Road disease YOLO detection system and method based on lightweight CNN

Abstract

The application discloses a road disease YOLO detection system and method based on a lightweight CNN, and particularly relates to the technical field of computer vision, acquires a road surface image frame and inputs the YOLO detection model based on the lightweight CNN, outputs a disease detection result set, and generates a disease observation record sequence containing a timestamp, a spatial position, a disease category, a confidence score, an event size representation quantity and an observation weight; spatial correlation and aggregation are performed based on the sequence, an event-level disease object is constructed, and an event evidence consistency credibility comprehensive quantity is generated based on the consistency of multiple observations; based on the event-level disease object, the severity of the event, the evolution trend, the road segment attribute and the maintenance resource state information are fused, the credibility comprehensive quantity is taken as a gating item, and the resource cost is taken as a penalty item, so that the urgency and the resource benefit comprehensive quantity are calculated, a work order generation request package is output, and deduplication dispatching and state closed-loop backfilling are performed based on the work order state and a disposal receipt.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and more specifically, to a road defect YOLO detection system and method based on lightweight CNN. Background Technology

[0002] Intelligent road defect detection technology is mainly used to automatically identify, locate, and quantitatively assess abnormal targets such as cracks, potholes, fissures, and subsidence in road surface images or videos, enabling more efficient and reliable closed-loop management of inspections and maintenance. This technology is widely used in scenarios such as daily inspections of municipal roads, maintenance priority assessments, construction quality verification, and rapid response to sudden defects. Among these, end-to-end detection methods, represented by the YOLO-based target detection framework, are often used for online identification by vehicle-mounted inspection terminals or edge devices due to their good real-time performance, reducing the intensity of manual inspections and improving coverage efficiency.

[0003] The existing technology has the following shortcomings:

[0004] Previous road defect detection systems had a rather crude design in the decision-making process from detection results to maintenance work orders. Common practices included triggering alarms or reports solely based on detection confidence thresholds, and simply sorting by category or roughly merging by the number of detections. This failed to incorporate key factors such as defect severity, temporal evolution trends, and road segment importance into a unified decision-making strategy. In application scenarios where a large number of minor cracks, texture interferences, or repair marks exist on the same road segment, and inspection frequency is high, the detection model often outputs a large number of bounding boxes and categories. Without subsequent event merging and severity assessment mechanisms, the system repeatedly reports a large number of low-risk targets as independent events, resulting in a flood of alarms and work orders. Simultaneously, fewer but more dangerous defects in urban areas, such as potholes, are not easily highlighted due to the large number of low-value results, leading to distorted priority in handling.

[0005] To address the above problems, this invention proposes a solution. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a road defect YOLO detection system and method based on lightweight CNN to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] The YOLO method for road defect detection based on lightweight CNN includes the following steps:

[0009] Acquire road surface image frames and input them into a YOLO detection model based on a lightweight CNN, outputting a set of disease detection results; generate a disease observation record sequence based on the disease detection result set, the disease observation record shall at least include timestamp, spatial location field, disease category identifier, confidence score, event size representation quantity and observation weight, and establish an index association between the disease observation record and the original image or evidence cropped image;

[0010] Spatial association and aggregation are performed based on disease observation record sequences to construct event-level disease objects and update their representative locations and dimensions. At the same time, a comprehensive quantity of event evidence consistency and credibility is generated based on the consistency of multiple observations. The consistency of multiple observations includes one or more of the following: confidence stability quantity, category consistency quantity, spatial compactness measure, morphological fidelity measure, and evidence sufficiency measure.

[0011] Based on event-level defects, the system integrates event severity, evolution trend, road segment attribute information, and maintenance resource status information. It uses the overall consistency and credibility of event evidence as a gate item and resource cost as a penalty item to calculate the urgency and resource benefit. Based on this, it outputs a work order request package. The resource cost is determined based on the maintenance resource status information.

[0012] Based on the work order status and disposal receipt, the system performs deduplication and status closed-loop feedback. For event-level disease objects that have been bound to work orders and are in the pending or disposal state, new work orders are not generated repeatedly, but their observation set, event evidence consistency and credibility comprehensive quantity, urgency and resource benefit comprehensive quantity are updated, and the event status is updated according to the disposal receipt.

[0013] In a preferred embodiment, a preset sampling inference strategy is used when acquiring road surface image frames. The sampling inference strategy includes a sampling time interval and a mileage interval.

[0014] In a preferred embodiment, before inputting the road surface image frame into the YOLO detection model based on a lightweight CNN, input normalization processing is performed on the road surface image frame. The input normalization processing includes at least one or more of the following: distortion correction, road surface region of interest selection, input size scaling, and pixel normalization.

[0015] In a preferred embodiment, when generating the disease observation record sequence, the spatial location field is obtained by converting the detection location in the disease detection result set based on the camera intrinsic parameters, camera extrinsic parameters, and road projection model; wherein, the road coordinate system is represented by mileage and lateral coordinates or planar coordinates.

[0016] In a preferred embodiment, the event size characterization is determined based on the defect category identifier. When the defect category identifier corresponds to a pothole, the event size characterization is determined by the coverage area or equivalent diameter of the detection location after projection onto the road surface. When the defect category identifier corresponds to a crack, the event size characterization is determined by the equivalent length, connectivity length, or slenderness ratio of the detection location after projection onto the road surface.

[0017] In a preferred embodiment, the observation weights are determined based on the imaging availability score; the imaging availability score is determined by one or more of the following: image blur level, percentage of overexposed or underexposed pixels, rain / fog intensity indication, and percentage of highly reflective areas; and a lower limit is set for the observation weights to avoid the weighted update failing due to the observation weights being zero.

[0018] In a preferred embodiment, spatial association and aggregation of disease observation record sequences includes establishing an index based on road spatial grids or mileage segments, merging disease observation records into existing event-level disease objects or triggering the creation of new event-level disease objects based on at least one of spatial distance and lateral deviation, and combining category compatibility thresholds to suppress erroneous merging of incompatible categories.

