A method for rapid detection of highway wave-shaped guardrails and related equipment

Abstract

The application provides a highway wave-shaped guardrail rapid detection method and related equipment. The method comprises the following steps: synchronously collecting laser profile, image and position mileage data of the guardrail; extracting guardrail profile features and comparing the guardrail profile features with preset reference features to generate candidate abnormal events; performing spatial coincidence degree verification through constructing continuous guardrail sections, and combining image visibility verification to confirm the abnormal events; and finally calculating confidence degree classification output detection records. Through the multi-source data synchronous collection and double verification mechanism, the method reduces the false positive rate and realizes the automatic and accurate detection of the guardrail abnormalities.

Description

Technical Field

[0001] This application relates to the fields of intelligent transportation and road maintenance technology, and in particular to a rapid detection method and related equipment for highway corrugated guardrails. Background Technology

[0002] Highway guardrails are important traffic safety facilities, and regular inspection of their structural integrity is crucial to ensuring driving safety. Some related technologies often use vehicle-mounted mobile inspection equipment, equipped with LiDAR, to continuously scan the guardrails, extract their three-dimensional contour features, and compare them spatially with standard contours to determine whether there are any abnormalities such as deformation, damage, or missing parts.

[0003] However, in real-world open road environments, mobile detection equipment inevitably encounters random point cloud noise caused by road bumps during operation. Simultaneously, the road environment often contains foreground objects such as passing vehicles, pedestrians, construction cones, or protruding vegetation. When these foreground objects briefly obscure the guardrail, the local point cloud scanned by the lidar may experience data loss or morphological distortion. Since related technologies typically rely solely on single-frame or local spatial contour deviations for anomaly detection, the system struggles to effectively distinguish between these "false anomalies" caused by temporary obstruction or random noise and genuine physical damage to the guardrail. This easily leads the system to misjudge numerous contour changes caused by environmental interference as guardrail anomalies, resulting in an extremely high false alarm rate. This high false alarm rate makes the detection system's output unreliable, requiring significant manual verification and removal of false records, thus reducing the overall efficiency of guardrail anomaly detection. Summary of the Invention

[0004] This application provides a rapid detection method and related equipment for highway corrugated guardrails, which can improve the overall operational efficiency of guardrail anomaly detection.

[0005] Firstly, this application provides a rapid detection method for highway wave-shaped guardrails. The method includes: simultaneously acquiring a sequence of laser contour frames, a sequence of image frames, and a sequence of location mileage for the guardrail using a unified time reference while a mobile detection device travels along the road, and recording the corresponding timestamps; extracting point cloud data within a preset distance and height range from the laser contour frame sequence to generate corresponding guardrail contour features; calculating the spatial deviation value between the guardrail contour features and a preset guardrail reference feature; if the spatial deviation value is greater than a preset deviation threshold, marking the corresponding guardrail contour feature as a candidate abnormal event; extracting the timestamps and location mileage data corresponding to the candidate abnormal events; and stitching together the laser contour frames corresponding to the timestamps and a preset number of adjacent laser contour frames along the travel direction of the mobile detection device. Construct a continuous guardrail segment; within the continuous guardrail segment, calculate the spatial overlap of candidate anomalies in adjacent laser contour frames. If the spatial overlap is greater than a preset overlap threshold, the candidate anomaly is confirmed to have passed the continuity check; retrieve the image frames corresponding to the timestamps of the candidate anomalies that have passed the continuity check, and identify the occlusion pixel areas and the visible pixel areas of the guardrail in the image frames; project the spatial coordinates of the candidate anomalies onto the image frames. If the projection point is located within the visible pixel areas of the guardrail, the candidate anomaly is confirmed to have passed the visibility check; extract the spatial deviation values ​​of the candidate anomalies that have passed the visibility check, and calculate the confidence level of the candidate anomalies; bind the confidence level, candidate anomalies, corresponding location mileage data, and image frames together, and output the guardrail anomaly detection record.

[0006] In the above embodiments, the system overcomes the limitations of single sensors being susceptible to interference in complex environments by constructing a multi-source data synchronous acquisition mechanism for laser contours, images, and location mileage. After extracting the guardrail contour and initially filtering out spatial deviations, the system introduces spatial continuity verification based on adjacent multi-frame point clouds and visibility verification based on image semantics. This multi-stage cross-validation combines the three-dimensional spatial perception of the LiDAR with the two-dimensional semantic recognition capabilities of the camera, eliminating false alarms caused by random noise or foreground occlusion within continuous segments. Finally, by calculating confidence levels based on comprehensive spatial deviations, a quantitative assessment of abnormal events is achieved, improving the reliability of the detection results.

[0007] In conjunction with some embodiments of the first aspect, in some embodiments, the extraction of point cloud data within a preset distance and height range from the laser contour frame sequence to generate corresponding guardrail contour features specifically includes: obtaining the installation height of the mobile detection device; obtaining the original distance data of each laser point in the laser contour frame sequence; removing laser points whose original distance data is less than the minimum detection distance or greater than the maximum detection distance to obtain a first filtered point cloud; calculating the actual ground height corresponding to each laser point based on the vertical offset of the device installation height and each laser point in the first filtered point cloud; comparing the actual ground height with the preset guardrail standard height; extracting laser points whose actual ground height is within the preset error allowable range of the guardrail standard height to obtain a second filtered point cloud; and extracting a three-dimensional contour model of the guardrail based on the second filtered point cloud, and extracting the waveform vertices and bottom fixed points of the guardrail from the three-dimensional contour model as guardrail contour features.

[0008] In the above embodiments, the system transforms the point cloud into the actual ground height by introducing the geometric relationship between the device installation height and the vertical offset. Using the standard height of the guardrail as a spatial constraint, the system can isolate the ground from background noise and obtain a clean 3D contour model. By directionally extracting the waveform vertices and bottom fixed points as feature references, the data processing dimensionality of spatial comparison is reduced, and the anti-interference capability of feature extraction is improved.

[0009] In conjunction with some embodiments of the first aspect, in some embodiments, the calculation of the spatial deviation value between the guardrail outline feature and the preset guardrail reference feature, and if the spatial deviation value is greater than the preset deviation threshold, then the corresponding guardrail outline feature is marked as a candidate abnormal event, specifically includes: obtaining the reference coordinates of the standard waveform vertex and the reference coordinates of the standard bottom fixed point in the preset guardrail reference feature; calculating the first spatial distance between the waveform vertex and the standard waveform vertex, and the second spatial distance between the bottom fixed point and the standard bottom fixed point, respectively, to obtain the spatial deviation value; when the first spatial distance is greater than the first preset threshold and the second spatial distance is less than the second preset threshold, it is determined that the guardrail has a deformation or offset deviation, and the corresponding guardrail outline feature is marked as a candidate abnormal event; when the first spatial distance or the second spatial distance cannot be calculated and the density of the second filtered point cloud is lower than the preset density threshold, it is determined that the guardrail has a damage or missing deviation, and the corresponding guardrail outline feature is marked as a candidate abnormal event.

[0010] In the above embodiments, the system constructs a multi-dimensional morphological deviation assessment model by calculating the spatial distance between the waveform's apex and the bottom fixed point. When the top is offset while the bottom is fixed, the system determines it as deformation or offset; when the feature point distance is missing and the point cloud density is below a threshold, the system uses density features to determine damage or missing points. This joint determination mechanism based on feature distance and point cloud density refines abnormal alarms into specific physical damage types, improving the fine-grainedness of abnormal event classification.

[0011] In conjunction with some embodiments of the first aspect, in some embodiments, the mobile detection device, while traveling along the road, synchronously acquires a sequence of laser contour frames, a sequence of image frames, and a sequence of location mileage from the guardrail using a unified time reference, and records the corresponding timestamps. Specifically, this includes: receiving positioning data in real time at a preset frequency through a GNSS positioning module to generate a sequence of location mileage containing latitude, longitude, and elevation; acquiring the wheel circumference and odometer component parameters of the mobile detection device, and calculating the fixed-distance exposure trigger timing in real time based on a preset interval exposure distance; when the travel distance of the mobile detection device reaches the fixed-distance exposure trigger timing, sending an exposure command to the camera to acquire a sequence of image frames; and controlling the lidar to acquire a laser contour frame within the same trigger cycle of sending the exposure command, and binding the sequence of location mileage, the image frame, and the laser contour frame to the same timestamp.

[0012] In the above embodiments, to address the issue of spatiotemporal misalignment in data during motion detection, the system calculates the timing for fixed-distance exposure in real time by combining wheel circumference and odometer parameters. When the trigger timing is reached, the system synchronously triggers camera exposure and LiDAR acquisition within the same cycle, binding multi-source data and location mileage to the same timestamp. This mechanism establishes fixed-distance synchronization logic at the underlying level, avoiding uneven data density caused by vehicle speed fluctuations and ensuring spatiotemporal consistency in heterogeneous data fusion.

[0013] In conjunction with some embodiments of the first aspect, in some embodiments, after the steps of synchronously acquiring the laser contour frame sequence, image frame sequence, and location mileage sequence of the guardrail at a unified time reference and recording the corresponding timestamps during the mobile detection device's travel along the road, the method further includes: acquiring initial travel mileage data of the mobile detection device during actual road travel; extracting the location mileage sequence output by the GNSS positioning module and calculating the actual spatial straight-line distance between adjacent positioning points; accumulating the actual spatial straight-line distance within a preset time window to obtain reference mileage data; calculating the mileage error between the initial travel mileage data and the reference mileage data; if the mileage error exceeds a preset calibration threshold, then using the reference mileage data to perform mileage calibration and update the location mileage sequence.

