Furthermore, the spectrum status parameters include received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status; the channel arrangement identifier includes frequency point number, channel bandwidth number, and probe frame sequence number; the frequency point number is obtained from low to high according to the candidate frequency points, the channel bandwidth number is obtained according to the bandwidth order of 20MHz, 40MHz, and 80MHz, and the probe frame sequence number is obtained according to the time sequence of the probe frames sent by the UAV patrol terminal; for example, CH03-5775-20-02 represents the 3rd candidate sidelink channel, with a candidate frequency of 5775MHz, a channel bandwidth of 20MHz, and a probe frame sequence number of 02.

[0058] Preferably, step S2 forms a sidelink spectrum state matrix that reflects the communicable status of unlicensed spectrum under the same road segment identifier, and eliminates candidate sidelink channels that have both high occupancy and high retransmission characteristics. This solves the problems of dense unlicensed spectrum devices, obvious co-channel interference, and easy loss of suspected defect segment index transmission in urban overpass and interchange ramp environments. It makes the selection criteria for candidate sidelink channels clear, the spectrum state arrangement structure clear, and provides reliable input for subsequent beam management and channel selection.

[0059] S3. Based on the collaborative inspection geometric matrix and the sidelink spectrum state matrix, determine the beam management parameter set, which includes beam pointing and beam switching order, and form the sidelink communication relationship according to the beam management parameter set. Note that the following should be noted in this step: S3.1 Extract the relative distance, relative altitude, and relative direction between the UAV patrol terminal and the vehicle-mounted patrol terminal from the collaborative patrol geometric matrix to generate the terminal relative orientation group.

[0060] It should be noted that the relative distance, relative height, and relative direction in this step have already been obtained in step S1.3. Step S3.1 extracts them from the 9th, 10th, and 11th columns of the collaborative patrol geometry matrix and arranges them in the order of road segment identification, relative distance, relative height, and relative direction.

[0061] It should also be noted that in this embodiment, the terminal relative orientation group only selects relative distance, relative altitude, and relative direction because beam management first needs to determine the spatial pointing relationship between the UAV patrol terminal and the vehicle-mounted patrol terminal; relative distance determines the transmit beamwidth and candidate side link channel priority, relative altitude determines the pitch angle direction of the transmit beam, and relative direction determines the horizontal angle direction of the transmit beam; relative velocity has already been generated in step S1.3 and entered into the collaborative patrol geometry matrix, but relative velocity is not used as the main arrangement item when extracting the terminal relative orientation group in step S3.1. The reason is that this step first completes the spatial orientation selection, and relative velocity can be used as an auxiliary reference when fine-tuning the beam switching order.

[0062] S3.2 From the sidelink spectrum state matrix, select the candidate sidelink channels to be retained in order of signal-to-noise ratio from high to low and channel occupancy rate from low to high, and generate the target sidelink channel group.

[0063] Specifically, from the sidelink spectrum state matrix, firstly, candidate sidelink channels with a signal-to-noise ratio (SNR) below 12dB are removed. Then, the remaining candidate sidelink channels are arranged from highest to lowest SNR. When the SNR difference between two candidate sidelink channels is no greater than 2dB, the candidate sidelink channel with the lower channel occupancy rate is placed first. When the SNR difference between two candidate sidelink channels is no greater than 2dB and the channel occupancy rate difference is no greater than 0.05, the candidate sidelink channel with the lower retransmission status is placed first. After arrangement, the top three candidate sidelink channels are selected to form the target sidelink channel group. When the number of remaining candidate sidelink channels is less than three, all remaining candidate sidelink channels are used to form the target sidelink channel group.

[0064] S3.3 Determine the beam direction based on the terminal relative orientation group, and determine the beam switching order based on the channel arrangement order of the target side link channel group, and generate a beam management parameter group.

[0065] In a preferred embodiment, the method for determining the beam direction based on the relative orientation group of the terminals is as follows: first, the horizontal beam direction is determined based on the relative direction, and the horizontal angle range falling within the relative direction is mapped to the transmitting beam set of the UAV patrol terminal and the receiving beam set of the vehicle-mounted patrol terminal; then, the pitch beam direction is determined based on the relative height and relative distance, and the pitch angle range is mapped to the transmitting beam set of the UAV patrol terminal and the receiving beam set of the vehicle-mounted patrol terminal.

[0066] Preferably, the horizontal beam pointing includes a forward beam, a left forward beam, a right forward beam, a left forward beam, a right forward beam, and a backward beam; the pitch beam pointing includes an upward pitch beam, a line-of-sight beam, and a downward pitch beam.

[0067] For example, when the relative direction is 35°, the UAV patrol terminal is located in the right front area relative to the vehicle-mounted patrol terminal, and the beam direction is determined to be the right front beam; when the relative height is 26.4m and the relative distance is 42.5m, the receiving direction from the vehicle-mounted patrol terminal to the UAV patrol terminal is the upward pitch direction, the receiving beam of the vehicle-mounted patrol terminal is selected as the upward pitch beam, and the transmitting beam of the UAV patrol terminal is selected as the downward pitch beam.

[0068] Preferably, in this embodiment, the channel arrangement order of the target side link channel group refers to the order after being arranged according to signal-to-noise ratio, channel occupancy rate, and retransmission status in step S3.2; while the beam switching order is determined according to the channel arrangement order of the target side link channel group, including the main communication beam order, the backup communication beam order, and the fallback communication beam order; the target side link channel arranged first corresponds to the main communication beam order, the target side link channel arranged second corresponds to the backup communication beam order, and the target side link channel arranged third corresponds to the fallback communication beam order.

[0069] Furthermore, the beam management parameter group includes road segmentation identifier, main communication beam pointing, backup communication beam pointing, fallback communication beam pointing, target side link channel group, and beam switching sequence.

[0070] S3.4. According to the beam management parameter group, pair the transmit beam and receive beam between the UAV patrol terminal and the vehicle-mounted patrol terminal to form a side link communication relationship.

