Urban infrastructure automatic detection method based on video image algorithm

Abstract

The invention is suitable for the technical field of urban facility detection, and provides an urban infrastructure automatic detection method based on a video image algorithm, and the method comprises the steps: obtaining real-time video stream data of multiple types of facilities in an urban block; performing disturbance processing to generate a plurality of groups of disturbance video frame sequences; calculating a similarity score between the feature vector of each video frame in the multiple groups of disturbance sequences and the weight vector of the pre-training defect recognition model; screening the video frame sequence according to a preset threshold value, and generating an updated video frame sequence; performing defect detection on the updated video frame sequence; matching the defect area with a preset facility template library, dividing a template image area, selecting a target area with the maximum parameter variation degree, calculating a position deviation value of the defect area relative to a reference point of the target area, and correcting defect positioning coordinates based on the position deviation value; and generating an urban infrastructure detection result according to the defect positioning coordinates. And the false detection rate of facility defect detection is effectively reduced.

Description

Technical Field

[0001] This invention relates to the field of urban infrastructure inspection technology, and in particular to an automated inspection method for urban infrastructure based on video image algorithms. Background Technology

[0002] Urban infrastructure such as road guardrails, drainage manhole covers, electromechanical equipment boxes, and sanitation facilities constitute the core carriers for ensuring the safe operation of cities. During long-term service, these facilities are susceptible to structural damage, displacement, deformation, corrosion, and aging due to environmental erosion, mechanical loads, and human factors. For example, a displaced manhole cover may cause a vehicle to overturn, and a broken guardrail may lead to a traffic accident. Such defects not only threaten public safety but also cause economic losses.

[0003] Currently, mainstream inspection methods mainly rely on manual inspections and random checks using fixed camera videos. Manual inspections require checking each facility individually, with a daily inspection efficiency of less than 200 facility units, which cannot meet the real-time monitoring needs of millions of facilities in megacities. Furthermore, video analysis systems based on static rules have a high false detection rate in complex scenarios such as rainy, foggy weather, and low-light conditions at night. They are particularly prone to errors in detecting progressive defects such as microcracks and early corrosion, leading to delays in maintenance decisions. Summary of the Invention

[0004] This invention provides an automated detection method for urban infrastructure based on video image algorithms to address the problem of high false detection rates of defects.

[0005] This invention provides an automated detection method for urban infrastructure based on video image algorithms, comprising:

[0006] Acquire real-time video stream data of various types of facilities within the urban area, including municipal, greening, drainage, electromechanical, and sanitation facilities;

[0007] The real-time video stream data is perturbed to generate multiple sets of perturbed video frame sequences;

[0008] Calculate the similarity score between the feature vector of each video frame in the multiple perturbation sequences and the weight vector of the pre-trained defect recognition model; filter the video frame sequences according to a preset threshold, delete video frames with similarity scores lower than the threshold, and generate an updated video frame sequence;

[0009] Defect detection is performed on the updated video frame sequence to identify defect areas of the various types of facilities and ancillary facilities;

[0010] The defect area is matched with a preset facility template library, the template image area is divided and the target area with the largest parameter variation is selected. The positional deviation of the defect area relative to the reference point of the target area is calculated, and the defect positioning coordinates are corrected based on the positional deviation.

[0011] The urban infrastructure inspection results are generated based on the defect location coordinates.

[0012] Furthermore, the multiple sets of perturbation video frame sequences include spatial perturbation frame sequences, color perturbation frame sequences, and noise-resistant perturbation frame sequences.

[0013] Furthermore, the pre-trained defect recognition model is a multi-scale feature fusion network, comprising:

[0014] The backbone feature extraction layer is used to extract multi-level spatial features of video frames using the ResNet-50 architecture;

[0015] The attention enhancement layer is used to superimpose a channel attention module at the output of the backbone feature extraction layer to generate a high-dimensional feature vector with adjusted weights.

[0016] The defect classification output layer is used to store learnable weight vectors corresponding to various facility defect types. These learnable weight vectors are generated through end-to-end training.

[0017] Furthermore, the calculation of the similarity score between the feature vector of each video frame in the multiple perturbation sequences and the weight vector of the pre-trained defect recognition model includes:

[0018] The initial feature vectors of each frame in the spatial perturbation frame sequence, color perturbation frame sequence, and noise-resistant perturbation frame sequence are extracted respectively.