[0014] In the above embodiments, to address the limitation of odometry being prone to cumulative drift, the system extracts the spatial straight-line distance between GNSS positioning points and accumulates it within a time window to generate a reference mileage. When the error between the initial and reference mileage exceeds a threshold, the position sequence is updated using the reference mileage. This mechanism introduces a global positioning benchmark to constrain the measurement error of the local odometry, suppresses position shifts detected over long distances, and improves the reliability of spatial positioning.

[0015] In some embodiments of the first aspect, the spatial coordinates of the candidate abnormal event are projected onto an image frame. If the projection point is located within the visible pixel area of ​​the guardrail, the candidate abnormal event is confirmed to have passed the visibility check. Specifically, this includes: obtaining the external parameter matrix between the camera acquiring the image frame and the lidar acquiring the laser contour frame sequence, as well as the camera's internal parameter matrix; using the external parameter matrix to transform the spatial coordinates of the candidate abnormal event from the lidar coordinate system to the camera coordinate system to obtain the three-dimensional coordinates in the camera coordinate system; using the internal parameter matrix to map the three-dimensional coordinates in the camera coordinate system to a two-dimensional image plane to obtain the projection point coordinates; determining whether the projection point coordinates fall within the boundary range of the visible pixel area of ​​the guardrail; if they fall within the boundary range and the projection point coordinates do not overlap with the pixel area of ​​the obstruction, the candidate abnormal event is confirmed to have passed the visibility check.

[0016] In the above embodiments, the system maps the three-dimensional coordinates of candidate anomaly events to a two-dimensional image plane using intrinsic and extrinsic parameter matrices, establishing a cross-modal spatial association. After acquiring the projection points, the system performs topological relationship determination between them and the visible areas of the guardrail and the occlusion areas in the image. This mechanism utilizes image semantic information to compensate for the shortcomings of point clouds in identifying foreground occlusion, and by eliminating projection points that overlap with occlusions, it reduces the false anomaly detection rate caused by interference from non-guardrail objects.

[0017] In conjunction with some embodiments of the first aspect, in some embodiments, the extraction of spatial deviation values ​​of candidate abnormal events that have passed visibility verification and the calculation of confidence levels for candidate abnormal events specifically include: obtaining the spatial overlap degree corresponding to the candidate abnormal event and the pixel distance from the projection point to the edge of the pixel region of the occluder; calculating a basic abnormal score based on the product of the spatial deviation value and a preset proportional coefficient; weighting the basic abnormal score using the spatial overlap degree and correcting the weighted basic abnormal score using the pixel distance to obtain a comprehensive confidence score; and matching the comprehensive confidence score with multiple preset confidence intervals to determine the confidence level of the candidate abnormal event.

[0018] In the above embodiments, the system integrates spatial deviation value, spatial overlap degree, and pixel distance from the projection point to the occlusion edge to construct a confidence scoring model. The system calculates a base score based on spatial deviation, uses multi-frame continuity for weighting, and corrects according to the visual occlusion critical distance, expanding a single geometric index into a composite evaluation system that includes spatiotemporal continuity and visual reliability. This achieves quantitative grading of detection results and reduces the screening cost of manual review.

[0019] In a second aspect, embodiments of this application provide a rapid detection device for highway corrugated guardrails. The rapid detection device for highway corrugated guardrails includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the rapid detection device for highway corrugated guardrails to perform the method described in the first aspect and any possible implementation thereof.

[0020] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on a highway corrugated guardrail rapid detection device, cause the highway corrugated guardrail rapid detection device to perform the method described in the first aspect and any possible implementation thereof.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on a highway corrugated guardrail rapid detection device, cause the highway corrugated guardrail rapid detection device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Understandably, the highway corrugated guardrail rapid detection device provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0023] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0024] 1. This application constructs a multi-source data synchronous acquisition mechanism for laser contours, images, and location mileage, combined with fixed-distance exposure triggering and mileage calibration, to avoid uneven data density and long-distance position shift caused by vehicle speed fluctuations, thus ensuring the spatiotemporal consistency of heterogeneous data fusion and the reliability of spatial positioning.

[0025] 2. This application introduces multi-stage cross-validation based on spatial continuity verification of adjacent multi-frame point clouds and visibility verification based on image semantics, which makes up for the limitations of single sensors being susceptible to interference, reduces false alarms caused by random noise or foreground occlusion by non-guardrail objects, and lowers the false anomaly detection rate.

[0026] 3. This application constructs a comprehensive confidence scoring model by integrating spatial deviation value, spatial overlap degree and visual occlusion distance to achieve quantitative classification of detection results and refine abnormal alarms into specific physical damage types. This not only improves the fineness of abnormal event classification, but also reduces the screening cost of manual review. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating a rapid detection method for highway corrugated guardrails in an embodiment of this application;

[0028] Figure 2 This is a schematic diagram of the physical device structure of a rapid detection device for highway corrugated guardrails in the embodiments of this application. Detailed Implementation

[0029] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0030] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0031] For ease of understanding, the method provided in this implementation is described in process below. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating a rapid detection method for highway corrugated guardrails in an embodiment of this application.

[0032] S101. During the movement of the mobile detection equipment along the road, the laser contour frame sequence, image frame sequence, and location mileage sequence of the guardrail are synchronously collected using a unified time reference, and the corresponding timestamps are recorded.

[0033] The unified time reference refers to the sharing of the same clock reference among all sensors and positioning modules within the system. This ensures that each acquisition event is assigned a comparable absolute time marker, guaranteeing a consistent spatiotemporal reference coordinate system for the laser contour frames, image frames, and location mileage sequences during subsequent data fusion. The laser contour frame sequence refers to the set of point cloud data output frame by frame by the lidar at a fixed or event-triggered frequency. Each frame contains the original distance values, reflected signal strength, and corresponding angle information for each ranging point within the current cross-section. The image frame sequence refers to the ordered set of images continuously acquired by the camera according to preset exposure rules. The location mileage sequence refers to the geographic location record stream output by the GNSS positioning module at a fixed frequency, containing latitude, longitude, elevation, and positioning quality indicators. The timestamp is an absolute time marker bound to each acquisition event, and its format can be Unix epoch time (accuracy not less than 1ms) or GNSS UTC time (accuracy up to sub-millisecond levels).

[0034] This step continues to execute after the mobile detection equipment starts data acquisition and serves as the data foundation for all subsequent analysis steps. Specifically, the establishment of a unified time reference supports the following three implementation methods: First, hardware-triggered synchronization, which uses the pulses per second (PPS) signal output by the GNSS module as the clock reference. The main control board sends trigger commands to the lidar and camera synchronously when the rising edge of the PPS arrives. The acquisition events of each sensor are aligned with this pulse moment, and the time synchronization error can be controlled within 5ms. This is suitable for high-precision detection scenarios equipped with a PPS hardware interface. Second, main control board software timestamp correction synchronization, where the main control board system clock uniformly allocates time tags, records the timestamp at the same time as sending the acquisition command, and aligns the time axis of multiple data streams according to the trigger log. This is suitable for low-cost equipment configurations that do not have a PPS interface. Third, post-processing registration synchronization, which performs nearest neighbor matching or linear interpolation alignment based on the timestamps and mileage information of each data stream after acquisition is completed, serving as a supplement or degradation scheme to the first two methods. The three methods described above cover different implementation paths from the hardware layer to the software layer, and from online processing to offline post-processing, so that this solution is not limited by a specific synchronous implementation.

[0035] Regarding vehicle speed management, this embodiment supports a detection speed range of 0 to 90 km / h. If a fixed time interval triggering method is used, when the vehicle speed increases from 60 km / h to 90 km / h, the road surface coverage spacing between adjacent frames will increase from approximately 1.7 m to approximately 2.5 m, posing a risk of missing small abnormal sections. To address this, this embodiment introduces a fixed-distance exposure triggering mechanism in a subsequent preferred embodiment, using travel distance rather than time as the triggering basis, fundamentally decoupling the correlation between sampling density and vehicle speed. Furthermore, when a suspected abnormal section is detected during the subsequent continuous verification phase, the system can temporarily narrow the trigger interval to implement encrypted sampling for that section, obtaining denser evidence frames for high-risk sections without affecting overall detection efficiency, thus balancing the collection efficiency of regular road sections with the detection accuracy of abnormal areas.

[0036] In some preferred embodiments, the above steps specifically include: receiving positioning data in real time at a preset frequency through a GNSS positioning module of a global navigation satellite system, generating a sequence of location mileage containing latitude, longitude and elevation; obtaining the wheel circumference and odometer component parameters of the mobile detection device, and calculating the fixed-distance exposure triggering time in real time in combination with a preset interval exposure distance;

[0037] When the travel distance of the mobile detection device reaches the fixed-distance exposure trigger time, it sends an exposure command to the camera to acquire a sequence of image frames; within the same trigger cycle of sending the exposure command, it controls the lidar to acquire a laser contour frame, and binds the position mileage sequence, image frame and laser contour frame to the same timestamp.