[0071] Specifically, pairing includes: According to the beam direction in the beam management parameter group, select the UAV transmitting beam towards the vehicle-mounted patrol terminal from the transmitting beam set of the UAV patrol terminal, and select the vehicle receiving beam towards the UAV patrol terminal from the receiving beam set of the vehicle-mounted patrol terminal to form the first beam pairing group. According to the beam switching order in the beam management parameter group, the vehicle-mounted transmitting beam towards the drone patrol terminal is selected from the transmitting beam set of the vehicle-mounted patrol terminal, and the drone receiving beam towards the vehicle-mounted patrol terminal is selected from the receiving beam set of the drone patrol terminal, thus forming a second beam pairing group. The first beam pairing group, the second beam pairing group, and the target side link channel group are combined to form a side link communication relationship.

[0072] It should be noted that step S3 combines the geometric orientation relationship formed in step S1 with the spectrum state relationship formed in step S2 to form a beam management parameter group that includes beam pointing, target side link channel group and beam switching order; it solves the problem of unstable communication direction between UAV patrol terminal and vehicle-mounted patrol terminal under conditions of overpass curves, interchange ramps, and changes in vehicle driving attitude, so that beam pairing has spatial basis, channel selection has spectrum basis, and the transmission path of suspected defect fragment index is clear.

[0073] S4. The road surface image data of road segments is identified by the defect identification model in the UAV inspection terminal, a suspected defect segment index is generated, and sent to the vehicle-mounted inspection terminal via sidelink communication. Note that the following should be noted in this step: S4.1. According to the road segmentation identifier, the road surface image data obtained by the UAV inspection terminal is segmented and arranged to generate road surface segmented image groups.

[0074] Specifically, the road surface image data is obtained by the airborne visible light camera of the drone inspection terminal. The image resolution of the airborne visible light camera is 3840×2160, the lens focal length is 24mm, the shooting height is 25m to 35m, and the gimbal pitch angle is -75° to -90°. The drone inspection terminal arranges the road surface image data falling within the same mileage range, the same lane direction, and the same road boundary area into road surface image groups according to the road segmentation markers.

[0075] Each road surface image group includes at least one overhead view of the road surface, preferably three consecutive overhead view images of the road surface, with the shooting interval between the three consecutive overhead view images being a flight distance of 1m to 3m.

[0076] S4.2 The defect identification model identifies crack areas, pothole areas, subsidence areas and repair damage areas in the road surface segment image group, and generates defect candidate area groups.

[0077] In this embodiment, the road damage identification model is constructed based on the Mask R-CNN instance segmentation algorithm. The model is deployed within the edge computing unit of the UAV inspection terminal. The model takes road surface segmented image groups as input and outputs a group of candidate road damage regions. Each road surface segmented image in the image group corresponds to a unique road segment identifier, and each image includes the UAV's spatial location, flight attitude, shooting direction, and image acquisition time. The model performs instance segmentation and identification on crack areas, pothole areas, subsidence areas, and repair damage areas in the road surface segmented image group, and outputs the damage category, the bounding rectangle position of the damage, the mask area of ​​the damage, and the confidence score of the damage category for each candidate road damage region, generating a group of candidate road damage regions.

[0078] Specifically, the disease identification model includes an image normalization layer, a backbone feature extraction network, a feature pyramid network, a candidate region generation network, a region feature alignment layer, a disease category classification branch, a disease boundary regression branch, and a disease mask segmentation branch. The image normalization layer performs size adjustment, pixel value normalization, and effective road region cropping on each road segment image in the road segment image group. The backbone feature extraction network adopts a ResNet-50 structure and includes convolutional layers, residual block group 1, residual block group 2, residual block group 3, and residual block group 4 in sequence. 4; The feature pyramid network receives the multi-scale feature map output by the backbone feature extraction network and forms a fused feature map of 4 scales; the candidate region generation network generates candidate disease boxes on the fused feature map; the region feature alignment layer performs scale alignment on the region features corresponding to the candidate disease boxes; the disease category classification branch outputs the category confidence of crack region, pit region, subsidence region, repair and damage region and background region; the disease boundary regression branch outputs the position of the bounding rectangle of the disease candidate region; the disease mask segmentation branch outputs the pixel-level mask region of the disease candidate region.

[0079] In a preferred embodiment, the image normalization processing layer specifically includes: first, adjusting each road surface segment image in the road surface segment image group to 1024×1024 pixels; then, normalizing the pixel values ​​of the red, green, and blue channels of the road surface segment images, with the normalized pixel values ​​ranging from 0 to 1; then, based on the road boundary positions corresponding to the road segment identifiers, excluding the areas outside the left guardrail, the areas outside the right guardrail, the non-road surface areas outside the bridge shadows, and the vehicle occlusion areas from the road surface segment images to obtain the effective road image area; wherein, the effective road image area retains the road surface layer, lane lines, expansion joints, repair areas, and shoulder edges; the vehicle occlusion areas are obtained from the vehicle outline segmentation results, which can be obtained by a conventional vehicle target detection model; the effective road image area serves as the input to the backbone feature extraction network, enabling the defect identification model to reduce misidentification of guardrails, bridge pier shadows, and vehicle outlines.

[0080] Furthermore, the backbone feature extraction network uses ResNet-50 initial parameters trained on the ImageNet public image dataset and is retrained on urban overpass and interchange ramp pavement defect sample images. The backbone feature extraction network outputs feature maps at four scales, corresponding to the 1 / 4, 1 / 8, 1 / 16, and 1 / 32 scales of the pavement patch images, respectively. The 1 / 4 scale feature map preserves crack edges, small repair damage edges, and lane line boundaries; the 1 / 8 scale feature map preserves pothole boundaries, repair damage block boundaries, and local subsidence shadows; the 1 / 16 scale feature map preserves larger subsidence areas and continuous repair damage areas; and the 1 / 32 scale feature map preserves the overall grayscale changes of the road patch and the curved background of the ramp pavement. The feature pyramid network fuses the feature maps at the above four scales from top to bottom to form four fused feature maps, so that crack areas, pothole areas, subsidence areas, and repair damage areas all have segmentable image features at different scales.