[0019] The initial feature vectors are fused and then input into the backbone feature extraction layer of the pre-trained defect recognition model to generate high-dimensional semantic feature vectors.

[0020] Calculate the similarity metric between the high-dimensional semantic feature vector and each weight vector in the defect classification output layer, and select the maximum similarity metric as the similarity score of the current video frame.

[0021] Furthermore, the step of filtering the video frame sequence according to a preset threshold, deleting video frames with similarity scores lower than the threshold, and generating an updated video frame sequence includes:

[0022] Generate a defect probability threshold based on the environmental state parameters of the current video stream;

[0023] Differentiated channel filtering conditions are set for the spatial perturbation frame sequence, color perturbation frame sequence, and noise-resistant perturbation frame sequence, respectively;

[0024] Based on the channel filtering conditions, video frames with unsatisfactory similarity scores in each sequence are deleted in parallel. Spatiotemporal consistency verification is performed on the deleted multi-channel sequences, and video frames that meet the cross-channel verification conditions are retained.

[0025] Generate an updated video frame sequence with timestamp alignment based on the verification results.

[0026] Furthermore, the defect detection performed on the updated video frame sequence to identify defective areas of the various types of facilities and ancillary facilities includes:

[0027] The updated video frame sequence is divided into multi-scale regions to generate a set of facility sub-regions;

[0028] Calculate the defect feature parameters for each facility sub-region, including texture complexity, edge breakage, and color anomaly.

[0029] The defect feature parameters are input into a pre-trained defect classifier, which outputs defect regions with labeled contours.

[0030] Furthermore, the step of matching the defective region with a preset facility template library, dividing the template image region, and selecting the target region with the greatest parameter variation includes:

[0031] Retrieve standard template images corresponding to the current facility type from the preset facility template library, and divide the standard template images into grid regions to generate multiple candidate template regions;

[0032] Calculate the parameter variation degree of each candidate template region, whereby the parameter variation degree characterizes the degree of drastic change in pixel features within the current region;

[0033] The candidate template region with the largest parameter variation is selected as the target region, and a mapping relationship between the target region and the defect region is established.

[0034] Furthermore, the calculation of the parameter variation degree of each candidate template region includes:

[0035] Extract the gradient magnitude matrix of the candidate template region, and calculate the variance of the gradient magnitude matrix as the degree of structural variability;

[0036] Extract the color channel histogram of the candidate template region, and calculate the entropy value of the color channel histogram as the color change degree;

[0037] The parameter variation is generated by weighted fusion of structural variation and color variation.

[0038] Furthermore, the calculation of the positional deviation of the defect area relative to the reference point of the target area, and the correction of the defect location coordinates based on the positional deviation, includes:

[0039] Set a geometric center reference point in the target area;

[0040] Locate the feature point in the defect region that corresponds to the geometric center reference point;

[0041] The two-dimensional Euclidean distance between the feature point and the geometric center reference point is calculated as the positional deviation. The defect area contour coordinates are translated according to the positional deviation to generate the corrected defect location coordinates.

[0042] Furthermore, the urban infrastructure monitoring results include facility maintenance decision-making instructions, safety early warning information, and smart parking guidance schemes.

[0043] As can be seen from the above technical solutions, the present invention has the following advantages:

[0044] This invention acquires real-time video stream data of various types of facilities within an urban area, perturbs the video stream data to generate multiple sets of perturbed video frame sequences, calculates the similarity score between the feature vector of each video frame in the multiple perturbed sequences and the weight vector of a pre-trained defect recognition model, filters the video frame sequences according to a preset threshold, deletes video frames with similarity scores below the threshold, and generates updated video frame sequences, performs defect detection on the updated video frame sequences, identifies defect areas of various types of facilities and ancillary facilities, matches the defect areas with a preset facility template library, divides template image regions and selects the target region with the largest parameter variation, calculates the positional deviation of the defect area relative to the reference point of the target region, and corrects the defect positioning coordinates based on the positional deviation, and finally generates urban infrastructure detection results based on the defect positioning coordinates. This invention significantly reduces the false detection rate in rainy and foggy weather through multi-channel perturbed sequence generation and dynamic threshold filtering mechanisms; and solves the spatial positioning distortion problem of key facilities such as manhole cover displacement and guardrail tilting based on the reference point mapping technology of the target region with the largest parameter variation, thereby improving subsequent operation and maintenance efficiency. Attached Figure Description