[0019] In a preferred embodiment, when updating the representative location and representative size of the event-level disease object, a weighted statistical update is performed based on the observation weight of the disease observation record, and the weighted effective evidence quantity is used to characterize the sufficiency of evidence for the calculation and gating of evidence sufficiency measurement.

[0020] In a preferred embodiment, the confidence stability measure is determined based on the degree of fluctuation of the confidence score, the category consistency measure is determined based on the degree of support of the disease category identifier in the observation set, the spatial compactness measure is determined based on the spatial dispersion of the spatial location field, the morphological fidelity measure is determined based on the consistency of changes in the event size representation or representative size, and the evidence sufficiency measure is determined based on the weighted effective evidence quantity and the minimum effective evidence quantity threshold; furthermore, the road segment attribute information includes at least the road grade, traffic flow grade and the type of surrounding sensitive points, and the maintenance resource status information includes at least the number of available maintenance teams, the team coverage area, the current pending workload, and the estimated arrival time delay.

[0021] In a preferred embodiment, the following modules are included:

[0022] The detection and documentation module is used to acquire road surface image frames and input them into a YOLO detection model based on a lightweight CNN, and output a set of disease detection results. Based on the set of disease detection results, a sequence of disease observation records is generated. The disease observation records include at least a timestamp, spatial location field, disease category identifier, confidence score, event size representation quantity, and observation weight, and an index association is established between the disease observation records and the original image or evidence cropped image.

[0023] The aggregation and evidence collection module is used to perform spatial association and aggregation based on disease observation record sequences, construct event-level disease objects and update their representative locations and representative dimensions. At the same time, it generates a comprehensive quantity of event evidence consistency and credibility based on the consistency of multiple observations. The consistency of multiple observations includes one or more of the following: confidence stability quantity, category consistency quantity, spatial compactness measure, morphological fidelity measure, and evidence sufficiency measure.

[0024] The priority order generation module is used to calculate the urgency and resource benefit based on the event-level disease object, integrate event severity, evolution trend, road segment attribute information and maintenance resource status information, and use the event evidence consistency and credibility comprehensive value as a gate item and resource cost as a penalty item, and output the work order generation request package accordingly; wherein the resource cost is determined based on the maintenance resource status information.

[0025] The deduplication and closed-loop backfeed module is used to perform deduplication of work orders and closed-loop status backfeed based on work order status and disposal receipt. For event-level disease objects that have been bound to work orders and are in the pending or disposal state, new work orders are not generated repeatedly, but their observation set, event evidence consistency and credibility comprehensive quantity, urgency and resource benefit comprehensive quantity are updated, and the event status is updated according to the disposal receipt.

[0026] The technical effects and advantages of this invention are as follows:

[0027] This invention performs input normalization on road surface image frames through a detection and filing module on the inspection terminal side, and inputs the image frames into a YOLO detection model based on a lightweight CNN to output a set of disease detection results. It further generates a disease observation record sequence containing timestamps, spatial location fields, disease category identifiers, confidence scores, event size representation quantities, and observation weights, and establishes an index association with the original image or evidence cropped image, thereby enhancing the traceability of evidence while ensuring real-time performance.

[0028] This invention uses an aggregation and evidence collection module to spatially correlate and aggregate disease observation record sequences, constructing event-level disease objects and continuously updating representative locations and dimensions. Simultaneously, it generates a comprehensive quantity of event evidence consistency and credibility based on the consistency of multiple observations. Consistency can be characterized by one or more of the following: confidence stability quantity, category consistency quantity, spatial compactness measure, morphological fidelity measure, and evidence sufficiency measure. This allows the same disease to be uniformly characterized under multiple frames and multiple observations, reducing interference from occasional false detections and duplicate detections, and improving the stability and usability of event-level results.

[0029] This invention integrates event severity, evolution trend, road segment attribute information, and maintenance resource status information under gating constraints by calculating the priority of work orders. It also uses resource cost as a penalty to calculate the urgency and the comprehensive amount of resource benefits and outputs a work order generation request package. This allows the work order priority to reflect both the reliability of event evidence and the benefits of handling the event, reducing blind work order assignment and resource crowding.

[0030] This invention utilizes a deduplication closed-loop recharge module to perform deduplication of work orders and status closed-loop recharge based on work order status and handling receipts. For event-level defects that are already bound to work orders and are in a pending or handling state, new work orders are not generated repeatedly. Instead, the observation set, the overall quantity of consistency and credibility of event evidence, the overall quantity of urgency and resource benefits are updated, and the event status is updated based on the handling receipts. This forms a closed-loop management of detection, order issuance and handling, improving maintenance response efficiency and management consistency. Attached Figure Description

[0031] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings;

[0032] Figure 1 This is a flowchart illustrating the YOLO detection method for road defects based on lightweight CNN of the present invention.

[0033] Figure 2 This is a schematic diagram of the YOLO road defect detection system based on lightweight CNN of the present invention. Detailed Implementation

[0034] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] Example 1: The present invention provides a YOLO-based road defect detection method using lightweight CNN, such as... Figure 1 As shown, it includes the following steps:

[0036] Step 1: Detect and create a file;

[0037] This step takes an onboard vision system deployed on a municipal road inspection vehicle or road maintenance inspection terminal as an example. The input is the road surface image and its positioning and attitude information acquired during the inspection, including road surface image frames acquired by forward or downward industrial cameras, acquisition timestamps, vehicle positioning and attitude information, camera intrinsic and extrinsic calibration parameters, sampling and inference strategies after system startup, and acquisition and inference according to the preset sampling strategy after the current inspection task is completed. The sampling strategy can be set as a sampling time interval or mileage interval to ensure that the road surface coverage density during the inspection can meet the needs of subsequent event-level aggregation and work order decision-making. In addition, considering the impact of imaging fluctuations caused by rain, fog, nighttime supplementary lighting, glare, and vehicle vibration, each frame of image is first subjected to relatively simple input normalization, such as distortion correction according to the camera calibration parameters and selection of the region of interest on the road surface, scaling the image to the input size required by the YOLO detection model and performing pixel normalization processing to ensure that the input distribution is basically the same under different terminals and different lighting conditions.