[0038] The preset frequency refers to the time frequency at which the GNSS positioning module continuously outputs positioning frames. In this embodiment, 10Hz is preferred, meaning that one frame containing fields such as latitude, longitude, elevation, number of satellites, and differential status is output every 100ms. This frequency can be adjusted within the range of 5Hz to 20Hz depending on the positioning board model. A higher output frequency helps improve the spatial resolution of the location mileage sequence and reduces the cumulative error between adjacent positioning points. Wheel circumference refers to the theoretical rolling circumference of the vehicle's drive wheels, which needs to be calibrated by round-trip measurements at a known reference distance before operation. A typical value is 1.8m to 2.2m. Odometer component parameters refer to the proportional coefficient that converts the odometer encoder pulse count into linear displacement, which needs to be determined by joint calibration of the wheel circumference and encoder resolution. Interval exposure distance refers to the actual road distance that the detection device should travel between two adjacent exposure triggers. It can be set within the range of 0.1m to 5m, with the preferred value determined by a combination of target detection resolution and the minimum lateral unfolding length of guardrail anomalies.

[0039] This step operates in a real-time loop throughout the data acquisition process. Specifically, the GNSS positioning module continuously outputs a positioning sequence at a frequency of 10Hz; simultaneously, the main control board reads the odometer pulses in real time and calculates the cumulative driving distance since the last trigger based on calibration parameters; when this distance reaches the preset interval exposure distance, the main control board synchronously sends an exposure command to the camera within the same trigger cycle and controls the LiDAR to acquire a frame of point cloud data. The data from both sensors are bound to the same timestamp and then associated with and stored with the corresponding GNSS positioning record.

[0040] The interval exposure distance can be derived from the target detection resolution requirements. Let the minimum lateral extension length of interest for guardrail anomalies be... (Unit: m), the overlap rate of adjacent image frames along the driving direction must be no less than [value missing]. The upper limit of the exposure distance. satisfy:

[0041]

[0042] For example, when the minimum length of the anomaly of interest is 0.5m and the inter-frame overlap rate is required to be no less than 50%, the upper limit of the exposure distance is 0.25m; when storage resources are sufficient, 0.25m to 0.5m can be selected, and when storage is limited, it can be appropriately relaxed to within 1m. The calibration steps for the wheel circumference and odometer component parameters are as follows: select a known distance reference segment of no less than 50m on a flat road surface, drive once in the forward direction and once in the reverse direction, record the total number of odometer pulses respectively, divide the average number of pulses of the two trips by the length of the reference segment to obtain the number of pulses per meter, and take the average of the two trips as the calibration result. The calibration error should not be greater than 0.3%.

[0043] Compared to timed exposure, distanced exposure ensures uniformity of sampling intervals along the driving direction from the ground up, eliminating the interference of vehicle speed fluctuations on the detection coverage density. This is particularly crucial for preventing missed detections of guardrail anomalies in highway scenarios. In some implementations, a LiDAR-triggered, camera-following-triggered mode can also be used, where a trigger signal is sent to the camera based on the LiDAR sampling period. The two implementations differ in the triggering entity, but both can ensure the spatiotemporal correspondence between image frames and point cloud frames at the same driving mileage position, which is an equivalent implementation of the distanced synchronization concept in this step.

[0044] In some preferred embodiments, after step S101 above, the method further includes:

[0045] Acquire initial mileage data of the mobile detection device during actual road travel; extract the sequence of position mileage output by the GNSS positioning module and calculate the actual straight-line distance between adjacent positioning points; accumulate the actual straight-line distance within a preset time window to obtain reference mileage data; calculate the mileage error between the initial mileage data and the reference mileage data; if the mileage error exceeds a preset calibration threshold, use the reference mileage data to perform mileage calibration and update the position mileage sequence.

[0046] The initial mileage data refers to the distance value independently accumulated by the wheel encoder or OBD interface of the testing equipment during actual driving, in meters, and continuously updated according to the odometer sampling frequency. This value is affected by factors such as tire wear, road slippage, and encoder cumulative error, and there is a certain degree of drift accumulation during long-distance driving. The reference mileage data refers to the estimated mileage value obtained by accumulating the spatial straight-line distances of adjacent positioning points in the GNSS positioning sequence segment by segment. It is based on the global positioning benchmark and is not affected by the vehicle's mechanical state, but is constrained by GNSS positioning accuracy and signal continuity. The preset time window refers to the length of the time interval for mileage accumulation, which can be set within the range of 5s to 60s, with 10s being preferred. If the window is too small, the reference mileage will jump frequently due to GNSS positioning noise; if the window is too large, the calibration response will be lagging and the error accumulation will increase. The preset calibration threshold refers to the maximum allowable deviation between the initial mileage and the reference mileage. It can be determined by the statistical distribution of mileage error in historical road test data (such as taking the mean plus three times the standard deviation), or it can be set according to engineering experience to a cumulative error of no more than 0.5m per 100m.

[0047] This step can be performed in real-time in a rolling manner during data acquisition, or it can be performed in batches to correct the entire location and mileage sequence after acquisition. Specifically, the system extracts the latitude and longitude coordinates of two adjacent positioning points in the GNSS positioning sequence and calculates the actual straight-line distance between the two points using the Haversine formula. within the preset time window The reference mileage is obtained by accumulating the segments within the range. :

[0048]

[0049] Then the initial mileage was calculated. and The difference, if If the preset calibration threshold is exceeded, the mileage field in the location mileage sequence within the corresponding time period is updated with the reference mileage to complete one rolling calibration. Thereafter, the updated value is used as the starting point for the next window.

[0050] In special circumstances where GNSS signal degradation occurs (such as fewer than four satellites, differential status anomalies, or loss of lock-up upon entering a tunnel), the system switches to a short-term odometer maintenance mode dominated by wheel speedometers or IMUs. This mode uses encoder pulses to continuously accumulate travel distance, ensuring the position sequence remains continuous during the lock-up period. Once effective satellite signals are restored, the odometer readings of the lost-lock section are corrected using the re-locked GNSS positioning anchor points, eliminating short-term drift accumulation. In some implementations, known control points in the road network (such as road mileage markers, bridge expansion joints, or tollbooth gantries) can be used to perform segmented forced alignment of the odometer sequence, further eliminating systematic offsets in long-distance detection by introducing absolute position anchoring. This mechanism organically integrates the high-frequency continuous output of the global GNSS reference with that of the local odometer, suppressing the accumulation of position errors in large-scale detection scenarios and providing reliable assurance for accurate geolocation of subsequent abnormal events.

[0051] S102. Extract point cloud data within a preset distance and height range from the laser contour frame sequence to generate the corresponding guardrail contour features.

[0052] The preset distance and height range refers to the joint constraint interval for spatial effectiveness screening of the original laser point cloud. It consists of two independent dimensions: distance (from minimum to maximum detection distance) and height (the allowable error band of the standard height of the guardrail). These two dimensions work sequentially to gradually eliminate background point clouds unrelated to the guardrail. Guardrail contour features refer to the structural features extracted from the screened point cloud that characterize the spatial morphology of the guardrail cross-section. These can be the vertex coordinates and bottom fixed point coordinates of the waveform cross-section, or the envelope boundary line of the cross-sectional curve, peak and valley line parameters, feature points in the column connection area, or a combination of the above features. Different feature forms are suitable for different point cloud density conditions and guardrail types.

[0053] This step is executed in real time when each frame of laser contour data arrives, providing standardized input for the spatial deviation determination of S103. Specifically, the point cloud screening follows a two-stage filtering logic of "distance first, then height": In the first stage, points whose distance values ​​exceed the effective range are removed based on the distance parameter, eliminating equipment self-occlusion, near-distance reflections from the road surface, and far-end background noise; in the second stage, the actual ground height of each point is calculated using the vertical offset between the equipment installation height and the laser point, retaining points whose height falls within the allowable error zone of the guardrail standard height, and eliminating low-level reflection points from the road surface and interference from high-flying birds, vegetation, etc.

[0054] In terms of parameter configuration modes, this embodiment supports both fixed parameter mode and automatic parameter loading mode. In fixed parameter mode, operators manually input distance and height parameters through the settings interface, suitable for routine operations where the guardrail type is known and the site environment is stable. In automatic parameter loading mode, the system automatically calls the corresponding parameter group from the preset parameter library based on the currently identified guardrail type (e.g., the standard height of a double-wave guardrail is approximately 1.1m, and the standard height of a triple-wave guardrail is approximately 1.2m), achieving parameter switching without manual intervention, suitable for scenarios where the guardrail type changes multiple times within the same detection stroke.

[0055] Furthermore, the criteria for selecting effective points are not limited to distance and height. Laser reflection intensity, semantic segmentation prior results, and back projection constraints of the guardrail region of interest in the camera image can all be included in the selection logic. This can further improve the accuracy of guardrail point cloud localization in complex scenarios such as large vehicle occlusion and green vegetation coverage, and achieve bidirectional guidance between image semantics and point cloud geometry.