[0045] Figure 1 This is a schematic flowchart of an embodiment of an automated detection method for urban infrastructure based on video image algorithms according to the present invention;

[0046] Figure 2 This is a comparison image of the defect detection area in this invention. Detailed Implementation

[0047] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “corresponding to,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0048] Example 1

[0049] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The following section will describe the automated urban infrastructure detection method based on video image algorithms in this application from a system implementation perspective. Please refer to... Figure 1 The method provided in this application includes the following steps:

[0050] Example 1

[0051] The implementation method in this embodiment can be implemented in a system, on a server, or on a terminal; no specific limitation is made. The method in this application will be described below from the perspective of system implementation. Please refer to... Figure 1 The method provided in this application includes the following steps:

[0052] S1. Acquire real-time video stream data of various types of facilities within the city block, including municipal, greening, drainage, electromechanical, and sanitation facilities;

[0053] This step achieves data acquisition through an integrated acquisition system on a mobile carrier. Specifically, it uses a lightweight vehicle-mounted platform installed on the roof of the inspection vehicle. This platform can include a multispectral high-definition camera group equipped with visible light and infrared dual sensors to continuously acquire video streams at a resolution of 1920×1080 and a rate of 30 frames per second. The infrared band is specifically designed for rainy, foggy weather and low-light environments at night. It can also include an AI edge computing unit to compress the raw video stream in real time and transmit it back to the cloud processing center via a 5G communication module. Furthermore, it can include a human-machine interaction terminal for operators to mark facility types and associate them with GPS coordinates.

[0054] For example, when a vehicle travels at a speed of 20-40 km / h, the system automatically captures dynamic images of facilities within a range of 3-15 meters on both sides of the road. The acquired video stream data includes three core types of information: facility visual data, spatiotemporal reference data, and facility topology metadata. The facility visual data covers the structural surface of municipal guardrails, the surface texture of drainage manhole covers, the vegetation coverage of green belts, the integrity of electromechanical boxes, and the physical form of sanitation facilities. Each frame of the spatiotemporal reference data has a millisecond timestamp and BeiDou / GPS dual-mode positioning coordinates. The facility topology metadata records the spatial relationships between adjacent facilities, such as the positional correlation between manhole covers and drainage ditches.

[0055] S2. Perturb the real-time video stream data to generate multiple sets of perturbed video frame sequences;

[0056] In this embodiment, the multiple sets of perturbation video frame sequences include spatial perturbation frame sequences, color perturbation frame sequences, and noise-resistant perturbation frame sequences. The spatial perturbation frame sequences are generated by randomly applying ±15° rotation, 0.8-1.2x scaling, or ±10% pixel translation transformations to the video frames. The color perturbation frame sequences are generated by adjusting the brightness, contrast, and color saturation parameters of the video frames. The noise-resistant perturbation frame sequences are generated by adding Gaussian noise, salt-and-pepper noise, or simulated rain / fog noise to the video frames.

[0057] By simulating perspective rotation / translation caused by vehicle bumps and scale changes (zoom) at different observation distances, the robustness of detection for spatial deformation scenarios such as tilted manhole covers and deformed guardrails is addressed. Color perturbations are used to reproduce realistic conditions such as day-night brightness differences, contrast changes in sunny and cloudy weather, and fading due to facility aging, avoiding misjudgments caused by seasonal color differences in green facilities and reflections from electromechanical boxes. By injecting multiple types of noise, the model's defect recognition stability is trained under complex conditions such as rainy and foggy weather, camera contamination, and transmission signal distortion, reducing the false negative rate in harsh environments. The parallel generation mechanism of these three types of perturbations expands a single video stream into multi-dimensional training samples, effectively improving the detection probability of defects in small samples.