[0038] After input preparation, the system inputs the normalized road surface image into a YOLO detection model based on a lightweight CNN for forward inference, obtaining a set of road defect detection results for that frame. The set of road defect detection results includes at least: a defect category identifier, bounding box position, and confidence score for each detected target, where the confidence score characterizes the degree to which the detected target belongs to the defect category. To facilitate the subsequent convergence of single-frame detected targets into event-level defect objects, the system further encapsulates each detected target as a defect observation record, completing two key assignments during generation: spatialization and quality enhancement. First, based on camera intrinsic and extrinsic parameters and the road surface projection model, the system converts the position of the detection bounding box in the image coordinate system into a candidate point in the road coordinate system, using this point as the spatial location field of the defect observation record. The road coordinate system can be represented using mileage and lateral coordinates or planar coordinates, enabling the correlation of observations of the same road segment in different frames and different inspection trips within a unified coordinate system. Secondly, the system generates event size representations for each disease observation record to characterize the disease scale and provide input for subsequent severity assessment. When the disease type is a pothole, the event size representation can be determined by the coverage area or equivalent diameter after the bounding box is projected; when the disease type is a crack, the event size representation can be determined by quantitative indicators such as the equivalent length, connectivity length, or slenderness ratio after projection. These size representations do not require complex 3D reconstruction and are preferably calculated in real-time on the vehicle-mounted device based on camera calibration and road surface geometric assumptions, thus meeting the requirements for lightweight deployment.

[0039] Meanwhile, to proactively reduce false detections and duplicate reporting caused by low-quality frames in subsequent event aggregation and deduplication dispatch, the system generates an imaging availability score when generating disease observation records, and uses this score as the quality field of the disease observation records. The imaging availability score refers to the usability of the current frame image in terms of sharpness, exposure, and environmental interference. It normalizes statistics such as the degree of image blur, the proportion of overexposed or underexposed pixels, rain / fog intensity, and the proportion of highly reflective areas to [0,1]. A value closer to 1 indicates a clearer and more stable image for that frame, while a value closer to 0 indicates that the frame is affected by blur, glare, or rain / fog, and its weight should be reduced in subsequent statistics and decisions. To allow this quality field to be directly used as observation weights in subsequent aggregation calculations, the system can map the imaging availability score to observation weights. The calculation method is as follows: ;

[0040] in, This represents the observation weight of the i-th disease observation record. This indicates the imaging usability score for the corresponding frame of the observation record. Indicates the lower bound of the weight and satisfies This is used to avoid situations where the weight is zero in extreme cases, causing subsequent weighted statistics to fail to update. For example... It is 0.1.

[0041] Subsequently, the system writes the disease observation records with fields such as timestamp, spatial location, disease category, confidence score, event size representation, and observation weight into the result buffer or local database, forming an index relationship with the original image or evidence cropped image, generating a traceable observation sequence output, which serves as a unified input for subsequent steps such as event-level disease object construction, evidence consistency verification, and work order priority generation. It can maintain a stable detection, aggregation, and decision-making link in road environments with high inspection frequency and complex diseases.

[0042] Step Two: Aggregated Evidence Collection;

[0043] This step processes the sequence of road defect observation records output from step one and their corresponding road spatial index information. Inputs include: the spatial location field (road coordinates) for each defect observation record, defect category identifier, confidence score, event size representation, imaging availability score, and the observation weights mapped from it. Optional inputs include road spatial grid or mileage segment division rules and positioning jitter baseline statistical parameters. Because the same road segment generates a large number of repeated observations under high-frequency inspection conditions, and the lightweight YOLO detection output exhibits positional drift and confidence fluctuations at the single-frame level, directly using single-frame detection targets as reporting units could easily lead to minor network cracks or repair marks being split into numerous independent events and reported excessively. Therefore, this step converges the defect observation records into event-level defect objects at the result layer and constructs a comprehensive event evidence consistency and credibility quantity based on the consistency of multiple observations, thereby providing stable, interpretable, and controllable event-level input for subsequent work order priority calculations.

[0044] In terms of event-level aggregation, the system constructs a defect event map at the result layer and uses event-level defect objects as management units. First, the system indexes defect observation records according to road spatial grids or mileage segments, classifying observation records falling into the same or adjacent grids as candidate association sets. Then, the system matches candidate association sets based on spatial distance and lateral deviation, preferentially using Euclidean distance or mileage difference in the road coordinate system as the spatial distance metric, while also considering the road's lateral position to distinguish different lanes or lateral areas. When a defect observation record meets preset thresholds in terms of spatial distance, lateral deviation, and category compatibility with an existing event-level defect object, the system merges the observation record into the observation set of that event-level defect object. When no existing event object can be matched, the system creates a new event-level defect object starting from that observation record and initializes its observation set and representative attributes. Through these spatial association and category compatibility constraints, the system can converge observations of the same defect repeatedly occurring in consecutive frames and multiple vehicle trips into the same event object, while avoiding the erroneous merging of defects from different lanes or locations.

[0045] To ensure the stability of the spatial location and size representation of event objects across multiple observation updates, the system uses a weighted statistical update of representative locations and sizes based on observation weights for the observation records of incorporated event objects. Specifically, for a certain event-level disease object e, its current observation set is denoted as... Where N represents the total number of observation records associated with the event object; let the road coordinates of the i-th observation record be denoted as . The observation weight is The event size representation quantity is The system can define the representative location of an event object as the weighting center. Its update method is ;

[0046] in, This indicates the representative position of event-level disease object e. Represents observation records The road coordinates, Represents observation records Observation weights; denominator This represents the weighted amount of valid evidence, used to avoid low-quality observations from overly influencing the location of events. Similarly, the system can define the representativeness of an event object as the weighted mean. The calculation method is as follows: ;

[0047] in, This represents the representative size of the event-level defect object e. Represents observation records The event size representation quantity. Through the above weighted update, the system introduces imaging availability into the event-level aggregation process without increasing the computational burden on the model, thereby reducing the disturbance of the position and size of event objects by glare, rain, fog or blurred frames.

