A method for detecting missing anti-glare plates based on UAV video recognition

Abstract

The application discloses a kind of based on the anti-glare board missing detection method of unmanned aerial vehicle video identification, belong to road facility inspection and computer vision technical field. Including: collecting unmanned aerial vehicle inspection video, determine effective identification area;Detect the anti-glare board target that exists in image, and filter the effective target in effective identification area;According to the transverse center coordinates of anti-glare board detection frame, establish target sequence, calculate adjacent target spacing and determine reference adjacent spacing;When the interval to be judged exceeds the product of reference spacing and abnormal threshold coefficient, the interval is determined as missing candidate position;Matching confirmation is carried out to missing candidate position in continuous multiple frames, and the anti-glare board missing judgment result is output.The application utilizes the spatial characteristics that anti-glare board is continuous, repeated arrangement along road direction, and combines multiple frame time sequence confirmation, effectively reduces false alarm caused by single frame accidental miss detection and imaging quality fluctuation, improves the accuracy and reliability of anti-glare board missing detection in unmanned aerial vehicle inspection scene.

Description

Technical Field

[0001] This invention belongs to the field of road facility inspection and computer vision technology, specifically relating to a method for detecting missing anti-glare panels based on UAV video recognition, applicable to the automated inspection and missing identification of anti-glare panel facilities in the central median of highways. Background Technology

[0002] Anti-glare panels are common traffic safety facilities in the median strip of highways. They are typically installed continuously and repeatedly along the road to reduce glare from oncoming traffic at night and improve driving safety. Due to factors such as vehicle collisions, facility aging, construction disturbances, inclement weather, and human-caused damage, anti-glare panels may become missing, tilted, or damaged. Missing anti-glare panels, in particular, disrupt the continuity and integrity of road safety facilities, reduce their anti-glare effect, and make it difficult to effectively block glare from oncoming vehicles at night, increasing glare interference for drivers of oncoming vehicles and thus affecting road traffic safety. Therefore, accurately and quickly detecting missing anti-glare panels on highways and promptly identifying and locating their missing positions is of great significance for ensuring road traffic safety and improving road maintenance efficiency.

[0003] Existing inspection methods mainly include manual inspection, vehicle-mounted inspection, and fixed monitoring image recognition. Manual inspection is inefficient, and the inspection range and timeliness of problem discovery are affected by personnel allocation; vehicle-mounted inspection is limited by road conditions, shooting angle, and driving route; fixed monitoring equipment has limited coverage and cannot meet the needs of continuous inspection of long-distance highways. Drones can collect continuous video images along the road or median strip, facilitating rapid inspection of anti-glare panels along the route and verification of suspected locations.

[0004] Existing technologies include solutions for recognizing anti-glare panel status based on vehicle-mounted or road images, as well as solutions for UAV road inspection, road facility anomaly identification, and pavement defect detection. However, existing solutions typically lack a definitive judgment mechanism that combines effective recognition areas, target arrangement relationships, and continuous frame verification to address issues such as the changing target size of anti-glare panels in UAV inspection videos, insufficient imaging quality in some areas, and occasional missed detections in single frames. This can easily lead to false alarms or missed detections. Summary of the Invention

[0005] This invention provides a method for detecting missing anti-glare panels based on drone video recognition, aiming to solve the following technical problems:

[0006] First, it addresses the problem that during highway maintenance and inspection, missing anti-glare panels are mainly discovered through manual or vehicle-mounted inspections, resulting in low inspection efficiency, insufficient coverage continuity, and untimely discovery and location of missing locations, making it difficult to meet the needs of daily inspections and rapid verification of long-distance highways.

[0007] Second, this addresses the issue that in UAV inspection video images, factors such as perspective during shooting, changes in flight attitude, and the oblique arrangement of anti-glare panels in the image cause inconsistencies in the size, appearance, and clarity of anti-glare panels at different locations within the same frame, leading to unstable detection results for some targets and making it difficult to form an effective target sequence for subsequent missing data analysis.

[0008] Third, it addresses the problem that existing detection methods mainly focus on classifying the state of a single target, failing to fully utilize the spatial pattern of continuous and repetitive arrangement of anti-glare panels along the road direction within the effective identification area, and making it difficult to infer missing candidate locations through abnormal spacing between adjacent anti-glare panel targets.