[0056] In some preferred embodiments, the above steps specifically include: obtaining the installation height of the mobile detection device, obtaining the original distance data of each laser point in the laser contour frame sequence, removing laser points whose original distance data is less than the minimum detection distance or greater than the maximum detection distance, and obtaining a first filtered point cloud;

[0057] Based on the installation height of the equipment and the vertical offset of each laser point in the first selection point cloud, the actual ground height corresponding to each laser point is calculated; the actual ground height is compared with the preset standard height of the guardrail, and laser points whose actual ground height is within the preset error allowable range of the standard height of the guardrail are extracted to obtain the second selection point cloud; based on the second selection point cloud, the three-dimensional contour model of the guardrail is extracted, and the waveform vertices and bottom fixed points of the guardrail are extracted from the three-dimensional contour model as the guardrail contour features.

[0058] The equipment installation height refers to the actual vertical distance from the center of the lidar scanning unit to the road surface. This must be measured and recorded on-site before operation, with a typical value of 1.2m to 1.5m, varying depending on the installation location (roof platform, side bracket, or dedicated cantilever). The minimum and maximum detection distances are the lower and upper bounds for distance filtering, respectively, and can be manually set within a range of 1m to 10m (the minimum value must not exceed the maximum). The minimum detection distance is typically configured to 1m to 3m to eliminate near-field reflections from the equipment itself and the road surface, while the maximum detection distance is typically configured to 5m to 10m to exclude background noise from distant buildings and vegetation on both sides of the road. The standard guardrail height is the design installation height for the target guardrail type; for double-wave guardrails, it can be set to 1.1m, and for triple-wave guardrails, it can be set to 1.2m. The preset error allowable range is the tolerance band for height filtering, with a default value of ±5cm. It can be manually adjusted within a range of ±3cm to ±8cm, and this range is usually determined by the statistical error distribution of the height of the on-site calibration samples.

[0059] This step sequentially completes four stages: distance filtering, height filtering, point cloud filtering, and feature extraction. Specifically, height calculation is based on the following geometric relationship: Let the first... The vertical offset of each laser point relative to the scanning unit is (Positive for downward), the equipment installation height is: Then the actual ground height at that point is:

[0060]

[0061] High-level filtering criteria are ,in This is the standard height for guardrails. This is within the allowable error range. When the number of valid points filtered by height in a single frame is lower than the preset minimum number of valid points threshold (which can be set to 5 to 10 points, with the specific value determined by the angular resolution of the LiDAR and the horizontal unfolding width of the guardrail), the frame is marked as a sparse point cloud frame and enters the subsequent damage or missing deviation judgment branch, instead of being directly discarded, in order to avoid misjudging real missing points as invalid frames due to temporary reasons such as occlusion.

[0062] For the second-selection point cloud after double screening, the system first uses statistical outlier removal or a neighborhood radius-based filtering algorithm to remove outlier points, and then extracts the guardrail contour features. Feature extraction supports two implementation methods: one is the "vertex-bottom point" extraction method, which identifies the highest stable point group reflecting the waveform peak in the cross-section as the waveform vertex and the lowest stable point group in the bottom installation area as the bottom fixed point, recording their three-dimensional coordinates respectively. This method is suitable for conventional working conditions with clear waveform contours and sufficient point cloud density. The other is the cross-sectional curve fitting method, which performs polynomial or spline curve fitting on the selected point cloud and extracts the peak height of the fitted curve. Valley ,wavelength Equal peak and valley line parameters and centerline offset are used as structured features, suitable for conditions where point clouds are sparse or where the installation has a systematic tilt. The form of the 3D contour model is not limited to a single-frame cross-sectional contour; it can also be a local 3D segment stitched together from multiple frames or a set of fitted structured parameters to adapt to the feature accuracy requirements of different subsequent anomaly judgments.

[0063] Furthermore, the equidistant distribution of guardrail posts along the roadside (standard post spacing is typically 2m or 4m) can be used to verify the structural consistency of feature extraction results: when the spacing between adjacent feature regions is consistent with the prior knowledge of the posts, the confidence level of the current contour recognition can be improved, and the probability of feature mislocalization caused by single-frame point cloud distortion can be reduced.

[0064] S103. Calculate the spatial deviation value between the guardrail outline feature and the preset guardrail reference feature. If the spatial deviation value is greater than the preset deviation threshold, mark the corresponding guardrail outline feature as a candidate abnormal event.

[0065] The preset guardrail reference features refer to the benchmark templates used for spatial comparison with the current guardrail outline features. These templates come from at least two sources: first, standard design size templates, which are reference coordinate sets pre-established based on the design specifications (wave height, wave pitch, installation height, etc.) of double-wave and triple-wave guardrails in national or industry standards. These are suitable for road sections where the testing environment is compliant and the guardrail installation meets standards. Second, dynamic historical sample templates, which are sliding baselines generated by statistically averaging the guardrail features of adjacent normal frames or historical normal road sections that have passed quality assessment during the same acquisition process. These templates can automatically adapt to changes in road slope or systematic installation deviations, and are suitable for complex conditions where there are segmental overall offsets in guardrail installation. The two types of templates can also be used in combination, with the standard template as the primary method and historical sample correction as a supplement.

[0066] Spatial deviation value refers to the quantified difference between the current frame guardrail contour features and the reference features. It can be the Euclidean distance between a single feature point, or a weighted comprehensive score of the distances of multiple feature points, or the Hausdorff distance. In this embodiment, the first spatial distance of the waveform vertices is used. Second spatial distance from the bottom fixed point As the main deviation component, candidate anomaly events refer to guardrail frame records that are marked as suspected anomalies at this stage but have not yet passed the dual verification of continuity and visibility. The attribute fields include trigger frame timestamp, location mileage, initial anomaly type label (deformation, offset, damage or missing) and spatial deviation value. Their status is "pending review" rather than "confirmed".

[0067] This step, executed frame-by-frame after the guardrail contour feature extraction is completed in S102, is the core node for decoupling the initial anomaly screening from subsequent multi-stage verification. Specifically, the system calculates the first spatial distance between the current frame waveform vertex and the reference waveform vertex. and the second spatial distance between the bottom fixing point and the reference bottom fixing point The exception type is determined according to the following two paths: when (First preset threshold) and When the second preset threshold is reached, it indicates that the top of the guardrail has deviated from the reference position while the bottom installation area is basically stable, which is judged as a deformation or offset anomaly; when or If the calculation cannot be performed due to insufficient valid points (below the minimum valid point threshold), and the single-frame point density of the second filter point cloud is lower than the preset density threshold, it indicates that the current area guardrail structure data is missing, and is judged as a damaged or missing anomaly.

[0068] The method for determining the preset threshold may include: collecting statistical deviation distribution data from normal road section samples, taking the mean and adding a certain multiple of the standard deviation (e.g., ...). Using this as the judgment boundary, this method can incorporate the combined impact of actual installation errors and acquisition noise into the threshold setting; or directly refer to the allowable deviation range of guardrails specified in road maintenance specifications (such as the horizontal displacement of the top of the guardrail not exceeding 50mm) as the threshold to ensure that the test results are aligned with the engineering acceptance standards. Preset deviation thresholds can also be set according to guardrail type (dual waveform / triple waveform) or road grade (expressway / national highway / provincial highway) to adapt to different detection accuracy requirements. The point cloud density threshold can be defined as the lower limit of the number of effective points in a single frame. Its specific value needs to be determined comprehensively based on the lidar sampling frequency, vehicle speed, and the minimum length of the missing section of concern for the guardrail: for example, when the lidar sampling frequency is 50Hz and the vehicle speed is 90km / h, the distance between adjacent frames is approximately 0.5m. If the minimum length of the missing section of concern is 0.5m, then the density threshold should ensure that at least one frame of data has a sufficient number of effective points, and the minimum number of effective points required per frame can be calculated accordingly.

[0069] It should be noted that candidate anomalies in this step only enter the verification queue and are not directly output as final anomaly records. Their validity still needs to be further confirmed by the continuity check in S105 and the visibility check in S107. This naturally filters out false alarms caused by occasional noise or occlusion in a single frame before the end of the process. To support subsequent cross-frame event merging and lifecycle tracking, each candidate anomaly can be assigned a unique spatiotemporal tracking identifier when it is generated in this step, recording the location of the first trigger frame, the mileage of the trigger frame, and the initial judgment result of the anomaly type. This provides an event-level index basis for the construction of continuous guardrail sections in S104 and the spatial overlap calculation in S105.

[0070] In some preferred embodiments, the above steps specifically include:

[0071] Obtain the reference coordinates of the standard waveform vertex and the standard bottom fixed point in the preset guardrail reference features; calculate the first spatial distance between the waveform vertex and the standard waveform vertex, and the second spatial distance between the bottom fixed point and the standard bottom fixed point, to obtain the spatial deviation value; when the first spatial distance is greater than the first preset threshold and the second spatial distance is less than the second preset threshold, it is determined that the guardrail has deformation or offset deviation, and the corresponding guardrail contour feature is marked as a candidate abnormal event;

[0072] When the first spatial distance or the second spatial distance cannot be calculated and the density of the second selected point cloud is lower than the preset density threshold, it is determined that the guardrail has a damaged or missing deviation, and the corresponding guardrail contour feature is marked as a candidate abnormal event.