[0048] After completing event-level aggregation and obtaining representative locations and dimensions, the system further constructs a comprehensive quantity of event evidence consistency and credibility based on the consistency of multiple observations. This is used to characterize whether the evidence for the event is stable and reliable. To this end, the system jointly characterizes the confidence level, category, spatial location, and size variations within the observation set. First, the system calculates the weighted mean of the event confidence level. With weighted variance Based on this, the confidence stability quantity is defined. : ,

[0049] ,

[0050] ;

[0051] in, Represents observation records Confidence score, It is an exponential function. The value range is (0,1], and the larger the variance, the more drastic the fluctuation. The closer it is to 0. Secondly, the system uses a class consistency measure. This characterizes whether the main category of an event is stable. For each category c, its weighted proportion is defined. for ;

[0052] in, Represents observation records Disease category identification, An indicator function that returns 1 if the condition within the parentheses is true, and 0 otherwise. This indicates taking the value corresponding to the largest proportion among all categories. A larger value indicates a more stable event category. Furthermore, this is a measure of the system's computational space compactness. To identify spatial dispersion caused by positioning drift or false detection. The system obtains representative locations. Then, calculate the spatial dispersion. : ,

[0053] ;

[0054] in, Describing the vector norm, This refers to the positioning jitter baseline pre-calculated for this road segment or positioning mode, used to distinguish between normal positioning fluctuations and abnormal dispersion. The positioning jitter baseline... The variance of positioning errors can be obtained by statistically analyzing historical positioning data under the same road segment and equipment conditions. For example, one could take... ; To prevent small quantities, for example =0.00000000001. Furthermore, the system's structural morphological fidelity metric... Used to constrain the reasonableness of dimensional changes in the same event across multiple observations: ;

[0055] in, Indicates the size dispersion. For representative dimensions of the event, To prevent small quantities, The larger the value, the more consistent the size variation.

[0056] In addition to the consistency characterization mentioned above, the system also introduces a measure of sufficiency of evidence. To reduce the excessive reporting of minor defects during frequent inspections, the system defines the weighted amount of valid evidence as... , and according to Calculate the sufficiency of evidence, where This indicates the minimum threshold for valid evidence, and it varies with the urgency of the event. Adaptive change.

[0057] To avoid cross-step dependencies and ensure process feasibility, the pre-urgency used in step two for the adaptive threshold is... Based solely on the event primary category and representativeness size obtained in step two, the system first determines the event primary category using the category consistency quantity. And configure category risk coefficients for different main categories. Then take representative dimensions Size-normalized mapping function Mapped to Thus obtain .

[0058] in, This can be achieved using piecewise linear normalization, for example: when hour ;when hour ;otherwise The system adopts accordingly. A minimum threshold for valid evidence is set so that high-risk events can obtain sufficient evidence with fewer observations, while low-risk events require more stable evidence to increase credibility; and a formal coarse urgency is obtained in step three. Then, the threshold was updated to And based on this, and the overall consistency and credibility of event evidence Perform a back-calculation calibration update to ensure consistency between evidence gating and subsequent priority calculations.

[0059] Finally, the system summarizes the above sub-quantities using a weighted geometric average to obtain the overall consistency and credibility of the event evidence: ;

[0060] in, to These are positive weighting coefficients used to reflect the importance of confidence stability, class consistency, spatial compactness, morphological fidelity, and sufficiency of evidence in different road environments; typical values ​​are... This is the equal-weighted geometric average. Through this step, the system outputs event-level disease objects and their representative locations and dimensions. Furthermore, an index is established between the event object and the original image or the cropped evidence image, enabling subsequent steps to prioritize and de-duplicate orders at the event level, thereby reducing the risk of minor defects flooding the screen and duplicate work orders from a mechanism perspective.

[0061] Step 3: Calculate order priority;

[0062] This step processes the event-level disease objects output from step two, along with their representative locations, representative dimensions, observation set statistics, and the overall consistency and credibility of the event evidence. The input also includes a road infrastructure information database and maintenance resource status information. The road infrastructure information database includes at least road segment attribute information such as road grade, number of lanes, traffic flow level, types of surrounding sensitive points, and management jurisdiction; the maintenance resource status information includes at least scheduling information such as the number of available maintenance teams, team coverage area, current pending workload, and estimated arrival delay. Because the number of minor cracks and repair marks on the same road segment is large under high-frequency inspection conditions, while the number of dangerous potholes is small but the risk is high, relying solely on reporting rules based on categories plus single-frame confidence thresholds can easily lead to work order overload and priority distortion. Therefore, this step uniformly quantifies event severity, evolution trend, road segment importance, and resource cost at the event-level granularity, and... Priority calculation is introduced as evidence gating to form a sortable queue of work order candidates and a structured work order generation request package.

[0063] To obtain the severity of an event, the system first determines the main category of the event. For a specific event-level defect object e, its observation set is as follows: Each observation record contains a category identifier. With observation weight The system uses a weighted average. Describe the degree of support for category c in this event: ;

[0064] in, This is an indicator function; it takes the value 1 when the observed category is c, and 0 otherwise. The system selects the category with the largest proportion as the main event category. And configure category risk coefficients for different main categories. This is used to reflect the differences in risk to driving safety and road structure posed by different types of defects. The system further increases the representativeness of events. Mapped to size normalization ,in As a monotonic mapping function, it can be implemented using piecewise linear or normalization based on historical quantiles, thereby unifying the dimensions of different road segments and sampling conditions; the aforementioned Piecewise linear normalization can be used, for example: when hour, ;when hour, ;otherwise ;in and It can be set based on historical data, for example , Based on the above quantities, the system defines event severity. for:

[0065] ;

[0066] in, The preferred value range is [0,1]. A larger value indicates a larger event scale and a higher type of risk.