[0009] Fourth, address the issue of insufficient stability in single-frame image judgment. That is, drone image jitter, temporary occlusion, changes in lighting, and occasional missed detections can easily distort the candidate results of missing single frames, leading to misjudgment or missed judgment of missing anti-glare panels.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] A method for detecting missing anti-glare panels based on drone video recognition includes the following steps:

[0012] Acquire video or continuous images by a drone flying along the extension of the median strip or anti-glare panel of a highway.

[0013] The video is preprocessed to obtain the image to be detected;

[0014] Based on the imaging position, size, and clarity of the anti-glare panel in the image, an effective recognition area is determined in the image to be detected.

[0015] The effective recognition area is input into the target detection model to identify the anti-glare panel targets that already exist within the effective recognition area, and to obtain the detection box position information and confidence level of each anti-glare panel target;

[0016] Valid anti-glare panel targets are selected based on confidence level, detection box position, and detection box overlap.

[0017] The selected anti-glare panel targets are arranged into a target sequence based on the horizontal center coordinates of the detection frame from small to large, thus forming the anti-glare panel target sequence.

[0018] Calculate the distance D between adjacent targets to be determined between adjacent anti-glare panel targets, and determine the reference adjacent distance d based on the statistical results of the distance between adjacent targets in the same frame or the local statistical results after removing outliers;

[0019] When the distance D between adjacent targets to be determined satisfies D>k×d, and the corresponding interval is located within the effective recognition area, the interval is determined as a candidate location for the missing anti-glare panel, where k is the abnormal threshold coefficient.

[0020] Inter-frame matching is performed on missing candidate positions in multiple consecutive frames. When the candidate results that are determined to be the same anti-glare panel interval position appear repeatedly in multiple consecutive frames or multiple adjacent frames, the anti-glare panel missing determination result is output.

[0021] Output the missing location, corresponding image frame, missing candidate region, inspection route, and time information.

[0022] Furthermore, the preprocessing includes at least one of frame extraction, diffusing, image enhancement, distortion correction, and region cropping.

[0023] Furthermore, the effective identification area is a local area on one side of the image selected based on the imaging position and clarity of the anti-glare panel in the image. The imaging size, appearance, and clarity of the anti-glare panel target within the effective identification area meet the requirements of target detection and spacing analysis.

[0024] Furthermore, the target detection model is a YOLO series target detection model, trained with the anti-glare panel as the only detection category.

[0025] Furthermore, the screening of effective anti-glare targets includes: filtering targets with confidence levels below a preset threshold, filtering targets located outside the effective recognition area, and filtering overlapping duplicate detection boxes based on the intersection-union ratio threshold used for model inference.

[0026] Furthermore, the distance D between adjacent targets to be determined is the center point distance between adjacent anti-glare targets, the distance of the detection frame boundary, or the projection distance along the road direction.

[0027] Furthermore, the reference adjacent spacing d is determined based on the median, average, or local statistical results after removing outliers of all adjacent target spacings within the same frame.

[0028] Furthermore, the inter-frame matching is as follows: the center point coordinates or interval positions of missing candidate positions in adjacent frames are matched. When the number of times the same candidate position that is successfully matched appears repeatedly in consecutive frames or adjacent frames reaches a preset threshold, it is confirmed that the same candidate position that is successfully matched has a missing anti-glare panel.

[0029] Furthermore, the determination of missing anti-glare panels also includes the cumulative number of times the candidate position appears within the sliding window.

[0030] Based on the above technical solutions, the invention utilizes the characteristic of anti-glare panels being continuously and repeatedly arranged along the central median strip of highways to transform the judgment of missing anti-glare panels into the judgment of abnormal arrangement among identified anti-glare panel targets. Since the missing location itself usually does not have a stable and clear target appearance, if "missing" is directly used as the target detection category, it is easily affected by image quality, temporary occlusion, and occasional missed detections, resulting in unstable detection results.

[0031] This invention first determines the effective recognition area suitable for target detection and missing data analysis based on the image size, appearance, and clarity of the anti-glare panel. Areas with unfavorable shooting angles, excessively small target sizes, indistinct texture features, or blurred edges are excluded from the primary detection and analysis areas of the current frame, thereby reducing the interference of low-quality imaging areas on subsequent target selection and spacing calculation.