[0081] The candidate region generation network sets candidate anchor boxes on four fused feature maps respectively; considering the differences in the shape of pavement defects of urban overpasses and interchange ramps, the scale of the candidate anchor boxes includes 16×16 pixels, 32×32 pixels, 64×64 pixels, 128×128 pixels and 256×256 pixels; the aspect ratio of the candidate anchor boxes includes 1:1, 2:1, 4:1, 8:1, 1:2 and 1:4. Among them, 16×16 pixels, 32×32 pixels, and aspect ratios of 4:1 and 8:1 are mainly adapted to narrow and long crack areas; 64×64 pixels, 128×128 pixels, and aspect ratios of 1:1 and 2:1 are mainly adapted to pit and repair damage areas; 128×128 pixels, 256×256 pixels, and aspect ratios of 1:1 and 1:2 are mainly adapted to subsidence areas and larger repair damage areas; the candidate region generation network outputs the disease foreground score and the candidate disease box position for each candidate anchor box, and retains candidate disease boxes with a disease foreground score of not less than 0.50.

[0082] Furthermore, the candidate disease boxes retained by the candidate region generation network are subjected to overlap rejection processing. The rules for overlap rejection processing are as follows: multiple candidate disease boxes with an overlap area ratio greater than 0.70 under the same disease category are compared, and the candidate disease box with the higher disease foreground score is retained, while the remaining candidate disease boxes are rejected; when two candidate disease boxes correspond to a crack area and a repaired damaged area respectively, and the overlap area ratio is greater than 0.70, both candidate disease boxes are retained and sent to the disease category classification branch for further classification; when two candidate disease boxes correspond to a pit area and a subsidence area respectively, and the overlap area ratio is greater than 0.70, the candidate disease box with the larger area is retained and sent to the disease mask segmentation branch.

[0083] Preferably, the above treatment ensures that the slender cracks are not directly covered by the road repair boundary, and that potholes and subsidence retain the basis for subsequent segmentation and judgment under approximately grayscale change conditions.

[0084] It should be noted that the region feature alignment layer adjusts the region features corresponding to each candidate disease box to 14×14 region classification features and 28×28 region mask features; the region classification features enter the disease category classification branch and the disease boundary regression branch; the region mask features enter the disease mask segmentation branch; the region feature alignment layer uses bilinear interpolation to preserve the fine edges of cracks and the transition texture of pit boundaries during the scale transformation of candidate disease boxes.

[0085] In a preferred embodiment, the disease category classification branch includes a fully connected layer 1, a fully connected layer 2, and a category output layer; the category output layer outputs five categories of confidence scores, namely crack area, pothole area, subsidence area, repair and damage area, and background area; the disease boundary regression branch outputs the location of the disease's bounding rectangle, which includes the horizontal coordinate of the top left pixel, the vertical coordinate of the top left pixel, the width of the bounding rectangle, and the height of the bounding rectangle; the disease mask segmentation branch includes four convolutional layers, one deconvolutional layer, and one mask output layer, which outputs a 28×28 disease mask probability map, and maps it to the original pixel coordinates of the road surface patch image after scale restoration processing to form the disease mask region.

[0086] It should also be noted that the training samples for the defect identification model in this embodiment are UAV overhead view road surface sample images of the main line of urban elevated bridges, interchange ramps, ramp merging areas, ramp diverging areas, areas adjacent to expansion joints, areas adjacent to bridge deck drainage channels, and guardrail shadow areas; the UAV overhead view road surface sample images are taken at a height of 20m to 40m, with an image resolution of not less than 3840×2160 and a shooting pitch angle of -70° to -90°; the training samples include crack area samples, pothole area samples, subsidence area samples, repaired and damaged area samples, and defect-free road surface samples.

[0087] Preferably, the total number of training samples is no less than 12,000, including no less than 3,000 samples of cracked areas, no less than 2,000 samples of pothole areas, no less than 1,500 samples of subsidence areas, no less than 2,500 samples of repaired and damaged areas, and no less than 3,000 samples of road surface without defects. The samples of road surface without defects include lane lines, expansion joints, drainage ditches, shadow boundaries, tire tracks, and bridge deck joints, so as to reduce the probability that the defect identification model will misclassify road ancillary structures as defect areas.

[0088] Specifically, the training samples are labeled using a pixel-level annotation method to form the training annotation results. The training annotation results include the annotation of the defect category, the annotation of the defect bounding rectangle, and the annotation of the defect mask. The annotation of the crack area forms a continuous mask along the visible edge of the crack. When the crack width is less than 3 pixels, it is annotated with a width of 3 pixels. The annotation of the pothole area forms a closed mask along the pothole damage boundary. The annotation of the subsidence area forms a closed mask along the continuous low-lying shadow boundary and the edge deformation line of the road surface. The annotation of the repair damage area forms a closed mask along the repair material boundary and the damage edge. Expansion joints, lane lines, drainage ditches, and guardrail shadows are not labeled as defect areas and are used as background areas in the defect-free road surface samples for training.

[0089] The training samples are divided into training image group, validation image group and test image group in a 7:2:1 ratio. The training image group is used for parameter learning of the disease identification model. The validation image group is used for selecting the disease category confidence threshold, overlap removal threshold and mask area threshold. The test image group is used for the final recognition performance check of the disease identification model. During the division, sample images corresponding to the same road segment, the same inspection date and the same disease location are only included in one of the training image group, validation image group and test image group to avoid the same disease from appearing repeatedly in the training and testing phases.

[0090] It should be noted that image enhancement processing is required before the training image group is entered into the disease identification model. Image enhancement processing includes brightness adjustment, contrast adjustment, rotation, random cropping, scaling, motion blur, and shadow overlay. The brightness adjustment range is 0.7 to 1.3, the contrast adjustment range is 0.8 to 1.2, the rotation angle range is -15° to 15°, the image size after random cropping is 1024×1024 pixels, the scaling range is 0.75 to 1.50, the motion blur kernel size is 3 to 7, and the shadow overlay transparency is 0.15 to 0.35. The above image enhancement processing corresponds to the strong light, bridge shadows, ramp turning perspective, slight drone shaking, and different flight altitudes in urban overpass inspections, so that the disease identification model can be adapted to the actual inspection images of urban overpasses and interchange ramps.

