Dangerous road section intelligent early warning method and device based on image acquisition and identification

Abstract

The invention provides a dangerous road section intelligent early warning method and device based on image acquisition and identification, and relates to the technical field of intelligent traffic safety early warning. The method comprises the following steps: acquiring a road image, a bridge image and a tunnel image acquired by image acquisition equipment and an unmanned aerial vehicle; respectively inputting the road image, the bridge image and the tunnel image into corresponding anomaly recognition models to generate anomaly recognition results; generating early warning information based on the abnormal recognition result, and automatically sending the early warning information to a display terminal, a user terminal and a management terminal along the road; and controlling the unmanned aerial vehicle to go to the corresponding abnormal area in response to the fact that the abnormal recognition result meets a preset condition, and performing field audio warning and traffic guidance by using the unmanned aerial vehicle. According to the scheme, the accuracy of road, bridge and tunnel anomaly recognition can be improved, and the response efficiency and the field handling capacity of traffic anomalies are improved.

Description

Background Art

[0002] In the existing field of traffic monitoring and early warning, it mainly relies on fixedly installed monitoring devices or manual inspections. These methods have a certain monitoring ability in fixed monitoring areas. However, due to the limited coverage and fixed monitoring angles, it is difficult to obtain dynamic information in complex traffic environments in a timely manner. Especially in scenarios with limited line of sight or high incidence of emergencies, there are often problems such as delayed detection and response. For example, in the bridge collapse disaster on a certain highway section, affected by continuous heavy rainfall, the bridge structure became unstable and collapsed in a short time, resulting in the interruption of highway traffic and serious secondary accidents. However, before the accident occurred, the background monitoring and management system failed to timely identify the early deformation signs of the bridge structure and did not capture early warning signals such as local settlement of the bridge deck, resulting in the failure of the early warning system to intervene in advance and finally forming a sudden and uncontrollable traffic safety incident.

[0003] With the continuous expansion of the road traffic network and the wide application of structures such as bridges and tunnels, the adaptability of traditional monitoring means in multi-type scenarios is increasingly insufficient, and it is difficult to meet the intelligent early warning requirements of multi-scenarios and real-time. On the other hand, in related road traffic early warning methods, usually the same model is used to process different types of traffic images, and the differences in appearance features, background interference, and abnormal manifestations between roads, bridges, and tunnels are not fully considered, resulting in poor image recognition effects.

[0004] In summary, the existing technology has problems of insufficient ability to identify abnormal states and difficulty in effectively adapting the recognition model to different regional characteristics when facing multi-type traffic scenarios, and it is difficult to meet the requirements of intelligent early warning in complex road environments.

[0005] It should be noted that the information disclosed in the above background art section is only used to enhance the understanding of the background of the present disclosure, and thus may include information that does not constitute the prior art known to those of ordinary skill in the art. Summary of the Invention

[0006] The purpose of the embodiments of the present disclosure is to provide an intelligent early warning method for dangerous road sections based on image acquisition and recognition, an intelligent early warning device for dangerous road sections based on image acquisition and recognition, an electronic device, and a computer-readable storage medium, so as to improve the accuracy of abnormal recognition and regional adaptability in multi-type traffic scenarios and realize intelligent early warning of abnormal states in complex road environments.

[0007] Other features and advantages of the present disclosure will become apparent through the following detailed description, or will be partially learned through the practice of the present disclosure.

[0008] According to the first aspect of the embodiments of the present disclosure, an intelligent early warning method for dangerous road sections based on image acquisition and recognition is provided, including:

[0009] Obtain road images, bridge images, and tunnel images collected by an image acquisition device and a drone;

[0010] Input the road images, bridge images, and tunnel images into corresponding anomaly recognition models respectively to generate anomaly recognition results, where the anomaly recognition models include a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model;

[0011] Generate a warning message based on the anomaly recognition results, and automatically send the warning message to display terminals, user terminals, and management terminals along the road;

[0012] In response to the anomaly recognition results meeting preset conditions, control the drone to fly to the corresponding anomaly area, and use the drone for on-site audio warning and traffic guidance.

[0013] According to a second aspect of the embodiments of the present disclosure, there is provided an intelligent warning device for dangerous sections based on image acquisition and recognition, including:

[0014] An image acquisition module for obtaining road images, bridge images, and tunnel images collected by an image acquisition device and a drone;

[0015] An anomaly recognition module for inputting the road images, bridge images, and tunnel images into corresponding anomaly recognition models respectively based on image sources and image structure features to generate anomaly recognition results;

[0016] A warning sending module for generating a warning message based on the anomaly recognition results and automatically sending the warning message to display terminals, user terminals, and user terminals along the road;

[0017] A drone guidance module for, in response to the anomaly recognition results meeting preset conditions, controlling the drone to fly to the corresponding anomaly area and using the drone for on-site audio warning and traffic guidance.

[0018] According to a third aspect of the embodiments of the present disclosure, there is provided an electronic device, including: a processor; and a memory, where computer-readable instructions are stored on the memory, and when the computer-readable instructions are executed by the processor, the intelligent warning method for dangerous sections based on image acquisition and recognition in the first aspect is implemented.

[0019] According to a fourth aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, the intelligent warning method for dangerous sections based on image acquisition and recognition in the first aspect is implemented.

[0020] The technical solutions provided by the embodiments of the present disclosure may include the following beneficial effects:

[0021] Exemplary embodiments of the present disclosure obtain road images, bridge images, and tunnel images through the cooperation of an image acquisition device and a drone, and input them into a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model for classification processing respectively. On the one hand, compared with the method of using a single model to recognize images of different traffic scenarios in the related art, it is possible to match the corresponding recognition model according to the structural attributes of the images, so as to better adapt to the visual characteristics of different regions during the recognition process, thereby reducing misjudgment or missed detection caused by insufficient model generalization. On the other hand, by combining the image acquisition device and the drone, compared with relying solely on the device to acquire images, it is possible to expand the spatial range of image acquisition, especially in areas with limited perspectives or difficult facility layout, providing a more flexible means of image acquisition. In addition, warning information is generated based on the recognition results and distributed through display terminals set along the road and terminals for users, which improves the transmission efficiency and usability of the warning information to a certain extent. Further, when the recognition result meets the preset conditions, the drone can be controlled to go to the relevant area, and the abnormal situation can be prompted by voice broadcast or other means to cooperate with the on-site traffic guidance, which helps to achieve a more direct warning response. Therefore, the technical solution in the present disclosure can improve the accuracy of anomaly recognition and regional adaptability in multi-type traffic scenarios, and realize intelligent warning of abnormal states in complex road environments.

[0022] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings herein are incorporated into the specification and constitute a part of this specification, showing embodiments consistent with the present disclosure, and are used together with the specification to explain the principles of the present disclosure. Obviously, the accompanying drawings in the following description are only some embodiments of the present disclosure, and those of ordinary skill in the art can obtain other drawings based on these drawings without creative efforts.