[0032] Within the effective recognition area, the target detection model is used to identify existing anti-glare panel targets and obtain the position information of each target. The detection results are then filtered to remove low-confidence targets, duplicate bounding boxes, or targets that clearly do not conform to the imaging characteristics of the anti-glare panel, forming a set of effective anti-glare panel targets for arrangement analysis.

[0033] The screened anti-glare panel targets are sorted according to the road extension direction or the main arrangement direction in the image to form an anti-glare panel target sequence, and the distance between any adjacent targets to be judged is denoted as D. Due to the influence of perspective, the imaging distance at different positions in the same image may vary. Therefore, the reference adjacent distance d can be determined based on the distance between adjacent targets near the position to be judged, or it can be determined using the median, average, local fitting results, or historical statistical results of the same frame distance. When calculating the reference adjacent distance d, distances that deviate significantly from the neighborhood level are eliminated or downweighted to avoid suspected missing gaps from participating in the reference value calculation. If D satisfies D > k × d, the corresponding interval is identified as an abnormal gap that does not conform to the continuous arrangement pattern of anti-glare panels and is determined as a candidate missing position. Here, k is the abnormal threshold coefficient, which can be set according to the anti-glare panel installation spacing, image scale changes, and inspection scenario. This method does not identify the appearance of the "missing" itself, but determines the candidate missing position based on the spatial relationship between the detected anti-glare panels.

[0034] This invention does not directly determine the absence of an anti-glare panel based on single-frame candidate results. The system matches candidate regions in adjacent frames; only candidate results appearing in a single frame are saved as results to be verified, and no missing conclusion is directly output; when candidate results determined by matching to be at the same anti-glare panel interval position appear repeatedly in multiple consecutive frames or multiple adjacent frames, it is confirmed that there is a missing anti-glare panel at that position, thereby reducing single-frame misjudgments caused by temporary occlusion, image blurring, or occasional missed detection.

[0035] The technical effects achieved by this invention are:

[0036] First, compared to directly classifying "missing anti-glare panels" as a detection category, this invention first detects multiple existing anti-glare panel targets, and then analyzes their arrangement relationship and changes in adjacent spacing. Since the missing location itself lacks stable appearance features, the abnormal gap judgment method allows the determination of the missing location to be based on the spatial relationship of the identified targets, which matches the scenario where anti-glare panels are continuously arranged along the road direction.

[0037] Second, compared to methods that rely solely on single-frame image judgment, this invention matches missing candidate positions across consecutive frames. If an anti-glare panel is accidentally not detected in a single frame, the system does not immediately output a missing conclusion; a missing panel is only determined when an abnormal gap at the same anti-glare panel interval appears repeatedly in multiple frames. This reduces single-frame misjudgments caused by flight jitter, partial occlusion, lighting changes, and occasional missed detections.

[0038] Third, compared with general road facility defect identification schemes, this invention combines the UAV highway inspection scenario, the repetitive arrangement characteristics of anti-glare panels, effective identification area constraints, arrangement anomaly analysis, and multi-frame confirmation to address the issues of anti-glare panel imaging scale changes and single-frame detection instability from the UAV perspective, and performs missing detection. For other traffic ancillary facilities that are continuously or approximately equally spaced along the road direction, such as guardrail posts, delineators, and mileage markers, the same missing inference approach can be adopted after adjusting the target detection category, reference spacing calculation method, and anomaly threshold.

[0039] Fourth, this invention can provide automated inspection results for highway maintenance, reduce the workload of manual segment-by-segment inspection, and improve the efficiency of long-distance road inspection and the timeliness of detecting missing problems. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of the anti-glare panel missing detection method of the present invention.

[0041] Figure 2 This is a schematic diagram of the effective identification region constraint in this invention. (a) represents the effective identification region; (b) represents the non-primary analysis region; the boundary between the effective identification region and the non-primary analysis region is the ROI boundary.

[0042] Figure 3 This is a schematic diagram comparing the normal arrangement and abnormal gaps of the anti-glare panel in this invention. Wherein (a) is the normal arrangement state, and d represents the reference adjacent spacing; (b) is the state with missing parts, and D represents the adjacent target spacing to be determined. Detailed Implementation

[0043] The implementation of the present invention will now be described with reference to the accompanying drawings. These embodiments are illustrative and do not constitute a limitation on the scope of protection of the present invention.