[0091] Furthermore, during the training of the defect identification model, the input image size was 1024×1024 pixels, the batch size was 2 to 4, the training epochs were 80, the initial learning rate was 0.002, and the learning rate decayed to 0.1 of the previous stage in the 40th and 60th epochs, respectively. The weight decay coefficient was 0.0001, and the momentum coefficient was 0.9. The parameters of the first two residual blocks of the backbone feature extraction network were frozen in the first 10 epochs of training, and the entire network participated in the training from the 11th epoch. The above training method enabled the defect identification model to gradually adapt to the texture of the elevated bridge pavement, the morphology of defects, and the background of the curved ramps while retaining the general image edge features.

[0092] As an example, the training loss of the disease identification model consists of classification loss, boundary regression loss, and mask segmentation loss; the confidence score of the disease category output by the disease identification model is obtained by the following formula:

[0093] in, For the first Category confidence of the category; The output of the disease category classification branch Category score; The output of the disease category classification branch Category score; For category number; 5 represents the category number; 5 represents the number of categories, corresponding to cracked areas, pitted areas, subsidence areas, repaired and damaged areas, and background areas, respectively; e is a natural constant.

[0094] As an example, the training loss of the disease identification model is obtained by the following formula:

[0095] Where L is the training loss of the disease identification model; Classification loss for disease category classification branches; The boundary regression loss is the boundary regression branch for the disease boundary regression; The masking loss is used for the disease masking segmentation branch; the classification loss is obtained based on the difference between the category label and the category confidence; the boundary regression loss is obtained based on the difference between the disease circumscribed rectangle label and the position of the disease circumscribed rectangle output by the disease boundary regression branch; and the masking segmentation loss is obtained based on the difference between the disease mask label and the disease mask probability map output by the disease masking segmentation branch.

[0096] The training loss described above enables the disease identification model to learn disease categories, disease boundaries, and disease pixel regions simultaneously, avoiding the problem of unclear disease locations caused by relying solely on classification results.

[0097] After the disease identification model completes training, the output threshold for each disease category is determined based on the validation image set. The confidence threshold for the disease category is 0.55 for crack areas, 0.60 for pit areas, 0.65 for subsidence areas, and 0.60 for repaired and damaged areas. The minimum mask area is 80 pixels for crack areas, 150 pixels for pit areas, 300 pixels for subsidence areas, and 200 pixels for repaired and damaged areas. When the confidence of a disease category is not lower than the corresponding disease category confidence threshold, and the pixel area of ​​the disease mask area is not lower than the minimum mask area of ​​the corresponding disease category, the region is included in the disease candidate region group.

[0098] In a preferred embodiment, when the road surface defect identification model identifies road surface segment image groups, it first inputs the road surface segment images in the road surface segment image group into the image normalization processing layer to obtain the effective road image region; then, the backbone feature extraction network and the feature pyramid network extract multi-scale road surface texture features of the effective road image region; then, the candidate region generation network generates candidate defect boxes; then, the region feature alignment layer obtains the region features corresponding to each candidate defect box; then, the defect category classification branch obtains the defect category and defect category confidence; the defect boundary regression branch obtains the position of the defect circumscribed rectangle; and the defect mask segmentation branch obtains the defect mask region; finally, based on the defect category confidence threshold and the minimum mask area threshold corresponding to each defect category, a defect candidate region group is formed.

[0099] In a preferred embodiment, each candidate disease region in the disease candidate region group includes a road segment identifier, image sequence number, disease candidate region number, disease category, disease category confidence level, disease circumscribed rectangle position, disease mask area, and disease area level; the disease circumscribed rectangle position includes the horizontal coordinate of the upper left corner pixel, the vertical coordinate of the upper left corner pixel, the width of the circumscribed rectangle, and the height of the circumscribed rectangle; the disease area level is obtained based on the area corresponding to the disease mask area in the road segment plane, and the area is less than 0.05m². 2 It was a small-area disease, covering an area of ​​0.05m². 2 up to 0.50m 2 This is a medium-sized disease, with an area greater than 0.50m². 2 The disease is large-area; for cracked areas, the crack length can be classified according to the main axis length and average width of the disease masking area; cracks with a length less than 0.5m are short cracks, cracks with a length of 0.5m to 2.0m are medium cracks, and cracks with a length greater than 2.0m are long cracks.

[0100] In a specific example, the road segment identifier corresponding to the road surface segment image group is GJ03-K12+360-K12+380-OUT-R-018, and the image sequence number is IMG02. The defect identification model identifies one crack region in this road surface segment image. The defect category confidence score of this crack region is 0.86. The location of the outer rectangle of the defect is: top left corner pixel x-coordinate 1260, top left corner pixel y-coordinate 840, outer rectangle width 320, outer rectangle height 75, and the pixel area of ​​the defect mask region is 1420 pixels. Due to the high confidence score of the defect category of the crack region... The score of 0.86 is higher than the disease category confidence threshold of 0.55 for the crack area, and the pixel area of ​​the disease mask area of ​​1420 pixels is higher than the minimum mask area of ​​80 pixels for the crack area. Therefore, the crack area is included in the disease candidate area group. The content of the disease candidate area group is GJ03-K12+360-K12+380-OUT-R-018, IMG02, A03, crack area, 0.86, top left pixel x-coordinate 1260, top left pixel y-coordinate 840, bounding rectangle width 320, bounding rectangle height 75, disease mask area, long crack.