[0067] To depict the evolution trend of events, the system maintains the trajectory of the change of representative dimensions of the events over time within a preset time window and calculates the growth rate. Specifically, the representative size of the current window's events is denoted as... The representative size of the previous window is The time interval between the two windows is The system can obtain the growth rate as follows: ;

[0068] in, To prevent small quantities, This means focusing only on positive growth that indicates a deterioration, avoiding the negative impact of size reduction after repairs on urgency. By using the growth rate, the system can distinguish between stable but slowly changing minor defects and rapidly worsening defects, prioritizing resources for events with escalating risk.

[0069] To reflect the importance of road sections, the system identifies representative locations of event-level defects. Mapping to the road infrastructure database, the corresponding road grade, traffic flow grade, and surrounding sensitive point attributes are obtained, and road weights are formed accordingly. Road weights can be constructed using a multiplicative combination method: ;

[0070] in, , , These are discrete gear mapping functions, used to map road grades. Traffic level With sensitive point type The mapping is to numerical weights; a larger weight indicates a higher value for handling the road segment or a greater safety impact. (Road grade mapping function) This can be set as follows: Expressway = 1.5, Arterial Road = 1.2, Secondary Arterial Road = 1.0, Local Road = 0.8. Traffic Flow Level Mapping Function It can be set as follows: High flow = 1.3, Medium flow = 1.0, Low flow = 0.7. Sensitive point type mapping function. The settings can be configured as follows: areas around schools or hospitals = 1.4, commercial areas = 1.2, and general areas = 1. By introducing road network management differences through road weights, the same defects will receive higher priority for treatment on main roads, bus routes, or roads around schools and hospitals, while a relatively conservative dispatching strategy will be maintained on low-traffic side roads.

[0071] Given the limited maintenance resources and the need for feasible dispatching, the system also incorporates resource costs into the unified decision-making process. The system calculates the estimated arrival time based on the event location and the team's coverage area. Distance from the expected journey The team congestion level B is obtained by combining the current team's pending task load. The team congestion level B can be a normalized value of the number of pending work orders or the occupancy rate; thus constructing the resource cost. : ;

[0072] in, , , A positive coefficient is used to balance the contribution of arrival delay, route distance, and team load to resource costs. , , It can be set according to the actual scheduling cost, with a typical value being [value to be filled in]. , , .

[0073] After obtaining the severity of the incident ,growth rate Road weight With resource costs Then, the system first calculates the coarse urgency without introducing evidence gating. This is used to characterize the intensity of the action drive for an event without considering the strength of the evidence: ;

[0074] in, , These are positive weighting coefficients, where , , For the growth rate compression function, it is preferable to use... In order to limit the amplifying effect of extreme growth rates on ranking, A positive smoothing parameter, for example The aforementioned coarseness and tightness On the one hand, it is used for the back-calculation and calibration of evidence gating parameters in step three: step two first relies on the pre-urgency. set up And calculate the sufficiency of evidence. Consistency and credibility of evidence in relation to events Step 3 involves obtaining the road weights. The formal calculation is performed after considering information such as growth rate. The minimum threshold for valid evidence has been updated to... Based on this, and Perform a back-calculation update; on the other hand... The final priority calculation used in this step ensures that evidence gating and priority ranking remain consistent under the same urgency caliber.

[0075] To incorporate the reliability of evidence into decision-making and mitigate order dispatch distortion caused by excessive posting of minor defects, the system will incorporate the overall consistency and credibility of event evidence output from step two. Urgency is introduced as a gating factor in the calculation, and resource cost is used as a penalty factor to obtain a combined urgency and resource benefit. : ;

[0076] in, This is a positive parameter used to adjust the strength of the influence of evidence credibility on priority. A larger value indicates a greater inclination to prioritize events with solid evidence, such as... ; A positive parameter is used to adjust the severity of resource cost penalties. A larger value indicates a greater tendency to suppress high-cost order assignments when resources are scarce, for example... .because It also integrates confidence stability, class consistency, spatial compactness, morphological fidelity, and sufficiency of evidence, therefore introducing... It can effectively suppress sporadic high-confidence but unstable targets in a single frame, making it easier to obtain high confidence for dangerous pits when the evidence is stable. Minor network cracks or repair marks are unlikely to be included in the high-priority queue when there is insufficient evidence or significant fluctuations, thus achieving resource control and risk priority.

[0077] Ultimately, the system based on The event-level defects are sorted to generate a candidate queue of work orders, and a work order generation request package is then output to the maintenance management platform. The work order generation request package includes at least the event identifier, representative location of the event, main event category, and event severity. ,growth rate Road weight The combined effect of urgency and resource benefits The system includes suggested processing timelines and evidence index information; the evidence index information preferably includes an index of multiple evidence images, key observation timestamps, and observation location trajectories for rapid review by the management end. The system can also... and Combined setting of processing type: when Exceeding the emergency repair threshold and When the value is above the confidence threshold, an emergency repair work order is generated; when Higher but When the data falls below the trusted threshold, a review work order is generated or a supplementary data collection strategy is triggered; when When the threshold is low, the event is added to the observation queue and only a status update is performed without creating a new work order. The emergency repair threshold can be set to 0.8, and the reliability threshold can be set to 0.7.

[0078] Through the above-mentioned event-level quantification and gating decision-making, this step realizes the executable conversion from detection results to maintenance work orders, and alleviates the problem of minor defects flooding the screen while serious defects are overwhelmed from a mechanism perspective.