[0073] Among them, the first spatial distance This refers to the three-dimensional Euclidean distance between the waveform vertex coordinates extracted in the current frame and the corresponding standard waveform vertex coordinates in the preset reference template, expressed in millimeters, quantifying the spatial offset of the top shape of the guardrail; the first preset threshold. Used for determination Whether it exceeds the normal fluctuation range can be determined by referring to the allowable deviation of the horizontal displacement of the top of the guardrail in the road maintenance specifications (e.g., 50mm), or by taking the deviation distribution of samples from normal road sections. The settings can be determined and preset according to guardrail type (double-wave / triple-wave) and road grade. Second spatial distance. This refers to the three-dimensional Euclidean distance between the coordinates of the bottom fixed point and the reference coordinates of the standard bottom fixed point, reflecting the positional stability of the installation area at the bottom of the guardrail; the second preset threshold. Typically significantly smaller than (Typical values ​​are 15mm to 25mm), ensuring that the deformation / displacement determination path is triggered only when the bottom is fixed and the top is offset, thus distinguishing between the two physical states of local deformation and overall displacement of the guardrail. The preset density threshold refers to the minimum requirement for the number of effective points within a single frame after double screening, which can be 5 to 10 points. The specific value needs to be determined comprehensively based on the angular resolution of the LiDAR, vehicle speed, and the minimum length of the missing section of the guardrail. For example, when the laser sampling frequency is 50Hz and the vehicle speed is 90km / h, the distance between adjacent frames is approximately 0.5m. If the minimum length of the missing section of the guardrail is 0.5m, then the density threshold should ensure that at least one frame of data contains a sufficient number of effective points, and the minimum number of effective points required per frame can be calculated accordingly.

[0074] This step uses the guardrail outline features extracted in S102 as input and completes candidate anomaly labeling along two parallel paths. Specifically, when and The deformation / offset path is triggered at that time. This combined condition accurately corresponds to the typical physical pattern where the top shape of the guardrail deviates from the normal waveform after an impact, while the bottom mounting base remains stable. It can effectively distinguish between the systematic overall offset of the mounting layer (in which case the top and bottom offset synchronously, and the combined condition is not met) and the deformation of the guardrail body. or When the number of valid points is insufficient for calculation and the point density of a single frame is simultaneously below a preset density threshold, a broken / missing path is triggered. Both conditions must be met simultaneously to exclude false triggers caused by distance calculation failure due to a single feature point being temporarily occluded. The criteria for determining insufficient point cloud density are not limited to the number of valid points. A low proportion of valid points, abnormally reduced laser reflection intensity, or high cross-sectional fitting residuals can all serve as equivalent triggering conditions, ensuring that this path remains applicable to different sources of degradation.

[0075] It should be noted that, since foreground occlusion is highly similar to the appearance of real structural defects at the point cloud level, the damage / missing candidate events marked in this step only enter the verification queue. They must be jointly confirmed by S105 continuity verification (real defects are continuously presented in multiple frames, while occlusion usually only affects a very few frames) and S107 visibility verification (occluders have clear foreground pixels in the image) before they can be finally distinguished. This is also one of the core bases that distinguishes the multi-stage verification mechanism from the single-step anomaly judgment scheme.

[0076] S104. Extract the timestamp and location mileage data corresponding to the candidate abnormal events, and along the driving direction of the mobile detection device, stitch together the laser contour frame corresponding to the timestamp and a preset number of adjacent laser contour frames to construct a continuous guardrail section.

[0077] The continuous guardrail section refers to a local three-guardrail observation segment formed by splicing together several laser contour frames extending forward and backward along the driving direction using candidate anomaly frames as spatial anchor points, after coordinate system transformation. This segment provides a spatially continuous reference for subsequent spatial overlap calculations. The actual coverage length of this section on the road surface is determined by both the frame selection window strategy and the laser sampling parameters. The preset number of adjacent frames refers to the total number of laser contour frames included in the splicing window according to the frame selection strategy; its reasonable value needs to comprehensively consider the lidar sampling frequency. Equipment travel speed With the desired segment coverage length ,satisfy Constraints are imposed to ensure that the selected frames have sufficient spatial coverage density for the target guardrail segment. Stitching refers to the process of transforming multiple frame point clouds from their respective sensor local coordinate systems to a unified segment reference coordinate system, which requires the location, mileage, and attitude information corresponding to each frame as the basis for the transformation.

[0078] This step is triggered immediately after candidate anomaly labeling is completed in S103. Its core task is to expand isolated single-frame anomaly candidates into a set of observations with spatiotemporal context, establishing a spatial comparison basis for continuity verification. Specifically, the selection of adjacent frames supports three window strategies: one is a fixed-frame-number window, taking frames before and after the candidate frame. The first method uses a configurable frame count of 3 to 10 frames, with 5 frames being the preferred count. This method is simple to implement, has low computational cost, and is suitable for typical scenarios where vehicle speed is relatively stable. The second method uses a fixed time window, taking the duration on both sides of the candidate frame timestamp. All frames within the window can be set from 0.1s to 0.5s, which is more stable than a fixed frame window when vehicle speed fluctuates significantly; the third is a fixed mileage window, which extends to both sides based on the mileage of the candidate frame position. All frames within the route can be set from 0.5m to 3m, preferably 1m. This provides the most direct estimation of the physical coverage of the guardrail and is the implementation most closely related to the abnormal physical extension length of the guardrail among the three strategies. The three strategies are equivalent implementations, and the system can choose one or combine them based on the odometer accuracy and vehicle speed range. In some implementations, strict frame-level stitching can be omitted, and local spatial comparison between frames can be performed using a sliding sequence association method, which can also meet the technical objectives of this step. The coordinate system for inter-frame stitching can be rigid body transformation based on pose translation matrix, GNSS / IMU fusion attitude transformation, or ICP / NDT fine-tuning registration when point cloud overlap is sufficient. The accuracy and computational cost increase progressively among the three, and the choice can be made based on deployment resources.

[0079] In terms of window symmetry design, forward frames (frames acquired after candidate frames) and backward frames (frames acquired before candidate frames) can be configured asymmetrically, giving forward frames a larger weight. This is because real structural anomalies usually persist in subsequent frames, while sporadic acquisition noise often disappears quickly in subsequent frames. Asymmetrical windows can concentrate more observation resources on the continuous verification of anomalies without increasing the total number of frames, thereby improving the ability to distinguish between real anomalies and transient noise in the continuous verification stage.

[0080] S105. Within a continuous guardrail section, calculate the spatial overlap of candidate abnormal events in adjacent laser contour frames. If the spatial overlap is greater than a preset overlap threshold, the candidate abnormal event is confirmed to have passed the continuity check.

[0081] Spatial overlap refers to a quantitative indicator of the spatial consistency of candidate anomaly events across multiple adjacent laser profile frames within a continuous guardrail section, reflecting the stability of the same anomaly region on the time axis. Its calculation can be based on the overlap rate of point clouds in adjacent anomaly regions (i.e., the ratio of intersection points to union points), the intersection-to-union ratio (IoU) of the cross-sectional profile, and the overlap of candidate feature points within the laser profile. The percentage of frames triggering anomalies, or the continuous stability of the trajectory of the anomaly center point, have different requirements for point cloud density and computational costs. These can be equivalently replaced by quantification methods such as the number of consecutive occurrences, similarity scores, or lifecycle length, all of which are optional implementations for continuity verification in this step. The preset overlap threshold refers to the minimum spatial overlap value for determining whether the continuity verification passes. It can be determined by the statistical distribution of historical samples of known normal road segments and typical abnormal road segments, and can be configured in the range of 0.4 to 0.8, with 0.6 being preferred. It can also be preset according to guardrail type and road grade.

[0082] This step is executed after the continuous guardrail segment is constructed in S104, and it is the core mechanism for filtering occasional noise in a single frame and suppressing false alarms. Specifically, taking the point cloud overlap rate as an example, let the candidate abnormal event be in the... Frame and the The corresponding abnormal region point sets in the frame are as follows and Then the overlap rate of the frame pair is:

[0083]

[0084] The system calculates the overlap rate of all adjacent frame pairs within a window and takes the average as the total spatial overlap, comparing it with a preset overlap threshold. A typical exclusion example illustrates this: if a candidate anomaly is triggered only in one frame and disappears in the subsequent three frames, its total overlap will be far below the threshold, and it will be judged as a pseudo-anomaly caused by road bumps, bird flight, or single-frame point cloud distortion; conversely, if the anomaly is consistently present in multiple consecutive frames, it passes the continuity check.

[0085] The continuity check also examines the rate of size change of candidate anomaly regions between frames: if the area of ​​anomaly regions gradually and monotonically expands or shrinks between adjacent frames, conforming to the physical laws of real guardrail structural deterioration, the requirement for the absolute value of overlap can be appropriately relaxed; if the size shows drastic random jumps, additional penalties are applied. This joint judgment mechanism enables the continuity check to go beyond simply "whether it continues to appear," further enabling it to distinguish between real gradual deformation and acquisition noise flicker, which is of substantial significance in reducing the burden of manual review in long-distance, high-speed detection scenarios.

[0086] S106. Retrieve the image frames corresponding to the timestamps of the candidate abnormal events that have passed the continuity check, and identify the pixel areas of the occluded objects and the visible pixel areas of the guardrail in the image frames.

[0087] The occlusion pixel region refers to the set of pixels in the image frame that are covered by foreground objects (including passing vehicles, pedestrians, construction cones, green plants, etc.) on the projection position of the guardrail. The occlusion types include complete occlusion, partial occlusion, and dynamic occlusion. Pixel-level precision masking is not required; annotation in the form of target detection boxes or outline approximations is sufficient for the functional requirements of this step. The visible pixel region of the guardrail refers to the set of pixels in the image frame that belong to the guardrail body and are not currently covered by occlusion objects, constituting the valid landing point range for the S107 visibility projection verification.