[0101] To reduce false detections caused by expansion joints, lane lines, drainage ditches, and guardrail shadows on urban elevated bridges and interchange ramps, the defect identification model needs to perform road ancillary structure exclusion processing before outputting defect candidate area groups. Specifically, road ancillary structure exclusion processing includes: when the overlap area ratio between the candidate defect area and the lane line area is greater than 0.60, and the defect type is crack area, the candidate defect area is marked as lane line edge interference area and not included in the defect candidate area group; when the overlap area ratio between the candidate defect area and the expansion joint area is greater than 0.60... If the area of ​​a candidate disease is 70 and the disease type is crack area, the candidate disease area will be marked as an expansion joint interference area and will not be included in the disease candidate area group; if the overlap ratio between the candidate disease area and the drainage channel area is greater than 0.60 and the disease type is subsidence area or repaired damaged area, the candidate disease area will be marked as a drainage channel boundary interference area and will not be included in the disease candidate area group; if the candidate disease area is located at the guardrail shadow boundary and the disease type confidence is less than 0.75, the candidate disease area will be marked as a shadow interference area and will not be included in the disease candidate area group.

[0102] Preferably, through step S4.2, the disease identification model not only provides the disease category judgment, but also provides the location of the disease's bounding rectangle and the disease mask area, enabling the subsequent step S4.3 to obtain the disease location and enabling step S5 to compare the location overlap between the disease candidate area group and the review disease area group. This step continuously connects the UAV top-view image acquisition conditions, disease sample sources, model structure, training annotation method, category output threshold, mask output rules, and road ancillary structure exclusion rules for urban overpasses and interchange ramps, solving the problem of only giving the model name without clarifying the model input, output, and training process.

[0103] S4.3 Combine the candidate disease area groups with road segmentation identifiers, disease locations, and disease categories to generate a suspected disease segmentation index.

[0104] In this embodiment, the location of the defect is obtained by combining the location of the circumscribed rectangle of the defect candidate region group, the defect mask area, and the road segment identifier corresponding to the road segment image group. First, the position of the defect candidate region in the image coordinate system is determined based on the location of the circumscribed rectangle of the defect. Then, by combining the intrinsic parameters of the airborne camera, the extrinsic parameters of the airborne camera, the spatial position of the UAV, and the shooting direction of the UAV, the defect candidate region is projected onto the road segment plane to obtain the defect location. The defect location includes the longitudinal distance within the road segment, the lateral distance within the road segment, the lane area to which it belongs, and the circumscribed range of the defect.

[0105] Preferably, the longitudinal distance within a road segment is based on the starting point of the road segment, and the lateral distance within a road segment is based on the right boundary of the road.

[0106] The disease category is output by the classification branch of the disease identification model and is determined in crack areas, pothole areas, subsidence areas, and repair damage areas. The candidate disease area group is combined with the road segment identifier, disease location, and disease category to generate a suspected disease segment index. The structure of the suspected disease segment index includes road segment identifier, suspected disease number, disease location, disease category, disease boundary, disease category confidence level, and image sequence number. For example, the suspected disease segment index is GJ03-K12+360-K12+380-OUT-R-018, D003, longitudinal 12.4m, transverse 3.2m, crack area, length 1.8m, width 0.03m, confidence level 0.86, IMG02.

[0107] S4.4 According to the side link communication relationship, the suspected disease segment index is sent from the UAV inspection terminal to the vehicle-mounted inspection terminal.

[0108] Specifically, the UAV patrol terminal first selects the target sidelink channel corresponding to the main communication beam sequence according to the target sidelink channel group in the sidelink communication relationship, and then sends the suspected defect segment index using the first beam pairing group; when the vehicle-mounted patrol terminal returns a confirmation frame, the UAV patrol terminal ends the transmission of the suspected defect segment index under that road segment identifier; when the vehicle-mounted patrol terminal does not return a confirmation frame, the UAV patrol terminal switches to the target sidelink channel corresponding to the backup communication beam sequence according to the beam switching sequence, and sends the suspected defect segment index again.

[0109] Preferably, the content transmitted in the suspected road defect segment index includes road segment identifier, suspected defect number, defect location, defect category, and defect category confidence level, instead of directly transmitting the complete road surface segment image group, thereby reducing unauthorized spectrum occupancy.

[0110] It should be noted that in step S4, the UAV inspection terminal first performs an overhead view of the elevated bridge and interchange ramps to identify suspected defects, forming a suspected defect segment index with road segmentation markings, defect locations, and defect types. This index is then sent to the vehicle-mounted inspection terminal via a side link communication relationship. This solves the problem of narrow field of view and difficulty in locating suspected defect areas in advance when the vehicle-mounted inspection terminal is inspecting alone. It allows for the preliminary screening of suspected defect areas, a reduction in the area to be verified by the vehicle, and a decrease in the amount of inspection calculations and communication data.

[0111] S5. The vehicle-mounted inspection terminal acquires close-range road surface image data within the road segments defined by the suspected defect segment index, performs verification and identification, and generates intelligent road surface defect inspection results. Note that the following should be noted in this step: S5.1 The vehicle-mounted inspection terminal determines the close-range verification area based on the road segmentation marks and the location of the defects in the suspected defect segmentation index.

[0112] Specifically, the vehicle-mounted inspection terminal first determines the route number, mileage range, lane direction, and road boundary area to which the suspected defect area index belongs based on the road segmentation markings; then, based on the longitudinal and lateral distances within the road segment at the defect location, it determines the close-up shooting position of the vehicle-mounted inspection terminal within that road segment; then, taking the defect location as the center, it takes 2 meters in front and behind in the longitudinal direction of the road, and 1 meter to the left and right in the lateral direction of the road, to form a close-up verification area.

[0113] Preferably, for the curved section of the interchange ramp, the longitudinal direction of the road is determined along the tangent direction of the ramp centerline, and the transverse direction of the road is determined along the normal direction of the ramp centerline.

[0114] S5.2 The vehicle-mounted inspection terminal acquires close-range road surface image data within the close-range verification area, and performs regional cropping of the close-range road surface image data according to the location of the defects to generate a verification image area group.

[0115] In this embodiment, close-range road surface image data refers to road surface image data obtained when the shooting distance between the vehicle-mounted close-range camera and the road surface is 0.6m to 2.5m; the vehicle-mounted close-range camera is installed at the front or side of the patrol vehicle at a height of 0.8m to 1.5m, with an image resolution of 3840×2160, and the shooting angle is directed towards the close-range verification area; the vehicle-mounted patrol terminal continuously acquires close-range road surface image data of the close-range verification area when the patrol speed is 5km / h to 20km / h, and the longitudinal distance between two adjacent close-range road surface image data is no more than 0.5m.