[0079] Step 4: Deduplication and closed-loop recharge;

[0080] This step processes the work order candidate queue and work order generation request package output from step three, as well as the event-level disease objects maintained in step two. Inputs include: the event identifier and representative location of the event-level disease object. Representative dimensions Consistency and credibility of event evidence (comprehensive score) The combined effect of urgency and resource benefits The system inputs the event's main category and observation set statistics, along with the work order status and handling receipt information returned by the maintenance management platform. Because the same event may be repeatedly observed within the handling cycle under high-frequency inspection conditions, without deduplication and state machine constraints, the system will continuously create new work orders, leading to duplicate dispatches, duplicate vehicle deployments, and resource congestion. Furthermore, the urgency of some high-risk events may increase dramatically as evidence accumulates; without an escalation channel, the handling of truly dangerous defects may be delayed. Therefore, this step establishes a work order deduplication and status closed-loop mechanism using the event-level defect object as the unique primary key. This allows the system to control the number of work orders without sacrificing high-risk response capabilities and uses handling receipts to traceably calibrate strategy parameters.

[0081] In terms of deduplication and dispatching, the system maintains the work order binding status for each event-level defect object and establishes a one-to-one or one-to-many controlled association between the event identifier and the work order identifier. For any event-level defect object e, when it is not yet bound to a valid work order, the system allows the creation of a new work order based on the work order generation request package in step three; when it is already bound to a work order and the work order is in a pending or processing state, the system will not create a new work order when the event is observed in subsequent inspections, but will instead merge the new observation record into the observation set of the event. And update the representative position of the event according to the rules in steps two and three. Representative dimensions Consistency and credibility of event evidence (comprehensive score) Combined with urgency and resource benefits To avoid erroneous upgrades caused by fluctuations in a single frame, the system preferentially uses statistical measures from multiple observations as the basis for upgrade triggering, for example, requiring... Only when the level remains consistently above the upgrade threshold or its increase exceeds a preset threshold in a series of valid observations will existing work orders be prioritized or have their processing time tightened. By implementing a constraint of updating only and not creating new ones, the system transforms repeated observations of the same event within the processing cycle into supplementary information and evidence enhancement for existing work orders, thereby suppressing minor issues from flooding the screen and duplicate work orders.

[0082] For event-level disease objects, a closed-loop status is established, with at least one status set: observed, dispatched and awaiting treatment, in progress, closed-loop, and false alarm archived, and status transition conditions are defined. Upon receiving a new observation update, the system, based on the event... and Reset the incident handling strategy: Above the emergency repair threshold When the value is below the trust threshold, the event is preferably placed in the "dispatch pending" or "processing" state. Above the confidence threshold When the event is below the confidence threshold, the event is preferred to enter the review dispatch strategy. A review evidence index can be added to the work order or a supplementary collection strategy can be triggered to improve the effective evidence of subsequent observations. When the value falls below the confidence threshold, the event remains in the observation state, and the observation set and trend value are no longer updated; a new work order is generated. To maintain consistency between the state machine and priorities, work order priority and processing time limit can be used as state attributes. Dynamic adjustments, for example, if the situation worsens during the handling of the event. When the risk level rises above the credibility threshold, the original work order will be upgraded and a change notification will be sent to prevent high-risk events from being moved to higher priority as evidence accumulates or risks increase, and to prevent serious problems from being suppressed in the data deluge for a long time.

[0083] Regarding the feedback of treatment receipts, after receiving a work order, the maintenance management platform sends back a treatment receipt. The treatment receipt should include information such as the work order identifier, arrival time, treatment result type, before-and-after photos or video evidence, and the level of on-site confirmation of the defect. The treatment result type should include repair completed, no treatment required (false alarm or repaired), and observation postponed. Upon receiving the treatment receipt, the maintenance management platform records it in the event-level defect object file based on the binding relationship between the work order identifier and the event identifier, and updates the event status: when the treatment result is repair completed, the event status is set to closed loop, and the closed loop time and before-and-after evidence are recorded; when the treatment result is a false alarm or no treatment required, the event status is set to false alarm archived, and the corresponding observation segment, imaging quality statistics, and category confusion information are recorded in the hard sample pool; when the treatment result is observation postponed, the system retains the event object and accumulates observation evidence and trend data during inspections. To facilitate backtracking, the system will generate traceable records for every change in the state, priority, and evidence of the event object, and associate them with timestamps, terminal sources, and evidence indexes to form a closed-loop evidence chain.

[0084] In terms of strategy calibration, the system performs controlled updates to parameters by handling feedback and difficult sample pools. For example, for false alarm archived events, the system statistically analyzes the distribution of each sub-quantity and imaging availability score in step two, and adjusts its overall evidence consistency and credibility. Weight parameters to A credibility threshold makes it easier to reduce credibility and enter the review or observation strategy under similar imaging conditions or similar texture backgrounds; for high-risk events that are closed-loop and confirmed on-site, extract the combined features of representative size, growth rate, road weight and resource cost, and adjust the urgency and comprehensive resource benefit in step three. The weight parameters or emergency repair thresholds make it easier to prioritize events under similar risk patterns. If the system supports offline training, the difficult sample pool can be used as incremental training or distillation input to update the parameters or post-processing strategies of the lightweight YOLO detection model. Version management links model updates with business effects for verification, making model updates rollbackable and verifiable.

[0085] Through the aforementioned deduplication and dispatching, status closure and feedback calibration mechanisms, the system uses event-level defects as the main thread to connect the entire process of detection, aggregation, decision-making, treatment and feedback. This ensures that repeated observations under high-frequency inspection conditions are used to enhance existing work order evidence rather than generate new work orders. At the same time, it ensures that dangerous defects can be upgraded to higher levels of treatment in a timely manner when the risk increases or the evidence is sufficient. Thus, under the reality of limited maintenance resources, the number of work orders can be controlled, the treatment priority can be explained and the closed-loop evidence can be traced.