[0088] This step is performed after the continuity check in S105 passes. Its purpose is to use image semantic information to make up for the inherent deficiency of laser point clouds in occlusion state perception, and to prepare image mask input for subsequent projection check. Specifically, the identification of occlusions and the visible areas of guardrails can be achieved through two approaches: The first is a semantic segmentation model, which uses a lightweight semantic segmentation network (such as a MobileNet-based encoding / decoding structure) to perform forward inference on image frames. Pixel sets corresponding to categories such as vehicles, pedestrians, vegetation, and cones are labeled as occlusion pixel areas, while pixel sets corresponding to guardrail categories are labeled as visible guardrail pixel areas. This approach is robust to complex scenes with diverse occlusion patterns and is suitable for deployment environments with ample computing resources. The second approach is a rule-based approach, which extracts the visible areas of guardrails based on the guardrail's prior lateral position in the image (constrained by the sensor's extrinsic matrix) combined with the guardrail's unique color (grayish-white metallic), texture (periodic wave stripes), and edge (waveform contour lines) features. Simultaneously, it extracts occlusion areas based on the size and dynamic characteristics of foreground targets (such as motion regions detected by differential detection between adjacent frames). This approach requires no model inference, performs stably in scenes with sufficient daylight and significant contrast between the guardrail and background, and has low computational cost. The two schemes can be flexibly selected based on hardware computing power. Alternatively, the semantic segmentation scheme can be used as the main path, and when its confidence level is lower than the threshold, it can automatically fall back to the rule-based method as a degradation strategy to ensure the continuity of the recognition process in all scenarios.

[0089] Furthermore, laser reflection intensity information can serve as a cross-modal aid for image recognition: when the laser reflection intensity in the corresponding area of ​​the guardrail shows a significant abnormal decrease and image segmentation fails to identify obvious foreground occlusions, this area can be marked as a suspected occlusion and used in conjunction with the image segmentation results to improve the ability of occlusion judgment to perceive occlusions with weak visual features (such as thin vegetation or light fog coverage). The source of image frames is not limited to a single static image; when acquisition conditions permit, the fusion results of adjacent consecutive frames can be used to enhance the stability of region recognition in scenes with weak texture or motion blur.

[0090] S107. Project the spatial coordinates of the candidate abnormal event onto the image frame. If the projection point is located within the visible pixel area of ​​the guardrail, then the candidate abnormal event is confirmed to have passed the visibility check.

[0091] Spatial coordinate projection refers to the process of mapping the three-dimensional position information of candidate anomalies recorded in the lidar coordinate system to the corresponding two-dimensional pixel plane of the image frame after transformation by the sensor's intrinsic and extrinsic parameter matrices. The projection result can be in the form of a single center point, a minimum bounding box, a contour line segment, or a complete projection area, depending on the spatial geometric characteristics of the candidate anomaly and the required judgment accuracy, and is not limited to a single point form. Visibility verification refers to the verification process of judging whether the guardrail position corresponding to the candidate anomaly is in an observable state at the image semantic level based on the projection result. Its core function is to distinguish between two situations: "the guardrail position has a real structural anomaly" and "the guardrail is occluded by a foreground object, causing laser misjudgment." It constitutes the core node in the method where laser three-dimensional spatial perception and image two-dimensional semantic perception mutually verify each other.

[0092] This step is executed after the image semantic mask is generated in S106. Specifically, the judgment logic is not limited to simply checking whether the projection point is located inside the visible pixel area of ​​the guardrail. It also needs to specifically handle the uncertain situation where the projection point is near the boundary of the area: when the pixel distance between the projection point (or the centroid of the projection area) and the edge of the pixel area of ​​the occluder or the boundary of the visible area of ​​the guardrail is less than the preset buffer width (which can be set to 5 to 20 pixels depending on the image resolution), the candidate event is marked as "boundary to be confirmed" instead of being directly passed or rejected. This mark will be used as a basis for de-weighting in the confidence scoring stage of S108 to deal with some pixel errors that exist in the image segmentation boundary itself. Under special working conditions of image acquisition quality degradation (such as the guardrail area contrast being lower than the usable threshold due to backlight, rain, fog, or lens contamination), the visibility verification is switched to degraded mode: the candidate event itself is retained and enters S108, but the image quality penalty factor is included in the score to avoid batch of real anomalies being missed due to temporary degradation of acquisition conditions.

[0093] Furthermore, the visibility determination rules can be adaptively selected based on the anomaly type: for missing anomalies, the focus is on checking whether there are continuous holes in the projection area that penetrate the background; for deformable anomalies, the focus is on judging whether the continuity of the projection contour edge is interrupted in the image; this differentiated rule fits the actual visual manifestation of different damage types. In some implementations, visibility verification can also be performed on adjacent multi-frame images or temporally fused images to overcome the limitations of single-frame images under short-term illumination changes or motion blur.

[0094] In some preferred embodiments, the above steps specifically include: obtaining the external parameter matrix between the camera acquiring image frames and the lidar acquiring laser contour frame sequences, as well as the camera's internal parameter matrix; using the external parameter matrix to transform the spatial coordinates of candidate abnormal events from the lidar coordinate system to the camera coordinate system to obtain three-dimensional coordinates in the camera coordinate system; and using the internal parameter matrix to map the three-dimensional coordinates in the camera coordinate system to a two-dimensional image plane to obtain the projection point coordinates.

[0095] Determine whether the coordinates of the projected point fall within the boundary of the visible pixel area of ​​the guardrail; if they fall within the boundary and the coordinates of the projected point do not overlap with the pixel area of ​​the obstruction, then the candidate abnormal event is confirmed to have passed the visibility check.

[0096] The extrinsic parameter matrix refers to a 4×4 homogeneous matrix that describes the rigid transformation relationship from the lidar coordinate system to the camera coordinate system, and includes rotational components. With translation components The internal parameter matrix is ​​obtained through joint calibration—typically using a checkerboard target or a laser-camera joint calibration board. The correspondence between laser points and image corner points is collected under at least 10 different orientations, and then solved using the least squares method. The optimal reprojection error after calibration is no greater than 1.5 pixels. Calibration must be performed again if the sensor installation state changes. The internal parameter matrix is ​​a 3×3 matrix describing the camera's imaging geometry, including the horizontal focal length. Vertical focal length Principal point coordinates The radial / tangential distortion coefficients are obtained through single-target calibration (such as Zhang's method, which involves acquiring no fewer than 20 checkerboard images at different angles).

[0097] This step is executed sequentially for each candidate event after the calibration parameters have been loaded into the system. The coordinate transformation is completed in two steps. The first step involves transforming the 3D coordinates of anomaly points in the lidar coordinate system using the extrinsic parameter matrix. Transform to camera coordinate system:

[0098]

[0099] The second step is to map the 3D points in the camera coordinate system to pixel coordinates using an intrinsic parameter matrix. :

[0100]

[0101] After obtaining the coordinates of the projection point, the system sequentially determines whether it falls within the boundary of the visible pixel area of ​​the guardrail and whether it overlaps with the pixel area of ​​the obstruction. When the projection point is less than the preset buffer pixel width from any area boundary, it is marked as an uncertain zone, and no clear pass conclusion is output. Instead, it is handled by the S108 confidence scoring stage.

[0102] In handling special operating conditions, when the contrast of the guardrail area in the image is lower than the usable threshold due to nighttime, backlighting, or rain / fog (which can be determined by the blur score of the Laplacian gradient operator), the visibility check is downgraded: the topological judgment result of the projection point's position is still recorded, but an image quality penalty factor is included in the score, and the candidate event continues to enter S108 instead of being directly rejected, to prevent batch underreporting. The identification of the visible pixel area of ​​the guardrail and the pixel area of ​​the occluder is not limited to pixel-level overlap judgment, but can also use target detection boxes, visible area ratio, multi-frame joint image judgment, or laser reflection intensity-assisted judgment, depending on computing resources.

[0103] Furthermore, the visibility rules can be adaptively switched based on the anomaly type: missing anomalies prioritize checking whether the projected area corresponds to a background penetration hole in the image; deformable anomalies prioritize checking whether the continuity of the projected contour edge is interrupted in the image, realizing anomaly type-driven differential judgment logic, which has a more refined differentiation effect in reducing type-related false alarms.

[0104] S108. Extract the spatial deviation values ​​of candidate anomalies that have passed the visibility check, and calculate the confidence level of the candidate anomalies.

[0105] The confidence level rating refers to the output label obtained by quantifying the reliability of candidate anomaly detection events that have passed the visibility check. It is the core quality evaluation field used to drive alarm priority and manual review ranking in the guardrail anomaly detection record. Its form is not limited to "confidence level" but can also be expressed as risk level, maintenance priority, or re-inspection recommendation level, serving different application scenarios for maintenance decisions. The input dimensions involved in the scoring are not limited to the single spatial deviation value, but also cover height deviation (the difference between the measured height of the top of the guardrail and the standard height), spatial overlap (inter-frame stability index from S105), image visibility ratio (the ratio of the visible pixel area of ​​the guardrail to the projected area), occlusion boundary distance (the pixel distance of the projection landing point from S107), acquisition posture stability (the roll and bump amplitude of the vehicle at the current moment), and point cloud density (the number of effective points in the second screening point cloud). Compared with a single geometric index, the joint support of multi-dimensional evidence can significantly improve the stability of the scoring conclusion under complex conditions such as sparse point cloud, vehicle speed fluctuation, or image occlusion threshold.