[0058] S3. Calculate the similarity score between the feature vector of each video frame in multiple perturbation sequences and the weight vector of the pre-trained defect recognition model; filter the video frame sequences according to the preset threshold, delete the video frames with similarity scores lower than the threshold, and generate the updated video frame sequences;

[0059] In this embodiment, the pre-trained defect recognition model is a multi-scale feature fusion network, including: a backbone feature extraction layer, used to extract multi-level spatial features of video frames using the ResNet-50 architecture; an attention enhancement layer, used to superimpose channel attention modules at the output of the backbone feature extraction layer to generate a high-dimensional feature vector with adjusted weights; and a defect classification output layer, used to store learnable weight vectors corresponding to various facility defect types, which are generated through end-to-end training.

[0060] Specifically, the pre-trained defect recognition model achieves defect recognition through transfer learning and multi-task training. First, the backbone network is pre-trained on the ImageNet dataset to learn general feature extraction capabilities. Second, fine-tuning is performed using an urban infrastructure defect dataset to optimize the weight allocation strategy of the channel attention module, focusing the network on key areas of the facility, such as manhole cover edges and guardrail connection points. Finally, end-to-end training generates learnable weight vectors for the defect classification output layer, with each vector corresponding to a feature prototype of a specific defect type, such as crack weight vectors and corrosion weight vectors. During the inference phase, the model suppresses background interference through an attention enhancement layer and performs similarity matching between facility features and weight vectors to achieve pixel-level defect localization.

[0061] In this embodiment, calculating the similarity score includes the following steps:

[0062] S311. Extract the initial feature vector of each frame in the spatial perturbation frame sequence, color perturbation frame sequence, and noise-resistant perturbation frame sequence respectively;

[0063] S312. After fusing the initial feature vectors, input them into the backbone feature extraction layer of the pre-trained defect recognition model to generate high-dimensional semantic feature vectors;

[0064] S313. Calculate the similarity metric between the high-dimensional semantic feature vector and each weight vector in the defect classification output layer, and select the maximum similarity metric as the similarity score of the current video frame.

[0065] Specifically, the initial feature vectors of the three types of perturbation sequences are extracted first: spatial perturbation frame sequence: SIFT key point descriptor (128-dimensional vector) is calculated to capture the deformation features of the facility structure; color perturbation frame sequence: HSV color space histogram (64-dimensional vector) is extracted to quantify color anomalies; noise-resistant perturbation frame sequence: LBP texture feature map (256-dimensional vector) is generated to characterize the degree of surface damage.

[0066] Secondly, feature fusion and transformation are performed. The three types of vectors are reduced to 64 dimensions using PCA. The vectors are then concatenated by channel weighting to generate a 192-dimensional combined feature vector. This vector is then input into the ResNet-50 backbone network and output as a 1024-dimensional high-dimensional semantic feature vector after passing through convolutional layers and attention modules.

[0067] Calculate the cosine similarity between the high-dimensional feature vector and all defect weight vectors, select the highest similarity value as the score of the current frame, and normalize the score range to [0,1], where 1 indicates a complete match with a certain type of defect feature.

[0068] In this embodiment, the generation of the updated video frame sequence includes the following:

[0069] S321. Generate a defect probability threshold based on the environmental state parameters of the current video stream;

[0070] S322. Set differentiated channel filtering conditions for spatial perturbation frame sequences, color perturbation frame sequences, and noise-resistant perturbation frame sequences respectively;

[0071] S323. Based on the channel filtering conditions, delete video frames in each sequence that do not meet the similarity score in parallel, perform spatiotemporal consistency verification on the deleted multi-channel sequences, and retain video frames that meet the cross-channel verification conditions.

[0072] S324. Generate an updated video frame sequence with timestamp alignment based on the verification results.

[0073] Specifically, environmental parameters include light intensity (0-100 klux), rain / fog level (1-5), and facility density (units / 100m). A defect probability threshold is calculated using a linear regression model, where the threshold T = 0.7 - 0.1 × rain / fog level + 0.05 × log(light intensity). Filtering criteria are set as follows: spatial perturbation frame sequence score > T, color perturbation frame sequence score > 0.85T, and noise reduction perturbation frame sequence score > 1.15T. Video frames that do not meet these criteria are then deleted.

[0074] Spatiotemporal consistency verification and sequence generation: 1. Based on the timestamp, check whether the same frame in the three channels is retained; 2. Only retain valid frames that are jointly determined by at least two channels; 3. For missing frames, use the average of the features of adjacent frames to generate compensation frames; 4. Reassemble the frame sequence according to the original time sequence to generate an updated sequence with timestamps.