[0116] Specifically, the region clipping method is as follows: First, based on the vehicle camera's intrinsic and extrinsic parameters, vehicle attitude, and defect location, the defect location is projected onto the image coordinate system of the close-range road surface image data; then, using the projected defect center location as the clipping center, the clipping range is determined according to the defect category; for crack areas, the clipping range is 0.5m outside the defect's outer boundary; for pothole areas, the clipping range is 0.8m outside the defect's outer boundary; for subsidence areas, the clipping range is 1.0m outside the defect's outer boundary; for repaired damaged areas, the clipping range is 0.6m outside the defect's outer boundary; the pixel size of a single verified image region after clipping is not less than 640×640; when the projected area is close to the image edge, it is padded inwards to 640×640.

[0117] For example, if the defect category in the suspected defect patch index is crack area and the defect outer range is 1.8m in length and 0.03m in width, then the vehicle-mounted inspection terminal expands outward by 0.5m in both longitudinal and transverse directions based on the defect outer range to form a cut-off range of 2.8m in length and 1.03m in width, and obtains the corresponding verification image area group from the close-range road surface image data.

[0118] S5.3 The vehicle-mounted inspection terminal performs disease verification and identification on the verification image area group, generates a verification disease area group, and compares the location overlap of the verification disease area group with the disease candidate area group.

[0119] Specifically, the defect verification and identification adopts the same instance segmentation model with the same category system as step S4.2. The category system includes crack areas, pothole areas, subsidence areas, and repair damage areas. Unlike step S4.2, the input for defect verification and identification is close-range road surface image data. Defect verification and identification retains more road surface texture details at the image scale and performs fine-grained identification of crack width, pothole boundaries, subsidence shadows, and repair damage edges. The defect verification area group includes road segmentation identifier, verification area number, verification defect outer range, verification defect mask area, verification defect category, and verification confidence.

[0120] Furthermore, the method for comparing the location overlap between the review defect area group and the defect candidate area group is as follows: First, the masked region of the review defect in the review defect area group is transformed from the image coordinate system of the close-range pavement image data to the road segment plane; then, the masked region of the defect in the defect candidate area group is transformed from the image coordinate system of the pavement segment image group to the same road segment plane; then, the overlap area, center offset distance, and defect category consistency of the review defect masked region and the defect masked region are compared in the road segment plane; the obtained location overlap comparison results include high overlap comparison results, medium overlap comparison results, and low overlap comparison results. Among them, a high overlap comparison result means that the overlap area ratio between the verified disease mask areas is not less than 0.50, the center offset distance is not greater than 0.5m, and the verified disease category is consistent with the disease category; a medium overlap comparison result means that the overlap area ratio between the verified disease mask areas is 0.20 to 0.50, or the center offset distance is 0.5m to 1.2m, and the verified disease category is consistent with the disease category or belongs to a similar category; a low overlap comparison result means that the overlap area ratio between the verified disease mask areas is less than 0.20, or the center offset distance is greater than 1.2m, or the verified disease category is inconsistent with the disease category.

[0121] Preferably, similar categories include pothole areas and repaired damaged areas, subsidence areas and repaired damaged areas, and crack areas are not classified as similar categories to other categories.

[0122] S5.4. Based on the results of location overlap comparison, determine the confirmed disease results, suspected disease results, or disease removal results, and combine the confirmed disease results, suspected disease results, or disease removal results with the road segmentation markers to generate intelligent road surface disease inspection results.

[0123] Specifically, the methods for determining confirmed disease results, suspected disease results, or disease elimination results shall be implemented according to the following rules: when the location overlap comparison result is a high overlap comparison result and the verification confidence level is not lower than 0.70, the disease result is confirmed; when the location overlap comparison result is a medium overlap comparison result and the verification confidence level is between 0.40 and 0.70, the suspected disease result is determined; when the location overlap comparison result is a low overlap comparison result, or the verification confidence level is lower than 0.40, the disease elimination result is determined.

[0124] Preferably, for misidentified areas caused by expansion joints, marking edges, and drainage ditch edges of urban overpasses, if the reviewed defect area group overlaps with the boundary of road ancillary facilities and the reviewed defect category is crack area, it shall be given priority to be identified as a suspected defect and marked as a manual review object in the intelligent inspection results of road defects.

[0125] Furthermore, the intelligent inspection results of road surface defects include road segmentation markings, defect result categories, defect types, defect locations, defect perimeters, verification confidence levels, location overlap comparison results, and inspection terminal sources. Among these, defect result categories include confirmed defect results, suspected defect results, and defect removal results. The inspection terminal sources include preliminary identification by drone inspection terminals and close-range verification by vehicle-mounted inspection terminals.

[0126] For example, the suspected defect segment index is GJ03-K12+360-K12+380-OUT-R-018, D003, longitudinal 12.4m, transverse 3.2m, crack area, length 1.8m, width 0.03m, confidence level 0.86, IMG02. In the defect area group obtained by the vehicle-mounted inspection terminal, the defect category is crack area, the confidence level is 0.82, and the overlap area ratio is 0.6. 4. If the center offset distance is 0.3m, the defect result is confirmed, and the following intelligent inspection result for road defects is generated: GJ03-K12+360-K12+380-OUT-R-018, confirming the defect result, crack area, longitudinal 12.4m, transverse 3.2m, length 1.8m, width 0.03m, verification confidence level 0.82, high overlap comparison result, preliminary identification by UAV inspection terminal and close-range verification by vehicle-mounted inspection terminal.

[0127] It should be noted that step S5 enables the vehicle-mounted inspection terminal to obtain close-range road surface image data within a clearly defined close-range verification area based on the suspected defect segment index. The final intelligent inspection result of road surface defects is formed through defect verification and location overlap comparison. This solves the problem of false detection caused by the shooting height, bridge shadows, guardrail obstruction, and road marking interference in the drone's overhead image. The defect identification result is verified by close-range image, the false detection area is eliminated, and the confirmed defect result has road segmentation and location basis.