[0086] Example 2: The design of the road defect YOLO detection system based on lightweight CNN of the present invention is based on the method in Example 1, specifically as follows... Figure 2 The following modules are shown:

[0087] The detection and documentation module is used to complete input standardization and lightweight YOLO detection and observation documentation on the inspection side. This module uses the onboard vision system of a municipal road inspection vehicle or road maintenance inspection terminal as the input side, receiving road surface image frames, acquisition timestamps, vehicle positioning and attitude information, camera intrinsic and extrinsic calibration parameters, and sampling inference strategies. It then performs acquisition and inference according to sampling time intervals or mileage intervals. For imaging fluctuations caused by rain, fog, nighttime lighting, glare, and vehicle vibration, it first performs distortion correction, road surface region of interest selection, input size scaling, and pixel normalization. The standardized images are then input into a lightweight CNN-based YOLO detection model to obtain a set of road defect detection results. This module further encapsulates single-frame detected targets into disease observation records. During generation, it writes disease category identifiers, bounding box positions, and confidence scores. Based on camera calibration and road projection models, it converts the detection positions into candidate points in the road coordinate system as spatial location fields, while also providing event size representations and generating imaging usability scores and observation weights mapped from them. Finally, it writes the disease observation records into the results buffer or local database, establishes an index relationship with the original images or evidence cropped images, and outputs a traceable observation sequence for subsequent event-level aggregation and dispatch decisions.

[0088] The aggregation and evidence collection module is used to converge the disease observation record sequence into event-level disease objects and form a comprehensive quantity of event evidence consistency and credibility. This module receives the road coordinate points, disease category identifiers, confidence scores, event size representation quantities, imaging availability scores, and observation weights of the disease observation records. It can also optionally receive road spatial grid or mileage segment division rules and positioning jitter baseline statistical parameters. In the results layer, a disease event map is constructed. After indexing by spatial grid / mileage segment, candidate observations are matched for spatial distance and lateral deviation. Combined with the category compatibility threshold, the observations are merged into existing event-level disease objects or trigger the creation of new event objects. To maintain the stability of the spatial location and size representation of event objects, this module employs weighted statistics based on observation weights for incorporated observations, continuously updating the representative location and size of event objects. Subsequently, based on the consistency of the observation set in terms of confidence, category, spatial location, and size changes, a comprehensive quantity of event evidence consistency and credibility is constructed, and an evidence sufficiency measure is introduced to suppress the excessive reporting of minor defects. The pre-urgency used for the adaptive threshold is determined only by the main event category and representative size, and can be back-calculated and updated after the formal coarse urgency is obtained in subsequent modules. This module outputs event-level defect objects and their representative locations, representative sizes, and the comprehensive quantity of event evidence consistency and credibility, while maintaining the index association between event objects and evidence images.

[0089] The priority order generation module is used to uniformly quantify severity, evolution trend, road segment importance, and resource cost at the event-level granularity, and generate a sortable queue of candidate work orders and work order generation request packets. This module receives event-level defect objects, representative locations, representative dimensions, observation set statistics, and a comprehensive quantity of event evidence consistency and credibility. It also receives road infrastructure information and maintenance resource status information. The road infrastructure information database includes at least information such as road grade, number of lanes, traffic flow level, types of surrounding sensitive points, and management jurisdiction. The maintenance resource status information includes at least information such as the number of available maintenance teams, team coverage, current pending workload, and estimated arrival delay. The module first determines the main category of the event and combines the category risk coefficient with size normalization to obtain the event severity. Then, it maintains the change trajectory of the representative dimensions of the event within a preset time window to obtain the growth rate. Subsequently, it maps the representative locations of the event to the road infrastructure information database to form road weights, and constructs resource costs by combining estimated arrival time, estimated travel distance, and team congestion. Based on this, a coarse urgency level is first obtained to complete the back-calculation calibration of the aforementioned evidence threshold. Then, the overall consistency and credibility of event evidence is introduced as a gating item into the priority calculation, and resource cost is used as a penalty item to obtain the overall urgency and resource benefit. Finally, a work order candidate queue is formed according to the order, and a work order generation request package is generated and output to the maintenance management platform. The request package carries at least the event identifier, representative location of the event, main category of the event, severity of the event, growth rate, road weight, overall urgency and resource benefit, suggested handling time limit, and evidence index information.

[0090] The deduplication and closed-loop feedback module is used to establish a work order deduplication, status closure, and treatment receipt feedback calibration mechanism based on event-level disease objects. This module receives the work order candidate queue, work order generation request packets, event-level disease objects and their event identifiers, representative locations, representative dimensions, a comprehensive measure of event evidence consistency and credibility, a comprehensive measure of urgency and resource benefits, event main category, and observation set statistics. It also receives work order status and treatment receipt information from the maintenance management platform. The module maintains the work order binding status for each event-level disease object. When an event is bound and the work order is in a pending or processing state, subsequent inspections will not create new work orders but will instead merge them into the observation set and update the event statistics and priority-related quantities according to predetermined rules. Furthermore, when multiple valid observations are stable, the module will trigger work order priority upgrades or treatment time limits tightening. The module also maintains an event state machine, including at least the following states: observation, dispatched and awaiting disposal, disposal in progress, closed loop, and false alarm archived. It performs state transitions and strategy adjustments based on the urgency, resource benefits, and credibility gating results. In terms of disposal receipt feedback, the module writes the work order identifier, arrival time, disposal result type, evidence before and after disposal, and the level of on-site disease confirmation into the event file. False alarms or samples that do not require disposal are included in the hard sample pool for controlled calibration of the overall consistency and credibility of event evidence and related parameters of urgency. When necessary, it supports offline training and version management, making updates verifiable and rollbackable, and forming a closed-loop evidence chain that connects detection, aggregation, decision-making, disposal, and feedback.