[0106] This step, executed independently for each candidate event after the visibility verification in S107, is an integration node that converges multi-source geometric evidence and image semantic evidence into actionable decision conclusions. Specifically, the comprehensive confidence score can adopt two schemes: one is a linear weighted scheme using a rule engine, which normalizes each input item and multiplies it by a preset weight, which can be obtained by training and fitting historical labeled samples or determined by expert experience in combination with road test data; the other is a model-based scheme, which directly inputs the above multi-dimensional feature vectors into a gradient boosting tree, lightweight neural network, or Bayesian classifier, outputs continuous scores, and then maps them to graded intervals. This method has a stronger ability to model nonlinear feature interactions and is suitable for large-scale deployment scenarios with sufficient sample accumulation.

[0107] After scoring is completed, the system matches the comprehensive score with multiple preset confidence levels (such as high, medium, and low levels, or four levels). High confidence levels are directly included in the anomaly report and trigger an alarm; medium confidence levels are entered into the manual review queue, along with suggestions on the urgency of repair; low confidence levels are retained as suspected records, and the main scoring dimension labels that triggered the event are simultaneously output (such as "suspected point cloud sparsity," "suspected acquisition posture fluctuation," "suspected occlusion threshold," etc., root cause diagnostic labels) for subsequent cross-day review or historical data comparison. By outputting the grading conclusion and root cause labels simultaneously, the system not only provides the final judgment result but also outputs the diagnostic basis supporting the conclusion, reducing the information acquisition cost of manual review of each item.

[0108] In some preferred embodiments, the above steps specifically include:

[0109] Obtain the spatial overlap degree corresponding to the candidate abnormal event, and the pixel distance from the projection point to the edge of the pixel region of the occluder; calculate the basic abnormal score based on the product of the spatial deviation value and the preset scaling factor;

[0110] The basic anomaly score is weighted using spatial overlap, and then corrected using pixel distance to obtain a comprehensive confidence score. The comprehensive confidence score is then matched with multiple preset confidence intervals to determine the confidence level of the candidate anomaly event.

[0111] Among them, the preset proportional coefficient This refers to the linear conversion factor that maps spatial deviation values ​​(in millimeters) to standardized scoring units. Its value is determined jointly by the typical distribution range of deviation values ​​in historically labeled samples and the target scoring upper limit. It can be preset according to guardrail type or road grade, or dynamically adjusted through sample regression fitting. Basic anomaly score. (in The spatial deviation value is the starting point of the scoring process. It only reflects the severity of candidate anomalies in the geometric space dimension and has not yet incorporated evidence of temporal stability and visual reliability. It serves as the input basis for subsequent weighting and correction operations. The comprehensive confidence score is the final quantitative conclusion obtained by sequentially fusing the three types of evidence. After normalization, the value falls within the [0,1] interval, which directly corresponds to the preset grading interval, facilitating the engineering setting of grading thresholds and cross-engineering consistency comparison.

[0112] This step is performed independently for each candidate event after the S107 visibility verification passes, converging the multidimensional evidence scattered across various verification stages into a single quantitative conclusion that can drive downstream alarm and review decisions. Specifically, the system completes the comprehensive confidence score in three steps: First, based on the spatial deviation value... Multiply by the preset scaling factor The basic abnormality score was obtained. This score only measures the severity of candidate anomalies at the geometric level and serves as the starting point for subsequent weighting and correction operations. The second step involves using the spatial overlap from S105. The spatiotemporal correction score is obtained by weighting the base scores. Spatial overlap reflects the stability of the same anomaly across multiple consecutive frames: for genuine structural damage, the anomaly persists across multiple frames. When the value approaches 1, the spatiotemporal correction score is positively amplified; for occasional candidates caused by random noise, the anomalies disappear rapidly in subsequent frames. As the value approaches zero, the score is naturally suppressed. Thus, temporal stability modulates the geometric severity. The third step involves using a pixel distance normalization correction factor. The spatiotemporal correction score is applied to obtain the final comprehensive confidence score. .in The pixel distance between the projection point and the edge of the occluding pixel region: The larger the value, the farther the projection point is from the occlusion boundary, and the more reliable the visual judgment. The closer a value is to 1, the less discount is applied to the score. The smaller the value, the more likely the projection point is near the edge of the occlusion, raising questions about its visual reliability. As the weight decreases, the score for that candidate event is correspondingly downgraded. Thus, the credibility of visual semantics is quantitatively integrated into the overall score.

[0113] The above linear weighting scheme is logically transparent, the parameters are interpretable, and the adjustment is convenient. Each weight and proportional coefficient can be initialized by fitting historical manually labeled samples or by road test experience. In large-scale deployment scenarios with sufficient sample size, multi-dimensional feature vectors can also be directly fed into gradient boosting trees or lightweight neural networks to replace manually designed linear rules, thereby automatically modeling the nonlinear interaction relationship between features and obtaining stronger classification accuracy.

[0114] After the scoring is completed, the system will Matches with preset graded intervals and triggers corresponding actions: High confidence (example) This will be directly included in the anomaly report and trigger an alarm. It is recommended to immediately arrange a follow-up maintenance check; medium confidence level (example: ) Enter the manual review queue with a repair urgency suggestion; low confidence (example: As suspected records, they are retained, and root cause diagnostic labels for the scoring dimensions are output simultaneously (such as "sparse point cloud suspected", "occlusion critical suspected", "collection posture fluctuation suspected", etc.), providing diagnostic basis for subsequent cross-project playback comparison and historical trend analysis. The grading threshold and scoring weight can be continuously adjusted with the posterior labels accumulated in the project, so that the accuracy of the scoring model in adapting to local road condition features can be continuously improved, forming an iterative closed-loop accumulation mechanism for detection experience.

[0115] S109. Bind the confidence level, candidate abnormal events, corresponding location mileage data and image frames to output the guardrail abnormality detection record.

[0116] Data binding refers to the process of structurally associating and combining multidimensional heterogeneous data generated in each processing stage from S103 to S108 for the same candidate abnormal event, using the candidate event's unique spatiotemporal tracking identifier as the primary key. This ensures that each guardrail abnormality detection record in the final output constitutes a self-consistent evidence unit, which can independently support manual review, engineering playback, and maintenance decisions without the need for cross-data source lookup. The guardrail anomaly detection record refers to the structured data object output after the entire process of this method. Its fields include at least: unique anomaly identifier (UUID format is recommended for cross-frame event merging and cross-project deduplication), trigger timestamp (Unix epoch time, accuracy not less than 1ms), geographic coordinates (longitude, latitude, elevation), route location (station number or cumulative mileage from the starting point of the project), guardrail side (left / right), anomaly type label (deformation, offset, damage or missing), confidence level and comprehensive score value, root cause diagnosis label, associated image frame (complete image path or embedded thumbnail), associated point cloud slice file path, version number of the trigger parameter used in this detection, and device status snapshot at the time of acquisition (number of satellites, differential status, vehicle speed, attitude angle).

[0117] This step, executed after confidence grading in S108, is the evidence aggregation and result encapsulation node in the entire detection process, ensuring that each output record possesses complete evidence integrity traceable from the acquisition site to the decision-making level. Specifically, the system uses the candidate event spatiotemporal tracking identifier generated in S103 as the primary key, sequentially associating confidence grading (from S108), location odometer information (from the synchronous acquisition sequence in S101 and its subsequent odometer calibration results), and image frames (from the visibility verification associated frames retrieved in S106). After supplementing point cloud slices, device attitude snapshots, and acquisition status fields, the record encapsulation is completed. When storage resources permit, video clips consisting of several frames before and after can also be synchronously bound to support dynamic continuous observation during playback.

[0118] Record storage supports at least two modes: First, a local project database mode, archived by project in SQLite or structured binary format, supporting retrieval by station number segment, anomaly type, or confidence level, as well as historical comparative analysis of anomaly trends across days and projects for the same road segment; second, a project playback package mode, packaging all anomaly records along with corresponding point cloud slices, image frames, and positioning trajectories into independent project playback files. Operators can reproduce the collection process of each anomaly in playback mode at any time point for quality verification and post-test manual annotation. Output formats are not limited to static record lists: high-confidence anomalies can be synchronously pushed to the map annotation layer of the management platform (colored and annotated to the electronic map according to their level) or trigger immediate alarm messages; medium-confidence anomalies can be exported as maintenance work orders for subsequent processing in the maintenance management system; low-confidence suspected records serve as input for cross-day re-inspection tasks, realizing complete lifecycle management of anomaly states from "detection" to "confirmation" to "closed-loop handling." Before output, the system performs event deduplication and cross-frame merging based on spatiotemporal tracking identifiers on all candidate records within the current detection range: candidate events triggered multiple times at the same physical location in adjacent frames are merged into a single record, with the frame with the largest deviation value used as the representative frame. At the same time, the number of merged frames and the continuous coverage mileage field are retained in the record, providing a basis for quantitatively assessing the damage scale in maintenance decisions, avoiding the overestimation of the number of anomalies due to repeated inclusion of the same damage in the statistics, and reducing the risk of misallocation of maintenance resources.