[0128] Reference Figure 2 Other aspects disclosed in this invention also propose an intelligent road surface defect inspection system based on artificial intelligence, including a drone inspection terminal, a vehicle-mounted inspection terminal, and a sidelink communication module; wherein: The UAV inspection terminal includes a first pose processing module, a spectrum status processing module, a beam management module, and a defect identification module; the vehicle-mounted inspection terminal includes a second pose processing module, a close-range verification module, and an inspection result generation module. The first pose processing module and the second pose processing module are connected in communication. They merge the pose information and relative motion physical parameter groups within the same road segment to generate road segment identifiers and collaborative inspection geometric matrices. The spectrum status processing module is connected to the side link communication module. The spectrum status processing module generates a side link spectrum status matrix based on the unlicensed spectrum signal status between the UAV patrol terminal and the vehicle-mounted patrol terminal. The beam management module is connected to the first pose processing module, the second pose processing module, the spectrum status processing module, and the side link communication module respectively. Based on the collaborative inspection geometric matrix and the side link spectrum status matrix, the beam management module determines the beam management parameter group, which includes the beam pointing and beam switching order, and controls the side link communication module to form the side link communication relationship according to the beam management parameter group. The defect identification module is connected to the side link communication module. The defect identification module identifies the road surface image data of road segments, generates a suspected defect segment index, and sends it to the vehicle-mounted inspection terminal via the side link communication module. The close-range verification module is connected to the inspection result generation module. The close-range verification module obtains close-range road surface image data within the road segment defined by the suspected defect segment index and performs verification and identification. The inspection result generation module generates intelligent inspection results of road surface defects based on the verification and identification results.

[0129] It should be recognized that embodiments of the present invention may be implemented or carried out by computer hardware, a combination of hardware and software, or by computer instructions stored in a non-transitory computer-readable storage medium.

[0130] The method can be implemented using standard programming techniques, including a non-transitory computer-readable storage medium configured with a computer program in the computer program, wherein the storage medium is configured such that the computer operates in a specific and predefined manner.

[0131] Each program can be implemented in a high-level procedural or object-oriented programming language to communicate with the computer system; however, if required, the program can be implemented in assembly or machine language.

[0132] In any case, the language can be either compiled or interpreted.

[0133] Furthermore, for this purpose, the program can run on programmed application-specific integrated circuits.

[0134] The processes described herein (or variations and / or combinations thereof) can be executed under the control of one or more computer systems configured with executable instructions, and can be implemented by hardware or a combination thereof as code (e.g., executable instructions, one or more computer programs, or one or more applications) that commonly executes on one or more processors. The computer program includes a plurality of instructions executable by one or more processors.

[0135] Furthermore, the method can be implemented in any suitable computing platform, including but not limited to personal computers, minicomputers, mainframes, workstations, networked or distributed computing environments, standalone or integrated computer platforms, or in communication with charged particle tools or other imaging devices.

[0136] Various aspects of the present invention can be implemented in machine-readable code stored on a non-transitory storage medium or device, whether portable or integrated into a computing platform, such as a hard disk, optical read and / or write storage medium, RAM, ROM, etc., such that it can be read by a programmable computer, and when the storage medium or device is read by the computer, it can be used to configure and operate the computer to perform the processes described herein.

[0137] Furthermore, machine-readable code, or parts thereof, can be transmitted via wired or wireless networks.

[0138] When such media includes instructions or programs that combine with a microprocessor or other data processor to implement the steps described above, the invention described herein includes these and other different types of non-transitory computer-readable storage media.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An intelligent inspection method for road surface defects based on artificial intelligence, characterized in that, include: Based on road mileage, lane direction, and road boundary location, the drone patrol terminal and the vehicle-mounted patrol terminal are assigned to the same road segment, and road segment identifiers are generated. The spatial position, flight attitude, and shooting direction of the UAV patrol terminal are combined with the spatial position, vehicle attitude, and shooting direction of the vehicle-mounted patrol terminal to generate a terminal pose combination. Based on the terminal pose combination, the relative distance, relative height, relative direction and relative speed between the UAV patrol terminal and the vehicle-mounted patrol terminal are determined, and a set of relative motion physical parameters is generated. According to the road segmentation identifier, the terminal pose combination and the relative motion physical parameter group are arranged in columns to generate a collaborative inspection geometric matrix; Based on the unlicensed spectrum signal status between the UAV patrol terminal and the vehicle-mounted patrol terminal, a sidelink spectrum status matrix is ​​generated; Based on the collaborative inspection geometry matrix and the side link spectrum state matrix, a beam management parameter group containing beam pointing and beam switching order is determined, and a side link communication relationship is formed according to the beam management parameter group. Forming the sidelink communication relationship includes: extracting the relative distance, relative altitude, and relative direction between the UAV patrol terminal and the vehicle-mounted patrol terminal from the cooperative patrol geometry matrix to generate a terminal relative orientation group; selecting candidate sidelink channels from the sidelink spectrum state matrix in order of signal-to-noise ratio from high to low and channel occupancy from low to high to generate a target sidelink channel group; determining the beam pointing according to the terminal relative orientation group and the beam switching order according to the channel arrangement order of the target sidelink channel group to generate the beam management parameter group; and pairing the transmit beams and receive beams between the UAV patrol terminal and the vehicle-mounted patrol terminal according to the beam management parameter group to form the sidelink communication relationship. The road surface image data of the road segment is identified by the defect identification model in the UAV inspection terminal, a suspected defect segment index is generated, and the index is sent to the vehicle-mounted inspection terminal through the side link communication relationship. The vehicle-mounted inspection terminal acquires close-range road surface image data within the road segment defined by the suspected defect segment index, performs verification and identification, and generates intelligent road defect inspection results.