[0024] Figure 1 Schematically shows a flowchart of an intelligent warning method for dangerous sections based on image acquisition and recognition according to some embodiments of the present disclosure.

[0025] Figure 2 Schematically shows a flowchart of generating a road anomaly recognition result according to some embodiments of the present disclosure.

[0026] Figure 3 Schematically shows a flowchart of generating a bridge anomaly recognition result according to some embodiments of the present disclosure.

[0027] Figure 4Schematically shows a flowchart of generating tunnel anomaly recognition results according to some embodiments of the present disclosure.

[0028] Figure 5 Schematically shows a schematic diagram of an intelligent warning device for dangerous sections based on image acquisition and recognition according to some embodiments of the present disclosure.

[0029] Figure 6 Schematically shows a schematic diagram of a computer system of an electronic device according to some embodiments of the present disclosure.

[0030] Figure 7 Schematically shows a schematic diagram of a computer-readable storage medium according to some embodiments of the present disclosure.

[0031] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts. Detailed Description of the Invention

[0032] Here, exemplary embodiments will be described in detail, and examples thereof are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. On the contrary, they are merely examples of devices and methods consistent with some aspects of this specification as detailed in the appended claims.

[0033] In addition, the drawings are only schematic diagrams and are not necessarily drawn to scale. The block diagrams shown in the drawings are only functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor devices and / or microcontroller devices.

[0034] In the present exemplary embodiment, first, an intelligent warning method for dangerous sections based on image acquisition and recognition is provided. The intelligent warning method for dangerous sections based on image acquisition and recognition can be applied to a terminal device. For example, the terminal device can be various electronic devices with an image processing unit, including but not limited to desktop computers, portable computers, smartphones, tablets, etc. In addition, the display terminals along the road can be LED displays, traffic lights, liquid crystal bulletin boards or other forms of visual output devices.

[0035] Figure 1 Schematically shows a flowchart of an intelligent warning method for dangerous sections based on image acquisition and recognition according to some embodiments of the present disclosure. Refer to Figure 1As shown, the intelligent warning method for dangerous sections based on image acquisition and recognition may include the following steps:

[0036] Step S110, obtaining road images, bridge images, and tunnel images collected by an image acquisition device and a drone;

[0037] Step S120, respectively inputting the road images, bridge images, and tunnel images into corresponding anomaly recognition models to generate anomaly recognition results, where the anomaly recognition models include a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model;

[0038] Step S130, generating a warning message based on the anomaly recognition result and automatically sending the warning message to display terminals and user terminals along the road;

[0039] Step S140, in response to the anomaly recognition result meeting a preset condition, controlling the drone to fly to the corresponding anomaly area and using the drone for on-site audio warning and traffic guidance.

[0040] In the intelligent warning method for dangerous sections based on image acquisition and recognition in the above embodiment, road images, bridge images, and tunnel images are obtained through the cooperation of an image acquisition device and a drone, and are respectively input into a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model for classification processing. On the one hand, compared with the method of using a single model to identify images of different traffic scenarios in the related art, it can match the corresponding recognition model according to the structural attributes of the images, so as to better adapt to the visual characteristics of different regions during the recognition process, thereby reducing misjudgment or missed detection caused by insufficient model generalization. On the other hand, by combining the image acquisition device and the drone, compared with relying only on the device to collect images, it can expand the spatial range of image acquisition, especially in some areas with limited perspectives or difficult facility layout, providing a more flexible means of image acquisition. In addition, generating a warning message based on the recognition result and distributing the information through display terminals set along the road and terminals for users improves the transmission efficiency and convenience of use of the warning message to a certain extent. Further, when the recognition result meets the preset condition, the drone can be controlled to fly to the relevant area and prompt the abnormal situation through voice broadcast, etc., to cooperate with the on-site traffic guidance, which helps to achieve a more direct warning response. Therefore, the above embodiment can improve the accuracy of anomaly recognition and regional adaptability in multiple types of traffic scenarios, and realize the intelligent warning of abnormal states in complex road environments.

[0041] Next, the intelligent warning method for dangerous sections based on image acquisition and recognition in this exemplary embodiment will be further described.

[0042] Step S110: Obtain road images, bridge images, and tunnel images collected by an image acquisition device and a drone.

[0043] Among them, the image acquisition device can represent a device for obtaining images of road, bridge, or tunnel scenes. In a specific implementation, image data of corresponding areas is collected by an image acquisition device set in the target area of a road, bridge, or tunnel. Among them, the image acquisition device can include a camera with an image acquisition function or a network video acquisition terminal, which can continuously or periodically collect visual information in a set monitoring area to obtain an image frame sequence. At the same time, through the image acquisition component configured on the drone, aerial image data of the road surface, bridge structure, or tunnel interior is obtained during the process of the drone performing an inspection task according to a preset flight path and acquisition frequency. During the acquisition process, the acquisition angle, resolution, and image acquisition interval can be dynamically adjusted in combination with the structural characteristics, traffic conditions, and environmental brightness conditions of the target area to ensure the effectiveness and clarity of the image data.

[0044] In addition, during the process of obtaining road images, bridge images, and tunnel images, the position information of the corresponding images can also be obtained synchronously. This position information can be provided by the positioning module configured on the image acquisition device or the drone, and is used to record the geographic coordinate data at the time of each frame of image acquisition. Exemplarily, the positioning module can be a global satellite navigation system positioning unit or a geographic information system interface module, etc. In addition, the position information can also be time-stamped and spatially bound with the image data to form an image data set with position information annotations, so as to be used in subsequent steps such as spatial positioning of abnormal recognition results, determination of early warning push areas, and drone navigation path planning, thereby improving the traceability of image data during the abnormal recognition and processing process.

[0045] Step S120: Input the road images, bridge images, and tunnel images into corresponding abnormal recognition models respectively to generate abnormal recognition results, where the abnormal recognition models include a road abnormal recognition model, a bridge abnormal recognition model, and a tunnel abnormal model.

[0046] Among them, the abnormal recognition result can represent the recognition output data obtained after the abnormal recognition model processes the input image. Its specific content can include the determination information on whether there is an abnormal state in the image, the type label of the abnormality, the confidence score, etc., and is used to characterize the potential risks or abnormal states related to traffic passing in the input image. The road abnormal recognition model can represent an image recognition model constructed for road images, used to recognize the abnormal states in the road area, classify and recognize road abnormal states such as road surface damage, obstacles, water accumulation, collapse, traffic accidents, guardrail loss, etc., and output the abnormal recognition result corresponding to the road image. The bridge abnormal recognition model can represent an image recognition model constructed for bridge images, used to recognize traffic abnormal states related to bridges such as bridge deck collapse, guardrail loss, main girder cracks, bearing settlement, deformation, displacement, etc. The tunnel abnormal model can represent an image recognition model constructed for tunnel images, used to recognize the abnormal states in the internal area of the tunnel that may affect traffic safety, and the abnormal states can include wall leakage, cracks, congestion, lighting abnormalities, traffic accidents in the tunnel, etc.