[0044] Example 1

[0045] To address the problems of relying on manual discovery for missing anti-glare panels in existing highway inspections, unstable identification of anti-glare panels in UAV images due to the influence of shooting angle and perspective, and the tendency for false alarms and missed alarms in single-frame judgment, this invention discloses an anti-glare panel missing detection method based on UAV video recognition. Figure 1 This is an overall flowchart of the anti-glare panel missing detection method of the present invention. Figure 2 To effectively identify the region constraint diagram, Figure 3 This diagram illustrates the comparison between normal and abnormally spaced anti-glare panels. (See attached image.) Figures 1 to 3 The method of the present invention includes the following steps:

[0046] Step S01: Unmanned Aerial Vehicle (UAV) Inspection Video Acquisition

[0047] The drone flies along the extension direction of the median strip or anti-glare panel of a highway, capturing video images containing the anti-glare panel. In one embodiment, the drone cruises along the extension direction of the anti-glare panel, so that the video images continuously contain multiple anti-glare panel targets, providing continuous video data for subsequent target sorting, spacing calculation, and multi-frame confirmation.

[0048] Step S02: Image frame extraction and preprocessing

[0049] The video captured by the UAV is frame-by-frame extracted to obtain continuous image frames, and then processed according to image quality, including shading, cropping, enhancement, distortion correction, or format conversion. This step is used to reduce the impact of UAV flight jitter, image exposure variations, and image distortion on subsequent target detection and distance calculation. In this embodiment, the system reads the UAV inspection video frame by frame through the video reading interface, and uses the currently successfully read image frame as the image to be detected; then, it calls the input preprocessing interface of the target detection model to perform scaling, edge filling, and normalization according to the model input size to generate model input data. The preprocessing results corresponding to the same input size are cached and reused during the processing of the current frame; if the cached results are not obtained, the current image frame is directly passed to the target detection interface, and the model internally completes the input preprocessing and inference.

[0050] Step S03: Determine the valid recognition area

[0051] See appendix Figure 2 Based on the drone's shooting angle, the location of the median strip, and the imaging quality of the anti-glare panel in the image, an effective recognition area (a) is determined for target detection and missing data analysis. In one embodiment, a local area on one side of the image can be selected as the effective recognition area based on the imaging position and sharpness of the anti-glare panel in the image. This effective recognition area represents the anti-glare panel target to be detected and is used for target detection and distance analysis, such as... Figure 2As shown, in this embodiment, the left 2 / 5 region of the image is considered. Other regions containing anti-glare targets are considered non-primary analysis regions (b) due to their smaller scale, weaker texture, or blurred edges. Since these non-primary analysis regions have smaller imaging scales or lower recognizability, their detection results are not included in the target ranking and missing data analysis for this frame. This embodiment, considering the imaging situation when the UAV collects data along the extension direction of the anti-glare target, designates the left-hand region of the image with larger and clearer anti-glare targets as the effective recognition region. Let the current frame width be w and the height be h, and the effective recognition region be set as... The effective identification region serves as the post-processing filtering range for the target detection results, while the target detection model still performs inference on the entire frame image to be detected. In this invention, the ROI boundary between the effective identification region (a) and the non-primary analysis region (b) can be set according to the shooting angle and imaging quality.

[0052] Step S04: Anti-glare plate target detection

[0053] This embodiment employs the YOLO object detection model. This model includes a feature extraction backbone network, a multi-scale feature fusion network, and an object detection head. The feature extraction backbone network extracts features from the anti-glare panel image, the multi-scale feature fusion network fuses target features at different scales, and the object detection head determines the location, category, and confidence score of the candidate detection boxes for the anti-glare panel. This embodiment uses the YOLO26 object detection model, with the inference image size parameter `img_size` set to 960. The object detection results include the location information, category information, and confidence score of the anti-glare panel detection boxes.