2. The intelligent road surface defect inspection method based on artificial intelligence according to claim 1, characterized in that, The collaborative patrol geometric matrix is ​​a 2x2 matrix; where: The first row of the collaborative patrol geometry matrix is ​​the UAV patrol terminal geometry parameter row, which is arranged in the following order: road segment identifier, terminal category identifier, spatial location, terminal height, heading angle, pitch angle, shooting direction, road boundary offset, relative distance to vehicle-mounted patrol terminal, relative height, relative direction, and relative speed. The second row of the collaborative patrol geometry matrix is ​​the vehicle-mounted patrol terminal geometry parameter row, which is arranged in the following order: road segment identifier, terminal category identifier, spatial location, terminal height, vehicle heading angle, vehicle pitch angle, shooting direction, road boundary offset, relative distance, relative height, relative direction, and relative speed with the UAV patrol terminal.

3. The intelligent road surface defect inspection method based on artificial intelligence according to claim 1, characterized in that, Generating the side link spectrum state matrix includes: According to the road segmentation identifier, the unlicensed spectrum channel between the UAV patrol terminal and the vehicle-mounted patrol terminal is divided into a candidate side link channel set; The received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status of each candidate sidelink channel in the candidate sidelink channel set are merged to generate a sidelink signal parameter group. Based on the channel arrangement order of the candidate sidelink channel set, the sidelink signal parameter group is arranged column by column to generate an initial spectrum state matrix; The candidate sidelink spectrum state matrix is ​​generated by removing the rows corresponding to the candidate sidelink channels that are simultaneously in the high-order sorting of channel occupancy and the high-order sorting of retransmission status in the initial spectrum state matrix.

4. The intelligent road surface defect inspection method based on artificial intelligence according to claim 3, characterized in that, The sidelink spectrum state matrix is ​​a multi-row, six-column matrix; where: The number of rows in the sidelink spectrum state matrix is ​​consistent with the number of candidate sidelink channels retained in the sidelink spectrum state matrix, and each row is a spectrum state parameter row for a retained candidate sidelink channel. The six columns of the sidelink spectrum status matrix are arranged in the order of road segmentation identifier, channel arrangement identifier, received signal strength, signal-to-noise ratio, channel occupancy rate, and retransmission status.

5. The intelligent road surface defect inspection method based on artificial intelligence according to claim 1, characterized in that, The pairing process includes: According to the beam direction in the beam management parameter group, a drone transmission beam directed toward the vehicle-mounted patrol terminal is selected from the transmit beam set of the drone patrol terminal, and a vehicle-mounted receive beam directed toward the drone patrol terminal is selected from the receive beam set of the vehicle-mounted patrol terminal, forming a first beam pairing group. According to the beam switching order in the beam management parameter group, a vehicle-mounted transmitting beam toward the UAV patrol terminal is selected from the transmitting beam set of the vehicle-mounted patrol terminal, and a UAV receiving beam toward the vehicle-mounted patrol terminal is selected from the receiving beam set of the UAV patrol terminal, forming a second beam pairing group. The first beam pairing group, the second beam pairing group, and the target side link channel group are combined to form the side link communication relationship.

6. The intelligent road surface defect inspection method based on artificial intelligence according to claim 1, characterized in that, Generating the suspected disease fragment index includes: According to the road segmentation identifier, the road surface image data obtained by the UAV patrol terminal is segmented and arranged to generate road surface segmented image groups; The defect identification model identifies crack areas, pothole areas, subsidence areas, and repair damage areas in the road surface segment image group, generating a defect candidate area group; The candidate disease area group is combined with the road segment identifier, disease location and disease category to generate the suspected disease segment index; According to the aforementioned side link communication relationship, the suspected disease segment index is sent from the UAV patrol terminal to the vehicle-mounted patrol terminal.

7. The intelligent road surface defect inspection method based on artificial intelligence according to claim 6, characterized in that, The intelligent inspection results of the road surface defects are generated, including: The vehicle-mounted inspection terminal determines the close-range verification area based on the road segmentation identifier and the location of the defect in the suspected defect segmentation index; The vehicle-mounted inspection terminal acquires close-range road surface image data within the close-range verification area, and performs regional cropping of the close-range road surface image data according to the location of the defects to generate a verification image region group. The vehicle-mounted inspection terminal performs disease verification and identification on the verification image region group, generates a verification disease region group, and compares the location overlap of the verification disease region group with the disease candidate region group; Based on the results of the location overlap comparison, the confirmed disease results, suspected disease results, or disease removal results are determined, and the confirmed disease results, suspected disease results, or disease removal results are combined with the road segmentation identifier to generate the intelligent road surface disease inspection results.

8. An intelligent road surface defect inspection system based on artificial intelligence, based on the intelligent road surface defect inspection method based on artificial intelligence as described in any one of claims 1 to 7, characterized in that, This includes drone patrol terminals, vehicle-mounted patrol terminals, and sidelink communication modules; among which: The UAV inspection terminal includes a first pose processing module, a spectrum status processing module, a beam management module, and a defect identification module; the vehicle-mounted inspection terminal includes a second pose processing module, a close-range verification module, and an inspection result generation module. The first pose processing module is communicatively connected to the second pose processing module. The two modules merge the pose information and relative motion physical parameter groups within the same road segment to generate road segment identifiers and collaborative inspection geometric matrices. The spectrum status processing module is connected to the side link communication module. The spectrum status processing module generates a side link spectrum status matrix based on the unlicensed spectrum signal status between the UAV patrol terminal and the vehicle-mounted patrol terminal. The beam management module is connected to the first pose processing module, the second pose processing module, the spectrum state processing module, and the side link communication module respectively. Based on the cooperative inspection geometry matrix and the side link spectrum state matrix, the beam management module determines a beam management parameter group that includes beam pointing and beam switching order, and controls the side link communication module to form a side link communication relationship according to the beam management parameter group. The defect identification module is connected to the side link communication module. The defect identification module identifies the road surface image data of the road segment, generates a suspected defect segment index, and sends it to the vehicle-mounted inspection terminal via the side link communication module. The close-range verification module is connected to the inspection result generation module. The close-range verification module obtains close-range road surface image data within the road segment defined by the suspected defect segment index and performs verification and identification. The inspection result generation module generates intelligent inspection results for road surface defects based on the verification and identification results.

Images