[0119] The rapid detection device for highway corrugated guardrails in the embodiments of this invention is described below from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 2 This is a schematic diagram of the physical device structure of a highway corrugated guardrail rapid detection equipment in the embodiments of this application.

[0120] It should be noted that, Figure 2 The structure of the highway corrugated guardrail rapid detection device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0121] like Figure 2 As shown, the highway wave-shaped guardrail rapid detection device includes a central processing unit (CPU) 201, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 202 or a program loaded from storage section 208 into random access memory (RAM) 203, such as performing the methods described in the above embodiments. The RAM 203 also stores various programs and data required for system operation. The CPU 201, ROM 202, and RAM 203 are interconnected via bus 204. An input / output (I / O) interface 205 is also connected to bus 204.

[0122] The following components are connected to I / O interface 205: input section 206 including audio input devices, push-button switches, etc.; output section 207 including a liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 208 including a hard disk, etc.; and communication section 209 including a network interface card such as a LAN (Local Area Network) card, modem, etc. Communication section 209 performs communication processing via a network such as the Internet. Drive 210 is also connected to I / O interface 205 as needed. Removable media 211, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., are installed on drive 210 as needed so that computer programs read from them can be installed into storage section 208 as needed.

[0123] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 209, and / or installed from removable medium 211. When the computer program is executed by central processing unit (CPU) 201, it performs the various functions defined in the present invention.

[0124] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0125] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0126] Specifically, the rapid detection device for highway corrugated guardrails in this embodiment includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the rapid detection method for highway corrugated guardrails provided in the above embodiment.

[0127] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the rapid detection device for highway corrugated guardrails described in the above embodiments; or it may exist independently and not assembled into the rapid detection device for highway corrugated guardrails. The storage medium carries one or more computer programs, which, when executed by a processor of the rapid detection device for highway corrugated guardrails, cause the rapid detection device to implement the rapid detection method for highway corrugated guardrails provided in the above embodiments.

[0128] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0129] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0130] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A rapid detection method for highway corrugated guardrails, characterized in that, The method includes: As the mobile detection equipment travels along the road, it synchronously collects the laser contour frame sequence, image frame sequence, and location mileage sequence of the guardrail using a unified time reference, and records the corresponding timestamps. Extract point cloud data within a preset distance and height range from the laser contour frame sequence to generate corresponding guardrail contour features; Calculate the spatial deviation value between the guardrail outline feature and the preset guardrail reference feature. If the spatial deviation value is greater than the preset deviation threshold, mark the corresponding guardrail outline feature as a candidate abnormal event. Extract the timestamp and location mileage data corresponding to the candidate abnormal events, and along the driving direction of the mobile detection device, stitch together the laser contour frame corresponding to the timestamp and a preset number of adjacent laser contour frames to construct a continuous guardrail section. Within the continuous guardrail section, the spatial overlap of the candidate abnormal event in the adjacent multiple laser contour frames is calculated. If the spatial overlap is greater than a preset overlap threshold, the candidate abnormal event is confirmed to have passed the continuity check. Retrieve image frames with timestamps corresponding to the candidate abnormal events that have passed the continuity check, and identify the pixel areas of occlusion and the visible pixel areas of the guardrail in the image frames; The spatial coordinates of the candidate abnormal event are projected onto the image frame. If the projection point is located within the visible pixel area of ​​the guardrail, the candidate abnormal event is confirmed to have passed the visibility check. Extract the spatial deviation values ​​of the candidate anomaly events that pass the visibility check, and calculate the confidence level of the candidate anomaly events; The confidence level, the candidate abnormal events, the corresponding location mileage data, and the image frames are data-bound together to output the guardrail anomaly detection record.

2. The method according to claim 1, characterized in that, The step of extracting point cloud data within a preset distance and height range from the laser contour frame sequence to generate corresponding guardrail contour features specifically includes: The installation height of the mobile detection device is obtained, the original distance data of each laser point in the laser contour frame sequence is obtained, and laser points whose original distance data is less than the minimum detection distance or greater than the maximum detection distance are removed to obtain the first filtered point cloud. Based on the installation height of the device and the vertical offset of each laser point in the first screening point cloud, calculate the actual ground height corresponding to each laser point. The actual ground height is compared with the preset standard height of the guardrail, and laser points whose actual ground height is within the preset error allowable range of the standard height of the guardrail are extracted to obtain the second screening point cloud; Based on the second filtered point cloud, the three-dimensional contour model of the guardrail is extracted, and the waveform vertices and bottom fixed points of the guardrail are extracted from the three-dimensional contour model as the contour features of the guardrail.

3. The method according to claim 2, characterized in that, The calculation of the spatial deviation value between the guardrail outline feature and the preset guardrail reference feature, and if the spatial deviation value is greater than the preset deviation threshold, then the corresponding guardrail outline feature is marked as a candidate abnormal event, specifically including: Obtain the reference coordinates of the standard waveform vertex and the reference coordinates of the standard bottom fixed point in the preset guardrail reference feature; The spatial deviation value is obtained by calculating the first spatial distance between the waveform vertex and the standard waveform vertex, and the second spatial distance between the bottom fixed point and the standard bottom fixed point. When the first spatial distance is greater than the first preset threshold and the second spatial distance is less than the second preset threshold, it is determined that the guardrail has deformation or offset deviation, and the corresponding guardrail contour feature is marked as the candidate abnormal event. When the first spatial distance or the second spatial distance cannot be calculated and the density of the second filtered point cloud is lower than a preset density threshold, it is determined that the guardrail has a damaged or missing deviation, and the corresponding guardrail contour feature is marked as the candidate abnormal event.

4. The method according to claim 1, characterized in that, The process of the mobile detection device synchronously acquiring the laser contour frame sequence, image frame sequence, and location mileage sequence of the guardrail using a unified time reference, and recording the corresponding timestamps, specifically includes: The system receives positioning data in real time at a preset frequency through a GNSS positioning module of the Global Navigation Satellite System, and generates a sequence of location mileages including latitude, longitude and elevation. The wheel circumference and odometer component parameters of the mobile detection device are obtained, and the fixed-distance exposure triggering time is calculated in real time by combining the preset interval exposure distance. When the travel distance of the mobile detection device reaches the fixed-distance exposure trigger time, an exposure command is sent to the camera to acquire the sequence of image frames; Within the same trigger cycle of sending the exposure command, the lidar is controlled to acquire a laser contour frame, and the sequence of location mileage, the image frame, and the laser contour frame are bound to the same timestamp.

5. The method according to claim 4, characterized in that, After the step of synchronously acquiring the laser contour frame sequence, image frame sequence, and location mileage sequence of the guardrail at a unified time reference during the movement of the mobile detection device along the road, and recording the corresponding timestamps, the method further includes: Acquire the initial mileage data of the mobile detection device during actual road driving; Extract the sequence of location mileage output by the GNSS positioning module, and calculate the actual straight-line distance between adjacent positioning points; The actual straight-line distance in space is accumulated within a preset time window to obtain reference mileage data; Calculate the mileage error between the initial mileage data and the reference mileage data. If the mileage error exceeds a preset calibration threshold, then use the reference mileage data to perform mileage calibration and update the sequence of location mileages.

6. The method according to claim 1, characterized in that, The step of projecting the spatial coordinates of the candidate abnormal event onto the image frame, and confirming that the candidate abnormal event passes the visibility check if the projection point is located within the visible pixel area of ​​the guardrail, specifically includes: Obtain the external parameter matrix between the camera that acquires the image frames and the lidar that acquires the laser contour frame sequence, as well as the internal parameter matrix of the camera; The spatial coordinates of the candidate anomaly event are transformed from the lidar coordinate system to the camera coordinate system using the external parameter matrix to obtain the three-dimensional coordinates in the camera coordinate system. The three-dimensional coordinates in the camera coordinate system are mapped to the two-dimensional image plane using the internal parameter matrix to obtain the coordinates of the projection point. Determine whether the coordinates of the projection point fall within the boundary range of the visible pixel area of ​​the guardrail; If the projection falls within the boundary range and the coordinates of the projection point do not overlap with the pixel area of ​​the occluder, then the candidate abnormal event is confirmed to have passed the visibility check.

7. The method according to claim 6, characterized in that, The step of extracting the spatial deviation values ​​of candidate anomalies that have passed the visibility check and calculating the confidence level of the candidate anomalies specifically includes: Obtain the spatial overlap degree corresponding to the candidate abnormal event, and the pixel distance between the projection point and the edge of the pixel region of the occluder; The basic anomaly score is calculated based on the product of the spatial deviation value and the preset proportional coefficient. The basic anomaly score is weighted using the spatial overlap, and the weighted basic anomaly score is corrected using the pixel distance to obtain a comprehensive confidence score. The comprehensive confidence score is matched with multiple preset confidence intervals to determine the confidence level of the candidate abnormal event.

8. A rapid detection device for highway corrugated guardrails, characterized in that, The highway corrugated guardrail rapid detection device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the highway corrugated guardrail rapid detection device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instruction is executed on the highway corrugated guardrail rapid detection device, the highway corrugated guardrail rapid detection device performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the highway corrugated guardrail rapid detection device, the highway corrugated guardrail rapid detection device performs the method as described in any one of claims 1-7.

Images