[0091] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0093] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0094] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0095] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A YOLO-based road defect detection method using lightweight CNN, characterized in that, Includes the following steps: Acquire road surface image frames and input them into a YOLO detection model based on a lightweight CNN, outputting a set of disease detection results; generate a disease observation record sequence based on the disease detection result set, the disease observation record shall at least include timestamp, spatial location field, disease category identifier, confidence score, event size representation quantity and observation weight, and establish an index association between the disease observation record and the original image or evidence cropped image; The observation weights are determined based on the imaging availability score; the imaging availability score is determined by one or more of the following: image blur level, percentage of overexposed or underexposed pixels, rain / fog intensity indication, and percentage of highly reflective areas; and a lower limit is set for the observation weights to prevent the weighted update from failing due to zero observation weights; Spatial association and aggregation are performed based on disease observation record sequences to construct event-level disease objects and update their representative locations and sizes. At the same time, a comprehensive quantity of event evidence consistency and credibility is generated based on the consistency of multiple observations. Consistency across multiple observations includes one or more of the following: confidence stability measure, class consistency measure, spatial compactness measure, morphological fidelity measure, and sufficiency of evidence measure; wherein, the sufficiency of evidence measure is determined by defining the sum of the observation weights of each observation record in the observation set as the weighted valid evidence measure, denoted as . Set a minimum threshold for valid evidence. This threshold is based on the urgency of the event-level disease object. Adaptive change; the evidence is adequately measured Calculated by the following formula: The aforementioned pre-urgency Determined based on the main event category and representative size; Based on event-level defects, the system integrates event severity, evolution trend, road segment attribute information, and maintenance resource status information. It uses the overall consistency and credibility of event evidence as a gate item and resource cost as a penalty item to calculate the urgency and resource benefit. Based on this, it outputs a work order request package. The resource cost is determined based on the maintenance resource status information. Based on the work order status and disposal receipt, the system performs deduplication and status closed-loop feedback. For event-level disease objects that have been bound to work orders and are in the pending or disposal state, new work orders are not generated repeatedly, but their observation set, event evidence consistency and credibility comprehensive quantity, urgency and resource benefit comprehensive quantity are updated, and the event status is updated according to the disposal receipt.

2. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: When acquiring road surface image frames, a preset sampling and inference strategy is adopted, which includes sampling time interval and mileage interval.

3. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: Before inputting road surface image frames into the YOLO detection model based on a lightweight CNN, input normalization processing is performed on the road surface image frames. Input normalization processing includes at least one or more of the following: distortion correction, road surface region of interest selection, input size scaling, and pixel normalization.

4. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: When generating the disease observation record sequence, the spatial location field is calculated from the detection location in the disease detection result set based on the camera intrinsic parameters, camera extrinsic parameters and road projection model; among them, the road coordinate system is represented by mileage and lateral coordinates or planar coordinates.

5. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: The event size characterization is determined based on the defect category identifier. When the defect category identifier corresponds to a pothole, the event size characterization is determined by the coverage area or equivalent diameter of the detection location after projection onto the road surface. When the defect category identifier corresponds to a crack, the event size characterization is determined by the equivalent length, connected length, or slenderness ratio of the detection location after projection onto the road surface.

6. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: Spatial association and aggregation of disease observation record sequences include establishing an index based on road spatial grid or mileage segment, merging disease observation records into existing event-level disease objects or triggering the creation of new event-level disease objects based on at least one of spatial distance and lateral deviation, and combining category compatibility thresholds to suppress the erroneous merging of incompatible categories.

7. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: When updating the representative location and representative size of event-level disease objects, a weighted statistical update is performed based on the observation weight of the disease observation record, and the weighted effective evidence quantity is used to characterize the sufficiency of evidence for the calculation and gating of evidence sufficiency measurement.

8. The YOLO detection method for road defects based on lightweight CNN according to claim 1, characterized in that: The confidence stability measure is determined based on the degree of fluctuation of the confidence score; the category consistency measure is determined based on the degree of support of the disease category identifier in the observation set; the spatial compactness measure is determined based on the spatial dispersion of the spatial location field; and the morphological fidelity measure is determined based on the consistency of changes in the event size representation or representative size. Furthermore, the road segment attribute information includes at least the road grade, traffic flow grade, and the type of surrounding sensitive points, and the maintenance resource status information includes at least the number of available maintenance teams, the team coverage area, the current pending workload, and the estimated arrival time delay.

9. A road defect YOLO detection system based on lightweight CNN, characterized in that, The detection system is used to implement the method according to any one of claims 1-8, and includes the following modules: The detection and documentation module is used to acquire road surface image frames and input them into a YOLO detection model based on a lightweight CNN, and output a set of disease detection results. Based on the set of disease detection results, a sequence of disease observation records is generated. The disease observation records include at least a timestamp, spatial location field, disease category identifier, confidence score, event size representation quantity, and observation weight, and an index association is established between the disease observation records and the original image or evidence cropped image. The aggregation and evidence collection module is used to perform spatial association and aggregation based on the disease observation record sequence, construct event-level disease objects and update their representative locations and representative dimensions, and generate a comprehensive quantity of event evidence consistency and credibility based on the consistency of multiple observations. Consistency across multiple observations includes one or more of the following: confidence stability measure, class consistency measure, spatial compactness measure, morphological fidelity measure, and evidence sufficiency measure. The priority order generation module is used to calculate the urgency and resource benefit based on the event-level disease object, integrate event severity, evolution trend, road segment attribute information and maintenance resource status information, and use the event evidence consistency and credibility comprehensive value as a gate item and resource cost as a penalty item, and output the work order generation request package accordingly; wherein the resource cost is determined based on the maintenance resource status information. The deduplication closed-loop backfeed module is used to perform deduplication dispatch and status closed-loop backfeed based on work order status and disposal receipt. For event-level disease objects that have been bound to work orders and are in the pending or disposal state, new work orders are not generated repeatedly, but their observation set, event evidence consistency and credibility comprehensive quantity, urgency and resource benefit comprehensive quantity are updated, and the event status is updated according to the disposal receipt.

Images