[0047] By inputting road images, bridge images, and tunnel images into the corresponding abnormal recognition models respectively to generate abnormal recognition results, it is possible to match dedicated recognition models for the image features of different traffic structure types, realize the corresponding adaptation of the model structure and image content during the recognition process, thereby improving the recognition accuracy of abnormal states in various traffic scenarios, reducing false alarms and missed alarms caused by insufficient model generalization, and helping to enhance the reliability and practicality of the system in multi-scenario traffic early warning.

[0048] Step S130, generate a warning message based on the abnormal recognition result, and automatically send the warning message to the display terminals, user terminals, and management terminals along the road.

[0049] Among them, the early warning information can represent the prompt content generated according to the abnormal recognition result and used to indicate the potential traffic risks in the target area. Its content can include fields such as abnormal type, abnormal location, image source, recognition time, risk level, etc., and can also include text information or multimedia content for display or voice broadcast. The display terminal can represent the device with information display function set along the road, bridge or tunnel, and can specifically include LED display screen, traffic indicator light, liquid crystal bulletin board or other forms of visual output devices, which are used to receive and display the early warning information to realize the risk prompt or traffic guidance for on-site traffic personnel. The user terminal can represent a mobile device or equipment with information receiving and displaying capabilities, and can specifically include a smart phone, a tablet computer, an in-vehicle information system installed with an early warning information receiving program or a display interface accessing a third-party application, which is used to display the early warning information related to the area of concern or the traffic path of the user to the user. The management terminal can represent the control device for receiving, displaying and processing the early warning information in the road monitoring system. The management terminal can be a monitoring platform of the traffic management center, an accident response dispatching platform or a command and decision-making terminal, etc., which has the data interaction ability with the front-end data acquisition device and the back-end task execution module, and supports the classification management of early warning information, the issuance of dispatching instructions and the visual control of the response process.

[0050] Sending the early warning information to the display terminals and user terminals along the road can realize the joint response of the abnormal recognition result and the information prompt. On the one hand, it helps to display the identified potential risks visually at the scene in the first time, improving the understanding of the abnormal situation by on-site traffic participants. On the other hand, through the distribution of early warning information by the user terminal, remote prompts for individual users can be realized, expanding the information reach range, reducing the response lag problem caused by information transmission delay or insufficient coverage, and thus enhancing the real-time performance of the early warning method and the traffic safety guarantee ability.

[0051] Step S140, in response to the abnormal recognition result meeting the preset conditions, control the unmanned aerial vehicle to go to the corresponding abnormal area, and use the unmanned aerial vehicle for on-site audio warning and traffic guidance.

[0052] Among them, the preset conditions can represent a set of judgment rules for determining whether an abnormal recognition result needs to trigger a response operation, and specifically can include one or more judgment rules such as abnormal type, risk level, abnormal area attribute, recognition confidence threshold, duration, etc. The audio warning can represent a voice broadcast emitted by an audio playback device carried by a drone to prompt on-site traffic participants. The content of the voice broadcast can include abnormal type announcements, risk area reminders, avoidance path suggestions, emergency evacuation instructions, etc., for enhancing the on-site reminder effect of abnormal information. Traffic guidance can represent the behavior of a drone to give direction reminders, path adjustment reminders, or area avoidance reminders to traffic participants in the area where an abnormality occurs or a designated warning area through audio broadcasts and light signals.

[0053] Using a drone for on-site audio warning and traffic guidance can achieve automated response and localized reminder after identifying a risky traffic state. On the one hand, by having the drone autonomously fly to the abnormal location for audio broadcasting, it can improve the dissemination efficiency of abnormal information on-site and enhance the immediate perception ability of traffic participants. On the other hand, combined with traffic guidance actions, it can assist in the temporary adjustment and evacuation of on-site traffic order, thereby enhancing the emergency response ability and risk intervention effect in traffic abnormal scenarios.

[0054] Next, the content in steps S110 to S140 will be described in detail.

[0055] In some embodiments, obtaining road images, bridge images, and tunnel images collected by an image acquisition device and a drone specifically includes the following technical steps: collecting bridge images through an image acquisition device pre-set on the bridge structure, where the bridge images include any one or more of bridge substructure images, bridge flexible structure images, bridge deck images, main beam images, vehicle images, pedestrian images, and bearing images; collecting tunnel images through an image acquisition device pre-set on the tunnel structure, where the tunnel images include any one or more of tunnel wall images, vault images, sidewall images, tunnel road surface images, cable trench images, fan images, vehicle images, lighting equipment images, and drainage structure images; obtaining road images through the drone at preset time intervals and inspection paths, where the road images include any one or more of road surface images, slope images, drainage facility images, and guardrail images; when the image acquisition device fails to complete the image acquisition of the corresponding area, collecting the bridge images or tunnel images of that area through the drone.

[0056] Among them, the image acquisition device can represent a device installed on a bridge or tunnel structure for image acquisition of a specified area, such as a fixed monitoring camera, a slide rail monitoring camera, etc. The image of the lower structure of the bridge can represent the image data covering the area below the bridge load-bearing system, including but not limited to the appearance images of piers, abutment caps, foundations, bent caps, etc. The image of the flexible structure of the bridge can represent the image data collected for the flexible components of the upper part of the bridge, such as suspension cables, tie rods, hanging cables, expansion joints, bridge deck paving, etc. The image of the bridge deck can represent the image data used to characterize the state of the bridge passage surface. The image of the main girder can represent the image data used to characterize the state of the main girder structure of the bridge. Through this image, structural problems such as surface cracks, erosion, pollution, and abnormal connections of the main girder can be identified. The image of the bearing can represent the image data used to characterize the state of the bearing part of the bridge. The image of the tunnel wall can represent the image data used to characterize the state of the tunnel sidewall area, which can include the lining structure or sprayed surface layer on the left and right sides of the tunnel. The image of the vault can represent the image data used to characterize the area of the tunnel top structure, which can be used to identify problems such as vault sagging, falling risk, abnormal equipment suspension, or lighting failure. The image of the cable trench can represent the image data collected by the image acquisition device for the cable trench or cable tray parts arranged along the tunnel structure, used to identify abnormal states such as cable trench deformation or cable exposure. The image of the fan can represent the image data obtained by shooting the fan device installed in the tunnel ventilation system by the image acquisition device, used to identify abnormal conditions such as ventilation equipment failure or bracket loosening. The image of the lighting equipment can represent the image data obtained by collecting the lighting fixtures installed in the tunnel and their supporting installation parts by the image acquisition device, used to identify abnormal phenomena such as lamp falling off or lighting failure. The image of the drainage structure can represent the image data used to characterize the state of the tunnel drainage system. The image of the slope can represent the image data obtained by shooting the slope areas on both sides of the road along the line by the image acquisition device, used to identify abnormalities such as slope cracks, slips, and local collapses. The image of the drainage facilities can represent the image data collected by the image acquisition device for the relevant parts of the road drainage system, which can include structures such as side ditches, drain holes, catch basins, or open ditches, used to monitor the smoothness of the drainage path, the integrity of the structure, and abnormal conditions such as siltation and blockage.