[0075] S4. Perform defect detection on the updated video frame sequence to identify defect areas of various types of facilities and ancillary facilities;

[0076] In this embodiment, the defective regions to be identified include the following:

[0077] S141. Divide the updated video frame sequence into multi-scale regions to generate a set of facility sub-regions;

[0078] S142. Calculate the defect feature parameters for each facility sub-region. The defect feature parameters include texture complexity, edge breakage, and color anomaly.

[0079] S143. Input the defect feature parameters into the pre-trained defect classifier and output the defect region with labeled contours.

[0080] Specifically, based on the timestamped update sequence generated above, multi-scale region division is first performed. Macro defects are identified at the original scale (256×256 pixels), medium defects are identified at the meso scale (128×128 pixels), and micro defects are detected at the fine scale (64×64 pixels). The zoning strategy is adapted to the facility form, such as using long strip zoning for municipal guardrails and concentric circle zoning for drainage well covers, to generate a set of facility sub-regions.

[0081] Next, defect feature parameters for each sub-region are calculated, including quantifying texture complexity through the contrast variance of the gray-level co-occurrence matrix, calculating edge breakage through the proportion of discontinuous pixels detected by Sobel edge detection, and determining color anomaly in HSV space based on the V channel standard deviation and a preset color gamut tolerance. The defect classifier receives a feature vector composed of texture complexity, edge breakage, and color anomaly, and outputs a defect existence determination and type label through a random forest decision model. When a defect is determined, the coordinates of the minimum bounding polygon are generated, and the type label (such as "structural crack") is bound to the contour and output. A dynamic threshold mechanism is built into the classifier decision boundary adjustment module, which adaptively optimizes the classification sensitivity based on the rain / fog level and light intensity.

[0082] The following example illustrates the defect detection principle using the crack detection of municipal storm drain grates:

[0083] Please see Figure 2 Under a cloudy and slightly hazy environment, only a blurry texture is visible on the surface of a cast iron rain grate. The detection system first divides the input video frame into multi-scale regions. At the micro scale (64×64 pixel grid), the region from the northeast corner of the grate at coordinates (120,90) to (184,154) is determined. By calculating the contrast variance of the gray-level co-occurrence matrix of this sub-region (0.38) and the proportion of non-continuous pixels at the Sobel edge (0.45), the texture roughness and geometric fracture characteristics of the crack source point are quantified. These parameters are input into a random forest defect classifier, which outputs a "structural crack" type label and the corresponding polygonal outline [[125,95],[128,97]...], accurately labeling the 3.5cm long radial crack region.

[0084] S5. Match the defect area with the preset facility template library, divide the template image area and select the target area with the largest parameter variation, calculate the positional deviation of the defect area relative to the reference point of the target area, and correct the defect positioning coordinates based on the positional deviation.

[0085] The defective area is matched with a pre-defined facility template library, the template image regions are divided, and the target region with the greatest parameter variation is selected, including the following:

[0086] S511. Retrieve a standard template image corresponding to the current facility type from the preset facility template library, and divide the standard template image into gridded regions to generate multiple candidate template regions;

[0087] S512. Calculate the parameter variation degree of each candidate template region. The parameter variation degree characterizes the degree of drastic change in pixel features within the current region.

[0088] 1. Extract the gradient magnitude matrix of the candidate template region and calculate the variance of the gradient magnitude matrix as the degree of structural variability;

[0089] 2. Extract the color channel histogram of the candidate template region and calculate the entropy value of the color channel histogram as the degree of color change;

[0090] 3. Weighted fusion of structural variation and color variation is used to generate parameter variation.

[0091] S513. Select the candidate template region with the largest parameter variation as the target region and establish a mapping relationship between the target region and the defect region.

[0092] First, standard template images (such as standard manhole cover images) matching the current defect type are retrieved from a pre-set facility template library. An adaptive grid algorithm is used to divide candidate regions: the grid side length is dynamically calculated based on the facility's physical dimensions (formula: grid side length = min(facility circumscribed rectangle side length) / 5), generating gridded candidate regions with a 20% overlap. Second, the parameter variability of each region is calculated: the gradient magnitude matrix is ​​extracted using the Sobel operator, and the variance is calculated as the structural variability. The region is then converted to HSV space to extract its histogram, and the entropy value is calculated as the color variability. A weighted geometric mean is used to fuse the two types of variability (structural weight 0.7, color weight 0.3). Finally, the candidate regions are sorted in descending order of variability, and high-variability regions are filtered using a dynamic threshold (>mean + 2 × variance). The largest cluster center is selected as the target region through DBSCAN spatial clustering, and a pixel-level mapping relationship between the target region and the defect region is established through affine transformation.