[0054] Step S05: Screening of test results

[0055] After model inference, the boundary positions, categories, and confidence scores of each candidate detection box are recorded. The frame number and timestamp are used as frame-level information and associated with the detection result of that frame. During object detection model inference, the confidence threshold and intersection-over-union (IoU) threshold are received from the runtime configuration. The coordinates of the top-left and bottom-right corners of the i-th candidate detection box are set as follows: Its horizontal center coordinates are The system filters valid anti-glare targets based on confidence level, detection box position, and overlap relationship. Targets with confidence levels below a preset threshold, targets located outside the effective recognition area, and overlapping duplicate detection boxes are filtered out based on the intersection-union ratio (IU / U) threshold used for model inference. Since adjacent anti-glare panels should generally not significantly overlap from the UAV's perspective, a duplicate box filtering rule reduces the likelihood of the same anti-glare panel being identified as multiple targets, preventing duplicate detection boxes from participating in target sorting and adjacent distance calculations. Only those meeting the criteria are retained. The detection bounding boxes. When there are fewer than three anti-glare panel detection bounding boxes remaining in the effective recognition area, no missing candidate analysis is performed in this frame; when there are three or more, target sorting and adjacent spacing calculation are performed.

[0056] Step S06: Anti-glare panel target sorting

[0057] Based on the horizontal center coordinates of the selected anti-glare panel detection frames, a target sequence is established from smallest to largest, which is used to calculate the horizontal spacing between adjacent detection frames. Step S07: Adjacent Spacing Calculation

[0058] Calculate the center point distance, detection frame boundary distance, or projection spacing along the road direction between adjacent anti-glare panel targets, and denote the distance between adjacent targets to be determined as D. Determine the reference adjacent spacing d based on the distance between adjacent targets near the location to be determined, or using the median, average, local statistical results after removing outliers, or historical statistical values ​​of the same frame spacing. This embodiment uses the lateral center coordinate spacing for calculation. Let n be the number of detection frames retained after filtering in step S05, and the lateral center coordinate sequence formed in step S06 be... ,according to Calculate the spacing between adjacent elements in the original sequence, and the number of spacings. .Will After copying, sort them from smallest to largest, and remove the smallest one. Take one interval, and then take the 20th percentile of the remaining interval as the reference adjacent interval d; when When the pixel value is zero, no missing candidate is generated in this frame.

[0059] Step S08: Determine the missing candidate position

[0060] When the distance D between adjacent targets to be determined satisfies D > k × d, and the corresponding interval is located within the effective recognition area, this interval is determined as a candidate location for missing anti-glare panels. This step does not directly identify the appearance of the "missing target," but determines the candidate location based on the distance relationship between the identified anti-glare panel targets. This embodiment sets a local large distance determination coefficient. and periodic spacing determination coefficient First calculate and ;when When local large spacing determination is enabled, when Periodic spacing determination is enabled; if neither of the two conditions is met, no missing candidate is generated for this frame. R max The meaning is: the ratio of the maximum value of the distance between the i-th adjacent targets to the reference adjacent distance d; R med The meaning is: the ratio of the median distance between the i-th adjacent targets to the reference adjacent distance d. For each in the original order Judge one by one: When the value is greater than 1.8, local large spacing judgment is enabled; When the value is greater than 1.4, periodic spacing determination is enabled. Only when all conditions for enabling the current frame are met will the corresponding interval be identified as a candidate location for missing anti-glare panels.

[0061] Step S09: Multi-frame temporal fusion and voting confirmation

[0062] The missing candidate positions in multiple consecutive frames are confirmed through multi-frame candidate position matching. The inter-frame matching involves matching the center point coordinates or interval positions of missing candidate positions in adjacent frames. When the number of times the same successfully matched candidate position appears repeatedly in consecutive or adjacent frames reaches a preset threshold, it is confirmed that the same successfully matched candidate position has a missing anti-glare panel. That is, if the candidate result determined to be the same anti-glare panel interval position appears repeatedly in multiple consecutive or adjacent frames, it is confirmed that the anti-glare panel is missing at that position; if the candidate result only appears in a single frame, it is saved as a result to be verified, and the missing conclusion is not directly output.

[0063] This embodiment uses the horizontally normalized position of the candidate interval within the effective recognition area to perform multi-frame candidate position matching and merging confirmation. Let... The right boundary of the left detection box. The left boundary of the right detection box. and To effectively identify the left and right boundaries of the region, the normalized position of the candidate interval is calculated using the following formula:

[0064] ;

[0065] And Limited to Within the specified range, the system writes the candidate positions for each frame into a sliding window of seven frames in length; if the range exceeds seven frames, the earliest frame is removed.