[0057] In specific implementations, the bridge deck, main girder, and bearing areas are collected by image acquisition devices pre - installed on the bridge structure. Among them, the image acquisition angle can be set as a directly - downward view angle, an oblique - side view angle, or a horizontal view angle according to the bridge structure form and monitoring requirements, so as to obtain bridge - deck images, main - girder images, and bearing images. The image acquisition method can be periodic snapshot or continuous video - stream mode, and the acquisition task is executed in combination with a preset time interval and trigger strategy. At the same time, tunnel images are obtained by image acquisition devices installed on both sides or the top of the tunnel structure, and the image content covers tunnel - wall images, vault images, and drainage - structure images. For different tunnel structures, the position orientation and exposure parameters of the image acquisition devices can be adjusted to ensure the clarity and stability of image data under different lighting and traffic environments. In addition, the drones equipped with image acquisition components perform regular flight tasks in the road area according to the preset time interval and inspection path, and road images are obtained during the flight. The image acquisition height, view angle, and resolution parameters can be set according to requirements such as road width and recognition accuracy.

[0058] When the image acquisition devices in the bridge or tunnel structure cannot complete the image acquisition task at a specific time period, in a specific area, or from a specific view angle, the drone can be controlled to go to this area to replace and complete the supplementary acquisition of bridge images or tunnel images to ensure the integrity and coverage continuity of the image data. The situations where the image acquisition devices cannot complete the acquisition task can include: the image quality deteriorates due to insufficient ambient light, strong - light interference, or rain and fog during a specific time period; there are monitoring blind spots in a specific area due to limited installation angles of the devices or being blocked by the structure; it is impossible to obtain effective images due to complex structure forms or limited observation directions from a specific view angle; and the image acquisition function fails due to factors such as stain coverage, lens damage, network anomalies, or power outages of the devices themselves, etc.

[0059] In some embodiments, the road images, bridge images, and tunnel images are respectively input into corresponding anomaly - recognition models to generate anomaly - recognition results, which specifically include the following technical steps: the road images are input into a road - anomaly recognition model constructed by a lightweight convolutional neural network and a spatial pyramid pooling structure to generate road - anomaly recognition results; the bridge images are input into a bridge - anomaly recognition model constructed by a multi - layer residual structure to generate bridge - anomaly recognition results; the tunnel images are input into a tunnel - anomaly recognition model constructed by an encoder - decoder semantic segmentation network to generate tunnel - anomaly recognition results; in response to the bridge - anomaly recognition result being cracks, deformation, and displacement, or the tunnel - anomaly recognition result being cracks and water seepage, anomaly - time - series data is constructed based on the historical image data and the current recognition results of the corresponding anomaly areas; the anomaly - time - series data is input into a pre - trained prediction model for trend analysis to evaluate the change trends of crack propagation, structural deformation amounts, and displacement trajectories.

[0060] Among them, the lightweight convolutional neural network can represent a convolutional neural network model with optimized structure and low computational cost, which has the characteristics of small number of parameters and high operation efficiency. The spatial pyramid pooling structure can represent a structure that performs pooling operations on feature maps at different scales and concatenates the pooling results at multiple scales, which can be used to enhance the model's perception ability of image targets of different sizes. The multi-layer residual structure can represent a feature extraction network formed by stacking multiple residual units, and each residual unit contains a main branch and a skip connection. The encoder-decoder semantic segmentation network can represent an image segmentation network that includes an encoding path and a decoding path. The encoding path is used to extract features and compress the image, and the decoding path is used to restore the spatial dimensions and output pixel-level classification results.

[0061] In this embodiment, the abnormal targets in road images usually have uneven scale distributions, large morphological differences, and a large acquisition volume. To ensure processing efficiency and take into account recognition accuracy, a lightweight convolutional neural network is used as the basic feature extraction backbone to improve the adaptability of the model in edge computing or resource-constrained scenarios. At the same time, a spatial pyramid pooling structure is introduced to perform feature fusion on multi-scale regions, which can enhance the model's comprehensive perception ability of local details and context relationships and improve the detection effect of road anomalies such as road surface damage, obstacles, water accumulation, subsidence, and traffic accidents. Bridge images usually contain multiple types of structural components, such as bridge decks, main girders, and bearings. The image content is complex and bridge anomalies mostly appear in the form of structural defects or occlusions. To enhance the extraction depth and stability of the model for structural features, a multi-layer residual structure is used to construct a bridge anomaly recognition model. The residual structure improves the feature transfer efficiency through cross-layer connections, which can effectively alleviate the gradient vanishing problem in deep networks and enhance the recognition ability of various bridge anomalies such as bridge deck subsidence, guardrail loss, main girder cracks, and bearing settlement in bridge images. Tunnel images have obvious regional boundary characteristics and require pixel-level regional division of the images. To achieve fine-grained semantic region recognition, an encoder-decoder semantic segmentation network is used to construct a tunnel anomaly recognition model. This network can extract regional features at different semantic levels, thereby achieving accurate recognition of tunnel anomalies such as wall leakage, congestion, lighting anomalies, traffic accidents in the tunnel, and overflow water accumulation.

[0062] In addition, in response to the situation where the bridge anomaly recognition result is a crack, deformation, or displacement, or the tunnel anomaly recognition result is a crack and water seepage, abnormal time-series data is constructed based on the historical image data of the abnormal area and the current recognition result. Specifically, the feature parameters at the same spatial position in multiple historical image frames of the corresponding structural area are extracted. These feature parameters can include crack length, width, opening angle, local contour change amount, water seepage area, and the central displacement value of the deformation area, and are aligned with the same type of recognition parameters in the current image frame to form abnormal time-series data with a time dimension. where x tdenotes the abnormal feature vector extracted at the t-th time point, T represents the current time series length, and x t may include features such as crack width, length, deformation amount, seepage area, etc.

[0063] Input the abnormal time series data into the pre-trained prediction model for trend analysis. This prediction model can be constructed using a long short-term memory network, which is used to simulate the dynamic evolution process of abnormal indicators in the time dimension and output the prediction results at future time points The LSTM unit can be represented by the following state transition formula:

[0064] h t = LSTM(x t , h t-1 ; θ)

[0065] where h t represents the hidden state output at the t-th moment, θ represents the set of model parameters, x t represents the abnormal feature input at the current moment, and h t-1 is the state at the previous moment.

[0066] After obtaining the prediction result , compare the current recognition value with the change amplitude of the predicted trend to determine whether the crack continues to extend, whether the structural deformation amount is in an increasing trend, whether there is a trend of increasing displacement in the target area, or the area change value of the seepage area in the tunnel. The above prediction results can be used to construct an abnormal evolution trend curve and provide a judgment basis for subsequent early warning level assessment and abnormal handling.