[0093] The corrected defect location coordinates include the following:

[0094] S521. Set a geometric center reference point in the target area;

[0095] S522. Locate the feature point corresponding to the geometric center reference point in the defect area;

[0096] S523. Calculate the two-dimensional Euclidean distance between the feature point and the geometric center reference point as the position deviation. Based on the position deviation, translate the contour coordinates of the defect area to generate the corrected defect location coordinates.

[0097] First, a geometric center reference point is set in the target area: the center of the circumcircle is calculated for circular facilities (such as manhole covers), and the intersection of the diagonals is used for rectangular facilities (such as electrical distribution cabinets); second, the corresponding feature points are located in the defect area: SIFT feature points are extracted and their Hausdorff distances to the reference point are calculated, and the point with the smallest distance is selected as the matching feature point; finally, the two-dimensional Euclidean distance between the feature point and the reference point is calculated as the position deviation, and the coordinates of all vertices of the defect contour are translated according to the deviation (new coordinates = original coordinates + Δd) to generate the corrected positioning coordinates. For example, the manhole cover crack contour is translated from the original coordinates (117, 93) by Δd = (3.2, 1.4) to the precise position (120.2, 94.4), achieving centimeter-level positioning accuracy.

[0098] S6. Generate urban infrastructure inspection results based on defect location coordinates.

[0099] This step encapsulates the corrected defect location coordinates and facility attribute information in a structured manner, generating standardized detection results that include geospatial location, defect type, and severity, and triggering three levels of decision-making instructions: Facility maintenance decision-making instructions: Automatically allocate maintenance resources based on defect type, such as triggering a reinforcement team to respond within 30 minutes if "manhole cover displacement > 3cm"; Safety warning information: Issue dynamic warnings based on location coordinates and road topology, such as simultaneously pushing "avoid and detour" information to the navigation APP within 50 meters of a cracked manhole cover; Smart parking guidance solution: Adjust parking spaces in real time for road sections associated with the defective facility, such as automatically locking parking spaces above displaced manhole covers and guiding vehicles to spare parking spaces within 200 meters.

[0100] The above embodiments reduce the false detection rate in rainy and foggy weather by generating multi-channel perturbation sequences and using dynamic threshold filtering; and solve the spatial positioning distortion problem of key facilities such as manhole cover displacement and guardrail tilt by using benchmark point mapping technology based on the target area with the largest parameter change, thereby improving the efficiency of subsequent operation and maintenance.

[0101] It is understood that those skilled in the art can combine various implementation methods in the above embodiments under the guidance of the above examples to obtain technical solutions with multiple implementation methods.

[0102] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automated detection method for urban infrastructure based on video image algorithms, characterized in that, include: Acquire real-time video stream data of various types of facilities within the urban area, including municipal, greening, drainage, electromechanical, and sanitation facilities; The real-time video stream data is perturbed to generate multiple sets of perturbed video frame sequences; Calculate the similarity score between the feature vector of each video frame in the multiple perturbation sequences and the weight vector of the pre-trained defect recognition model; filter the video frame sequences according to a preset threshold, delete video frames with similarity scores lower than the threshold, and generate an updated video frame sequence; Defect detection is performed on the updated video frame sequence to identify defect areas of the various types of facilities and ancillary facilities; The defect area is matched with a preset facility template library, the template image area is divided and the target area with the largest parameter variation is selected. The positional deviation of the defect area relative to the reference point of the target area is calculated, and the defect positioning coordinates are corrected based on the positional deviation. The urban infrastructure inspection results are generated based on the defect location coordinates.

2. The automated detection method for urban infrastructure based on video image algorithms according to claim 1, characterized in that, The multiple sets of perturbation video frame sequences include spatial perturbation frame sequences, color perturbation frame sequences, and noise-resistant perturbation frame sequences.