[0066] The current candidate position is compared with the center position of an existing candidate position group. If the absolute difference between the two is less than or equal to 0.025, the current candidate position is added to that candidate position group, and the center position of the group is updated accordingly. If no candidate position group meets the condition, a new candidate position group is created. If the cumulative count of a candidate position group within the sliding window reaches four, the candidate position group is updated from the pending confirmation state to the confirmed state.

[0067] Step S10: Output detection results

[0068] The system outputs the location of the missing anti-glare panel, the corresponding image frame, the candidate missing region, the cumulative number of times the candidate location appears within the sliding window, and the confirmation status, as well as the inspection route and time information. It can also generate inspection reports or early warning records for manual review, maintenance dispatch, and on-site repair scheduling. In this embodiment, the frame-level output records the frame sequence number and timestamp; the target detection results save the anti-glare panel detection frame location, category, and confidence level; and the missing panel analysis results save the candidate location, candidate region, cumulative number of occurrences, and confirmation status.

[0069] In the system implementing the above method, the method can be executed by a UAV image acquisition module, an image preprocessing module, an effective recognition area determination module, an anti-glare panel target detection module, a detection result filtering module, an anti-glare panel target sorting module, an adjacent distance calculation module, a missing candidate judgment module, a multi-frame confirmation module, and a result output module. The system includes: a drone image acquisition module for acquiring video or continuous images of the highway anti-glare panel area; an image preprocessing module for frame extraction, enhancement, cropping, or shading; an effective recognition region determination module for determining the image region used for screening and arranging the anti-glare panel detection results; in this embodiment, the effective recognition region is the left 2 / 5 of the image determined based on the anti-glare panel imaging position and clarity; an anti-glare panel target detection module for inferring from the entire frame image to be detected and outputting candidate detection boxes for the anti-glare panels; a detection result screening module for screening anti-glare panel targets located within the effective recognition region based on confidence level, detection box overlap, and lateral center coordinates; an anti-glare panel target sorting module for forming an anti-glare panel target sequence; an adjacent distance calculation module for calculating reference adjacent distance and abnormal distance; a missing candidate judgment module for identifying abnormal gaps; a multi-frame confirmation module for performing time-series voting confirmation of candidate missing positions; and a result output module for outputting missing positions, screenshots, and inspection results.

[0070] The target detection model of this invention is trained and validated using a self-built anti-glare panel image dataset. The training objective is to identify existing anti-glare panels in UAV inspection images and output the corresponding detection boxes and confidence scores. The target detection model uses YOLO26s, and the inference image size parameter `img_size` is set to 960. Anti-glare panel absence is not treated as an independent training category but is determined based on the anti-glare panel target detection results through effective identification of region constraints, adjacent spacing judgment, and multi-frame confirmation. Under the corresponding test environment, the target detection model achieves an mAP of 0.88 and an inference speed of 210 FPS. The above metrics reflect the recognition performance and processing speed of the target detection model in this embodiment, but are not equivalent to the accuracy of anti-glare panel absence determination, nor do they constitute a limitation on the implementation conditions of this invention.

[0071] Example 2

[0072] In practical applications, inspection personnel first plan the drone inspection route, causing the drone to fly along the extension direction of the highway median strip or anti-glare panel, and collect video or continuous images containing continuous anti-glare panel targets. After the collection is completed, the video images are transmitted in real time or imported offline to the ground station, server, or inspection analysis system.

[0073] The system performs frame extraction and preprocessing on the video images. Based on the drone's shooting angle, the imaging quality of the anti-glare panel, and the location of the median strip, it determines the effective recognition area of ​​the current frame, identifies the anti-glare panel targets within that area, and filters out valid targets. See appendix. Figure 3 The system establishes a target sequence based on the horizontal center coordinates of each detection frame, from smallest to largest, and denotes the distance between adjacent targets to be judged as D, and determines the corresponding reference adjacent distance d. When D satisfies D > k × d, the corresponding interval is listed as a missing candidate position. k represents the anomaly threshold coefficient. The system then matches the candidate positions in consecutive frames; when the candidate results for the same anti-glare panel interval position appear repeatedly, the anti-glare panel missing result is output.