[0067] In some embodiments, as shown in Figure 2 , inputting the road image into the road anomaly recognition model constructed by the lightweight convolutional neural network and the spatial pyramid pooling structure to generate the road anomaly recognition result specifically includes the following technical steps:

[0068] Step S210, perform size adjustment and pixel value normalization processing on the road image to obtain preprocessed image data, and input the preprocessed image data into a convolutional network structure composed of multiple depthwise separable convolutional units to extract corresponding multi-scale image features layer by layer to generate a first feature data set.

[0069] Specifically, perform size adjustment and pixel value normalization processing on the original road image I to obtain preprocessed image data where I(x, y, c) represents the pixel value at the horizontal position x, vertical position y, and channel index c in the image, μ c represents the average value of channel c, and σ c represents the standard deviation of channel c. The process of normalization processing can be expressed as:

[0070]

[0071] Input the standardized image into a convolutional network structure containing multiple depthwise separable convolution units, perform a layer-by-layer feature mapping operation, extract the spatial features of the image at different semantic levels, and generate the first feature data set F1. Among them, F1(x′, y′, d1) represents the first type of image feature at the position (x′, y′) and the channel index d1.

[0072] Step S220: Perform a multi-scale pooling operation on the first feature data set according to a set spatial pyramid structure, extract local statistical features at different scales, and generate a second feature data set.

[0073] Specifically, according to the set spatial pyramid pooling structure, perform a multi-scale pooling operation on the first feature data set F1. Let the scale set be S = {s1, s2, …, s k}, where s i represents the i-th pooling scale. For each scale s i , perform a pooling operation and output the local statistical features at this scale Concatenate the pooling features at all scales to obtain the second feature data set F2. The concatenation process can be expressed as:

[0074]

[0075] Among them, represents concatenation along the channel dimension, represents the pooling output feature at scale s i , and k represents the number of pooling scale levels set in the spatial pyramid structure.

[0076] Step S230: Concatenate the first feature data set and the second feature data set along the channel dimension to generate an enhanced feature data set, and the enhanced feature data set has the same spatial size as the first feature data set.

[0077] Specifically, perform a fusion process on the first feature data set F1 and the second feature data set F2 to generate an enhanced feature data set F E . To achieve spatial alignment, extend F2 to the same spatial dimension as F1 through a replication operation to obtain an extended feature The fusion operation process can be expressed as:

[0078]

[0079] Among them, represents the extended second - type feature data, d2 represents the new channel index after splicing, and F E is the finally enhanced image feature representation.

[0080] Step S240: Input the enhanced feature data set into the recognition sub - network with a fully - connected structure and a classification function, perform feature flattening and class classification processing, and output the abnormal class label corresponding to the road image.

[0081] Specifically, input the enhanced feature data set F E into the recognition sub - network including a flattening structure, a fully - connected layer, and a classification function. First, perform flattening processing on F E to obtain the feature vector v, where v(i) represents the i - th element in the flattened vector. Then input the feature vector v into the linear classifier to output the abnormal class label y. The classification function is expressed as:

[0082]

[0083] where y represents the recognized abnormal class label, j represents the predefined class index, w j represents the weight vector corresponding to class j, b j represents the bias term, represents the inner product of the feature vector and the weight. In this embodiment, the abnormal class label y can be a classification identifier used to characterize the abnormal type in the road image. The specific form can include but is not limited to preset abnormal state types such as road surface damage, obstacles, water accumulation, collapse, traffic accidents, etc. Each class label corresponds to a unique class index j, which is used to identify the belonging type of the recognition result in the predefined abnormal class set.

[0084] In some embodiments, as shown in Figure 3 the input of the bridge image into the bridge abnormal recognition model constructed by a multi - layer residual structure to generate the bridge abnormal recognition result specifically includes the following technical steps:

[0085] Step S310: Perform regional separation on the bridge image through a segmentation algorithm based on edge gradient and spatial texture features, and extract the structural region images including the bridge sub - structure, bridge flexible structure, main girder, bridge deck, and bearing.

[0086] Specifically, perform regional separation processing on the bridge image I B and adopt an image segmentation method that fuses edge gradient and spatial texture features to extract the bridge structural region. The edge response intensity is defined as:

[0087]

[0088] where I B(x, y) represents the pixel value at the position (x, y) in the bridge image. and represent the horizontal and vertical gradients of the bridge image at (x, y) respectively. The texture response function is defined as:

[0089]

[0090] where f m (x, y) represents the response value of the m-th texture channel at the pixel point (x, y), and α m is the weighting coefficient of this channel, and M is the total number of texture feature channels. The joint scoring function is expressed as:

[0091] S B (x, y) = λ1G B (x, y) + λ2T B (x, y)

[0092] where S B (x, y) represents the structural area extraction scoring value, and λ1 and λ2 are weight parameters. Based on S B (x, y), the images of the main girder, bridge deck, and bearing areas are extracted.

[0093] Step S320: Input the structural area image into a feature extraction network with a multi-layer residual structure, extract local edge features, joint features, and deformation contour features in different channels respectively, and generate multi-channel residual feature data.

[0094] Specifically, take the structural area image as the input R B , and input it into a feature extraction network with a multi-layer residual structure to extract corresponding structural features in different channels. The output features of the residual structure are defined as:

[0095]

[0096] where are the weights of the two-layer convolution kernels respectively, is the corresponding bias term, φ(·) represents the activation function, * represents the two-dimensional convolution operation, and F B represents the extracted multi-channel residual feature data. There are three channels in this multi-channel residual feature data. One channel represents the edge feature F B1 , another channel represents the joint feature F B2 , and there is also a channel representing the contour deformation feature F B3 .

[0097] Step S330: Perform cross-channel feature comparison operation on the multi-channel residual feature data, obtain the differential feature data between different structural channels through feature difference calculation, and fuse the differential feature data with the original channel features in the multi-channel residual feature data to generate enhanced differential features.

[0098] Specifically, perform pairwise comparison on the three-channel residual features, and extract the differential features between structures. The differential features are defined as:

[0099] Δ Bij (x,y) = |F Bi (x,y) - F Bj (x,y)|

[0100] where F Bi (x,y) and F Bj (x,y) respectively represent the feature response values of channels i and j at the pixel point (x,y), and Δ Bij (x,y) represents the pixel-level differential response between the two. By concatenating all the original channel features and the differential channel features, an enhanced differential feature map is obtained:

[0101]

[0102] where Concat(·) represents the concatenation operation in the channel dimension, represents the final bridge enhancement feature map for recognition. F B1 can represent the edge feature map extracted from the bridge area image through the first channel in the multi-layer residual structure, F B2 can represent the joint feature map extracted from the bridge area image through the second channel in the residual structure, F B3 can represent the contour deformation feature map extracted from the bridge area image through the third channel in the residual structure, Δ B12 can represent the pixel-by-pixel difference map between the edge feature map F B1 and the joint feature map F B2 , Δ B13 can represent the pixel-by-pixel difference map between the edge feature map F B1 and the contour deformation feature map F B3 , Δ B23 can represent the pixel-by-pixel difference map between the joint feature map F B2 and the contour deformation feature map F B3 .