3. The automated detection method for urban infrastructure based on video image algorithms according to claim 1, characterized in that, The pre-trained defect recognition model is a multi-scale feature fusion network, including: The backbone feature extraction layer is used to extract multi-level spatial features of video frames using the ResNet-50 architecture; The attention enhancement layer is used to superimpose a channel attention module at the output of the backbone feature extraction layer to generate a high-dimensional feature vector with adjusted weights. The defect classification output layer is used to store learnable weight vectors corresponding to various facility defect types. These learnable weight vectors are generated through end-to-end training.

4. The automated detection method for urban infrastructure based on video image algorithms according to any one of claims 1-3, characterized in that, The calculation of the similarity score between the feature vector of each video frame in the multiple perturbation sequences and the weight vector of the pre-trained defect recognition model includes: The initial feature vectors of each frame in the spatial perturbation frame sequence, color perturbation frame sequence, and noise-resistant perturbation frame sequence are extracted respectively. The initial feature vectors are fused and then input into the backbone feature extraction layer of the pre-trained defect recognition model to generate high-dimensional semantic feature vectors. Calculate the similarity metric between the high-dimensional semantic feature vector and each weight vector in the defect classification output layer, and select the maximum similarity metric as the similarity score of the current video frame.

5. The automated detection method for urban infrastructure based on video image algorithms according to claim 4, characterized in that, The step of filtering video frame sequences according to a preset threshold, deleting video frames with similarity scores lower than the threshold, and generating an updated video frame sequence includes: Generate a defect probability threshold based on the environmental state parameters of the current video stream; Differentiated channel filtering conditions are set for the spatial perturbation frame sequence, color perturbation frame sequence, and noise-resistant perturbation frame sequence, respectively; Based on the channel filtering conditions, video frames with unsatisfactory similarity scores in each sequence are deleted in parallel. Spatiotemporal consistency verification is performed on the deleted multi-channel sequences, and video frames that meet the cross-channel verification conditions are retained. Generate an updated video frame sequence with timestamp alignment based on the verification results.

6. The automated detection method for urban infrastructure based on video image algorithms according to claim 1, characterized in that, The defect detection process for the updated video frame sequence, identifying defective areas of the various types of facilities and ancillary facilities, includes: The updated video frame sequence is divided into multi-scale regions to generate a set of facility sub-regions; Calculate the defect feature parameters for each facility sub-region, including texture complexity, edge breakage, and color anomaly. The defect feature parameters are input into a pre-trained defect classifier, which outputs defect regions with labeled contours.

7. The automated detection method for urban infrastructure based on video image algorithms according to claim 1, characterized in that, The step of matching the defective region with a preset facility template library, dividing the template image region, and selecting the target region with the greatest parameter variation includes: Retrieve standard template images corresponding to the current facility type from the preset facility template library, and divide the standard template images into grid regions to generate multiple candidate template regions; Calculate the parameter variation degree of each candidate template region, whereby the parameter variation degree characterizes the degree of drastic change in pixel features within the current region; The candidate template region with the largest parameter variation is selected as the target region, and a mapping relationship between the target region and the defect region is established.

8. The automated detection method for urban infrastructure based on video image algorithms according to claim 7, characterized in that, The calculation of the parameter variation degree of each candidate template region includes: Extract the gradient magnitude matrix of the candidate template region, and calculate the variance of the gradient magnitude matrix as the degree of structural variability; Extract the color channel histogram of the candidate template region, and calculate the entropy value of the color channel histogram as the color change degree; The parameter variation is generated by weighted fusion of structural variation and color variation.

9. The automated detection method for urban infrastructure based on video image algorithms according to claim 8, characterized in that, The calculation of the positional deviation of the defect area relative to the reference point of the target area, and the correction of the defect location coordinates based on the positional deviation, includes: Set a geometric center reference point in the target area; Locate the feature point in the defect region that corresponds to the geometric center reference point; The two-dimensional Euclidean distance between the feature point and the geometric center reference point is calculated as the positional deviation. The defect area contour coordinates are translated according to the positional deviation to generate the corrected defect location coordinates.

10. The automated detection method for urban infrastructure based on video image algorithms according to claim 1, characterized in that, The urban infrastructure inspection results include facility maintenance decision-making instructions, safety early warning information, and smart parking guidance schemes.

Images