[0074] The output may include a screenshot of the missing location, the corresponding video frame, candidate missing regions, the number of frames in which the candidate results appear, or the confidence level calculated from multiple frames, as well as the inspection route and time information. Maintenance personnel can use this information for manual review and to arrange on-site repairs.

[0075] In summary, this embodiment completes the anti-glare panel missing determination in the following order: First, target detection is performed on existing anti-glare panels in the entire frame image to be detected; then, the detection results are filtered based on the effective recognition area; a target sequence is established based on the filtering results, and adjacent distance analysis is performed to determine the missing candidate positions; finally, the missing candidate positions in multiple consecutive frames are matched and confirmed. The target detection model is only used to obtain the detection results of existing anti-glare panels; the conclusion of anti-glare panel missing is formed by the adjacent distance judgment and multi-frame confirmation.

[0076] It should be further noted that the above embodiments are merely for understanding the technical solution of the present invention and are not intended to limit the scope of protection of the present invention. Any obvious adjustments and modifications made to the technical solution of the present invention that fall within the inventive concept should also fall within the scope of protection of the present invention.

Claims

1. A method for detecting missing anti-glare panels based on UAV video recognition, characterized in that, Includes the following steps: Acquire video footage collected by a drone flying along the extension direction of the median strip or anti-glare panel of a highway. The video is preprocessed to obtain the image to be detected; Based on the imaging position, size, and clarity of the anti-glare panel in the image, an effective recognition area is determined in the image to be detected. The effective recognition area is input into the target detection model to identify the anti-glare panel targets that already exist within the effective recognition area, and to obtain the detection box position information and confidence level of each anti-glare panel target; Valid anti-glare panel targets are selected based on confidence level, detection box position, and detection box overlap. The selected anti-glare panel targets are arranged into a target sequence based on the horizontal center coordinates of the detection frame from small to large, thus forming the anti-glare panel target sequence. Calculate the distance D between adjacent targets to be determined between adjacent anti-glare panel targets, and determine the reference adjacent distance d based on the statistical results of the distance between adjacent targets in the same frame; When the distance D between adjacent targets to be determined satisfies D>k×d, and the corresponding interval is located within the effective recognition area, the interval is determined as a candidate location for the missing anti-glare panel, where k is the abnormal threshold coefficient. Inter-frame matching is performed on missing candidate positions in multiple consecutive frames. When the candidate results that are determined to be the same anti-glare panel interval position appear repeatedly in multiple consecutive frames or multiple adjacent frames, the anti-glare panel missing determination result is output. Output the missing location, corresponding image frame, missing candidate region, inspection route, and time information.

2. The method according to claim 1, characterized in that, The preprocessing includes at least one of frame extraction, decluttering, image enhancement, distortion correction, and region cropping.

3. The method according to claim 1, characterized in that, The effective recognition area is a local area on one side of the image selected based on the imaging position and clarity of the anti-glare panel in the image. The imaging size, appearance, and clarity of the anti-glare panel target within the effective recognition area meet the requirements of target detection and spacing analysis.

4. The method according to claim 1, characterized in that, The target detection model is a YOLO series target detection model, trained with the anti-glare panel as the only detection category.

5. The method according to claim 1, characterized in that, The selection of effective anti-glare targets includes: filtering targets with confidence levels below a preset threshold, filtering targets located outside the effective recognition area, and filtering overlapping duplicate detection boxes based on the intersection-union ratio threshold used for model inference.

6. The method according to claim 1, characterized in that, The distance D between adjacent targets to be determined is the distance between the center points of adjacent anti-glare targets, the distance of the detection frame boundary, or the projection distance along the road direction.

7. The method according to claim 1, characterized in that, The reference adjacent spacing d is determined based on the median, average, or local statistical results after removing outliers among all adjacent target spacings within the same frame.

8. The method according to claim 1, characterized in that, The inter-frame matching is as follows: the center point coordinates or interval positions of missing candidate positions in adjacent frames are matched. When the number of times the same candidate position that is successfully matched appears repeatedly in consecutive frames or adjacent frames reaches a preset threshold, it is confirmed that the same candidate position that is successfully matched has a missing anti-glare panel.

9. The method according to claim 1, characterized in that, The determination of missing anti-glare panels also includes the cumulative number of times the candidate position appears within the sliding window.

Images