[0103] Step S340: Input the enhanced differential features into the recognition sub-network including a fully connected structure and a pattern matching classification function, and output the abnormal category label corresponding to the bridge image.

[0104] Specifically, the enhanced difference feature map is flattened into a feature vector v B , and is input into an identification sub-network including a fully connected structure and a pattern matching classification function. The classification function is defined as:

[0105]

[0106] where y B represents the identification result label of the bridge image, v B represents the flattened feature vector, is the weight vector corresponding to the j-th type of bridge anomaly, is the corresponding bias term, and j is the index of the predefined bridge anomaly category. In this embodiment, the anomaly category label y B corresponding to the bridge image may include deck collapse, guardrail missing, main beam crack, bearing settlement, etc.

[0107] In some embodiments, as shown in Figure 4 , inputting the tunnel image into a tunnel anomaly identification model constructed by an encoder-decoder semantic segmentation network to generate a tunnel anomaly identification result specifically includes the following technical steps:

[0108] Step S410: Input the tunnel image into a semantic segmentation network including a multi-level encoding layer and a decoding layer structure, perform pixel-level semantic segmentation on each region in the tunnel image, and generate a segmentation label image including a wall region, a vault region, and a pipeline region.

[0109] Specifically, input the tunnel image I T into a semantic segmentation network including a multi-level encoding layer and a decoding layer structure, and perform pixel-level semantic classification on different structural regions in the image. The encoding path for extracting semantic features can be defined as:

[0110]

[0111] where I T (x, y) represents the input value of the pixel point (x, y) in the tunnel image, E(·) represents the encoding operation, ψ(·) represents the downsampling process, represents the encoded semantic feature map. The decoding path restores the high-dimensional features to the original spatial dimension through skip connections and outputs the segmentation label image:

[0112]

[0113] where P T (x, y, c) represents the probability of predicting the c-th type of structure (such as wall, vault, pipeline) at the pixel point (x, y), and L T (x, y) is the final semantic segmentation label image.

[0114] Step S420: Based on the tunnel wall region and vault region marked in the segmented label image, extract the segmentation probability distribution of the corresponding pixels, and perform a significance analysis on the segmentation probability distribution to obtain an anomaly response image.

[0115] Specifically, on the basis of the segmentation label map L T , extract the segmentation probability distributions corresponding to the wall and vault regions, and perform a significance analysis. Let the c1 class be the wall and the c2 class be the vault, and their probability mappings are:

[0116] Q T (x,y) = max(P T (x,y,c1), P T (x,y,c2))

[0117] where Q T (x,y) represents the maximum structure probability value within the region of interest structure. Input this probability map into the significance function

[0118]

[0119] where represents the mean value of all Q T (x,y), and S T (x,y) is the significance response map, indicating the degree of structural confidence deviation of this pixel point.

[0120] Step S430: Extract the regions with significance higher than the set threshold in the anomaly response image as candidate anomaly regions, and extract the contour curvature and gray gradient features for each candidate anomaly region to obtain a set of region feature vectors.

[0121] Specifically, according to the significance response map S T (x,y), take the regions greater than the threshold τ as candidate anomaly regions. The definition of the region contour curvature feature extraction is:

[0122]

[0123] where B(i) represents the coordinate of the i-th boundary point on the contour curve, s is the curve arc length, and N is the number of boundary points. The gray gradient feature calculation is:

[0124]

[0125] where (x j ,y j ) represents the position of the j-th pixel point within the anomaly region, is the image gradient vector of this point, and M is the total number of pixels within the region. Finally, CT Combined with G T to form the region feature vector v T =[C T , G T .

[0126] Step S440: Input the set of region feature vectors into a classification network including a feature aggregation structure and a discriminant function, and output an anomaly class label corresponding to the tunnel image.

[0127] Specifically, input the region feature vector v T into a classification network including a feature aggregation structure and a discriminant function, and output a tunnel anomaly class label. The classification function is defined as:

[0128]

[0129] where y T represents the anomaly class label corresponding to the tunnel image, represents the weight vector of the k-th type of anomaly, is the bias, and k is the class index number. In this embodiment, the anomaly class label y T corresponding to the tunnel image may include wall leakage, congestion, lighting anomaly, traffic accidents in the tunnel, overflow and waterlogging, etc.

[0130] In some embodiments, generating a warning message based on the anomaly recognition result and sending the warning message to a display terminal and a user terminal along the road specifically includes the following technical steps:

[0131] First, perform data parsing on the anomaly recognition result, extract the anomaly class, anomaly location, and structure type fields, and generate a set of original warning contents. Among them, the anomaly location field may represent identification data for describing the geographical location of an anomaly event in the traffic structure space. The structure type field may represent identification data for indicating the traffic structure category corresponding to the anomaly recognition result, including but not limited to roads, bridges, and tunnels. In addition, the structure type field may also be further divided into multi-level structure labels such as "drone-road image" and "camera-bridge image" according to the specific detection image source device.

[0132] In specific implementation, obtain the anomaly class field from the anomaly recognition result to represent the corresponding structural anomaly type; extract the anomaly location field, including geographical coordinates, road number, or structure segment information, for marking the anomaly occurrence location; parse the structure type field to indicate the traffic structure category to which the anomaly belongs, including roads, bridges, or tunnels. Combine the above fields to form a set of original warning contents for subsequent warning message generation and push processing.

[0133] Then, integrate the original warning content set with road numbers, road locations, timestamps, and image recognition sources to construct a warning information data structure with location fields and structure fields.

[0134] Specifically, first extract the abnormal category field, abnormal location field, and structure type field from the original warning content set. Secondly, introduce the road number corresponding to the abnormal location field as the identification information of the road or structural unit. Further record the time information corresponding to the abnormal recognition to generate a timestamp field; and combine the image acquisition path or recognition model source to extract the image recognition source field, which is used to indicate the source type of the warning data. Organize the above fields according to a predetermined data structure, where the location field consists of the abnormal location field and the road number field, which is used to identify the specific spatial location where the abnormality occurs; the structure field consists of the structure type field and the image recognition source field, which is used to describe the traffic structure type to which the abnormality belongs and the recognition source.

[0135] Next, based on the location field in the warning information data structure, match the road-side display terminals corresponding to the abnormal location, and combine the user subscription strategy to identify the user terminals that meet the conditions, generating a set of terminal push objects.

[0136] Among them, the user subscription strategy can represent the warning reception preference configuration preset by the user in the client or management platform, which is used to screen the warning information content that the user is interested in. Specifically, the user subscription strategy can include one or more conditional fields, such as: target road section or area range, structure types of interest (such as bridges, tunnels, etc.), warning level thresholds, receiving time periods, etc. Rule matching can be performed based on the fields in the warning information data structure and the user subscription strategy to determine whether to push the current warning to the corresponding user terminal. The set of terminal push objects can represent a set of target terminal identification collections with push qualifications determined after location field matching and user subscription strategy screening after the generation of the current warning information. This set can include the device identifications of road-side display terminals, such as electronic information screen numbers, road section display addresses, etc., and can also include user-side device identifications, such as APP client IDs, mobile device Tokens, in-vehicle system codes, etc.

[0137] In specific implementation, first extract the abnormal geographical coordinates and road number information in the location field, call the preset terminal mapping table to determine the road-side display terminals that match the location range. Subsequently, according to the structure type field and the abnormal category field, compare with the area range, structure type of interest, and warning level preference set in the user subscription strategy to screen out the user terminals that meet the push conditions. Finally, combine the road-side display terminals and the user terminals that meet the screening conditions to form a set of terminal push objects, which is used to perform subsequent warning information distribution operations.

[0138] Finally, encapsulate the warning information data structure according to the configuration rules of the terminal push object set, and send it to the display terminal and the user terminal respectively for visual display.

[0139] First, for various terminal types in the terminal push object set, read their corresponding content presentation capabilities, data interface requirements, and push format templates as the basis for configuration rules. For the display terminal, extract the abnormal category field and abnormal location field from the warning information data structure, perform field clipping and format conversion, and generate a warning broadcast text containing a brief description of the abnormality, road abbreviation, and current time based on the configuration rules. For example, for an event such as "main beam crack in the G108 bridge section", it is encapsulated as "Abnormal main beam of G108, please pay attention and avoid", and the display duration field and blink style configuration are added.

[0140] For the user terminal, combine the structure type field and the image recognition source field to generate a push format, including abnormal category description, full road name, map location link, thumbnail of the abnormal structure, which can be cropped from the original image, timestamp of occurrence, etc. fields, and encapsulate them into a JSON or protocol message structure. According to the user terminal identifier, send the encapsulated content to the corresponding user terminal through the push channel. After each terminal receives the data, call the local visualization module to display the information: the display terminal refreshes the screen content according to the field rendering logic. The user terminal displays the abnormal information in the form of a pop-up window, notification bar, or map annotation, and supports the user to view the image details or perform feedback operations.

[0141] In some embodiments, the controlling the drone to go to the corresponding abnormal area in response to the abnormal recognition result meeting the preset conditions specifically includes the following technical steps:

[0142] First, parse the abnormal recognition result, extract the abnormal type, occurrence location, and image source, and compare the abnormal type with the preset conditions to determine whether the trigger condition is met.

[0143] Exemplarily, specifically, obtain the anomaly recognition results output by the road anomaly recognition model, bridge anomaly recognition model, or tunnel anomaly recognition model. The included anomaly category field is used to represent the identified structural anomaly types, such as road potholes, bridge main girder cracks, or tunnel vault water seepage, etc. Further extract the anomaly location field, including the geographical coordinates provided by the image acquisition device or UAV positioning system (such as longitude and latitude values in WGS-84 format), the corresponding road number (such as "G108 - eastward K12+300"), and possible structural positioning auxiliary fields. At the same time, parse the image recognition source field, which can be generated based on the content of the image acquisition record. For example, "UAV-CAM-01" indicates that the image is collected by the UAV numbered 01, and "FIXED-BRIDGE-03" indicates that the image is obtained by the image acquisition device installed at the position numbered 03 of the bridge structure.

[0144] After completing the field parsing, perform a field-level matching of the anomaly category field with the preset condition set stored locally. The preset condition set is set by the management system and can be dynamically adjusted. It usually includes the anomaly types to be responded to (such as structural cracks, leakage, road surface collapse, etc.), the corresponding structural types (bridges, tunnels, or roads), the trigger level (such as medium and above), and the response methods, etc. For example, when the anomaly type in the recognition result is tunnel vault water seepage and the structural type is a tunnel, and it is retrieved in the preset conditions that: "structural type = tunnel, anomaly type ∈ {vault water seepage, wall crack}, trigger level = all", it is considered to meet the trigger condition and enter the subsequent response process, such as the automatic generation of flight scheduling for the anomaly area or on-site broadcast tasks.

[0145] Then, in response to the anomaly type meeting the trigger condition, calculate the three-dimensional navigation path of the target anomaly area based on the occurrence location and the current UAV's staying position, and generate a set of track data.

[0146] Specifically, first, based on the anomaly location field included in the anomaly recognition result, extract the geographical coordinate information of the target anomaly area, including longitude, latitude, and an optional relative height parameter, for constructing the target point P target =(x t , y t , z t ). At the same time, obtain the spatial staying position coordinates P start =(x s , y s , z s ) of the currently standby UAV.

[0147] Subsequently, the preset flight mission planning module is called, and in the path planning process, an improved 3D A* algorithm or a trajectory smoothing algorithm based on B-spline curves can be used to generate a continuous sequence of flight nodes. Each flight node contains information such as spatial coordinates, flight altitude, heading angle, and target speed, which is used to guide the flight behavior of the UAV in different flight segments. Finally, the sequence of path nodes from P start to P target is encoded into a set of track data, where each track segment can be expressed as:

[0148] T i ={(x i ,y i ,z i ),θ i ,v i}

[0149] where (x i ,y i ,z i ) represents the coordinates of the track point, θ i represents the heading angle, and v i represents the target flight speed of this segment. All track segments are combined in sequence to form a complete set of track data, which is used for the subsequent control system to send to the UAV flight control unit to perform autonomous flight tasks.

[0150] Finally, the set of track data is sent to the UAV flight control system, the flight parameters are configured and the flight mission is started to control the UAV to fly to the target abnormal area.

[0151] Specifically, the generated set of track data is packaged in the format of a mission instruction and sent to the flight control system of the target UAV. The flight control system analyzes the track point positions, heading angles, and flight speed parameters in the set of track data, sets the flight mode, track tracking strategy, and obstacle avoidance configuration parameters to complete the initialization of the flight mission. Subsequently, the flight control system starts the autonomous flight mission, and the UAV flies sequentially according to the received sequence of track points, adjusts the attitude and altitude in real time, and flies along the planned path to the target abnormal area.

[0152] In some embodiments, the above intelligent warning method for dangerous road sections based on image acquisition and recognition may further include the following technical steps:

[0153] First step, obtain road infrared images, bridge infrared images, and tunnel infrared images and generate an infrared image data set. Among them, by obtaining road infrared images, bridge infrared images, and tunnel infrared images and generating an infrared image data set, it can effectively make up for the insufficient perception ability of visible light images in night or low-illumination environments, and help improve the detectability of abnormal states at night. On the other hand, infrared images can be used to capture temperature anomaly features that are invisible to the naked eye, such as internal heat accumulation in the main beam of a bridge, cold spot areas caused by leakage on the tunnel wall surface, or local overheating areas on the road surface, etc., so as to achieve supplementary detection of early concealed anomalies.

[0154] Second step, input the infrared image data set into an infrared anomaly recognition model, extract the feature of the heat value distribution of each pixel in the image, and generate a corresponding infrared feature data set.

[0155] Among them, the infrared anomaly recognition model can represent an image analysis model used to extract and recognize the heat value distribution features in road infrared images, bridge infrared images, and tunnel infrared images. In this embodiment, the infrared anomaly recognition model can be constructed by using a shallow thermal perception network that combines a multi-scale convolution structure and a local response normalization mechanism. The heat value distribution can represent the infrared radiation intensity value corresponding to each pixel in the infrared image, reflecting the surface temperature characteristics of this position.

[0156] In specific implementation, input the infrared image data set I IR into the infrared anomaly recognition model. First, perform normalization processing on the heat value distribution of each pixel position in the image to obtain a thermal response matrix:

[0157]

[0158] Among them, T(x,y) represents the normalized heat value, and I IR (x,y) represents the original infrared radiation intensity of the infrared image at the pixel position (x,y), μ T represents the mean value of the heat values of the entire image, and σ T represents the standard deviation of the heat values, which is used to suppress errors caused by differences in the imaging environment or sensor sensitivity.

[0159] Subsequently, input T(x,y) into a shallow thermal perception network that combines a multi-scale convolution structure and a local response normalization mechanism. In the first stage of the network, convolution operations are performed in parallel using convolution kernels with different receptive fields to extract the local and global response features of the hot spot area:

[0160]

[0161] Among them, represents the convolution feature map at the kth scale, and W kDenote the convolution kernel parameters of different sizes, * represents the convolution operation, b k Denote the bias term, and ReLU is the activation function.

[0162] To enhance the model's sensitivity to local hot spot anomalies, local response normalization operations can be introduced to perform normalization processing within each convolution feature map channel:

[0163]

[0164] Among them, Denote the normalized feature response, α, β, γ are hyperparameters of the normalization coefficients, n represents the width of the normalization channel window, and K represents the total number of feature maps.

[0165] Concatenate the normalized feature maps of all scales along the channel dimension to construct a unified infrared feature data set:

[0166]

[0167] Among them, F IR Denote the finally generated infrared feature data set, which contains temperature spatial response information at multiple scales and serves as the input basis for subsequent thermal anomaly analysis and classification judgment.

[0168] In the third step, perform local temperature difference analysis on the infrared feature data set, identify the image regions with abnormal heat value distributions, and combine the structural type labels of the images for result annotation to generate a set of thermal anomaly targets.

[0169] Specifically, based on the heat value responses at each pixel position in the infrared feature data set F IR , calculate the temperature statistical features of each local region. Select a sliding window of r×r for region division, and calculate the local average temperature value at the center point (x, y) of each window:

[0170]

[0171] Among them, μ IR (x, y) represents the local temperature mean centered at (x, y), and F IR (x + i, y + j) represents the heat value at the pixel (x + i, y + j) in the infrared feature response map, and r is the side length of the window.

[0172] Furthermore, calculate the local standard deviation of the temperature change within this region:

[0173]

[0174] Compare the standard deviation σ IR (x, y) of each local region with the average temperature fluctuation level of the entire image Make a comparison to determine whether it constitutes an abnormal region of calorific value distribution:

[0175]

[0176] Among them, M IR (x, y) indicates whether the position (x, y) in the thermal anomaly mask image is marked as abnormal, λ is an empirical threshold coefficient (such as taking 1.5 - 2.0), represents the mean value of the temperature standard deviation of the entire image.

[0177] Subsequently, combine the structure type identifier corresponding to each image (such as "road", "bridge" or "tunnel") with the current image source annotation field, and jointly encode the positions of the marked abnormal regions and the corresponding structural semantics to generate a set of thermal anomaly targets:

[0178]

[0179] Among them, (x k , y k ) represents the central position of the k-th marked thermal anomaly region, type k represents the corresponding structure type identifier, and the set of thermal anomaly targets represents the set of all target regions identified as having abnormal thermal distribution in the current infrared image.

[0180] Fourth step, perform classification operations and temperature threshold judgment operations on the set of thermal anomaly targets to generate infrared anomaly recognition results corresponding to the road infrared image, bridge infrared image, and tunnel infrared image.

[0181] Specifically, perform thermal feature extraction operations on each abnormal region in the set of thermal anomaly targets to obtain the average temperature, temperature difference range, and calorific value comparison information with neighboring regions of the region. According to the structure type corresponding to the infrared image, call the preset temperature judgment rules to classify and judge the calorific value performance. For example, if the calorific value of a certain region in the bridge infrared image is significantly lower than the surrounding regions and lower than the lower limit of the temperature response of the bridge structure, it can be determined as "low temperature anomaly"; if a certain region in the road image shows concentrated high heat performance and exceeds the corresponding threshold range, it is marked as "high temperature anomaly". If there is a temperature jump region distributed along the surface of the structure in the tunnel image, it can be identified as "abnormal thermal distribution". After the judgment is completed, attach the corresponding anomaly type label to each abnormal region, and retain its position identifier and structure type identifier. Multiple annotation results are combined to form an infrared anomaly recognition result, which is used to clarify the types and distribution positions of thermal anomalies existing in the current infrared image, and serves as the basis for generating subsequent warning information.

[0182] In this embodiment, by introducing an infrared image processing process, the abnormal recognition ability in road, bridge and tunnel scenarios at night or in low-light environments is significantly improved. On the one hand, by acquiring and uniformly processing road infrared images, bridge infrared images and tunnel infrared images, a multi-scene thermal perception image data set with temperature information characteristics can be formed, providing a stable data basis for subsequent recognition. On the other hand, combining the infrared abnormal recognition model to extract features from the pixel-level heat value distribution, and on this basis, performing local temperature difference analysis and structure type annotation, which helps to accurately locate the heat value abnormal area and clarify its structural semantics. Further through classification operations and temperature threshold judgments, various types of thermal abnormal states can be effectively recognized, such as local overheating, cold spot aggregation or abnormal heat distribution, etc., thereby improving the accuracy, robustness and applicable time coverage of abnormal detection, and enhancing the practicality and response ability of the early warning method under all-weather conditions.

[0183] In addition, in other embodiments of the present disclosure, displacement change data and vibration response data of the target area can also be collected by radar detection devices pre-set on the bridge structure and the tunnel structure, and whether the bridge structure and the tunnel structure are abnormal can be determined according to the displacement change data and the vibration response data.

[0184] Among them, the radar detection device can refer to a detection device that is set at key parts of the bridge structure and the tunnel structure and realizes non-contact measurement based on transmitting electromagnetic waves and receiving reflected wave signals. The radar detection device can be a continuous wave radar, a frequency modulated continuous wave radar or a synthetic aperture radar device, etc. The displacement change data can refer to the relative displacement amount calculated based on the phase change amount of the reflected wave of the target area obtained by the radar detection device at different time points, which can be used to reflect the dynamic displacement information of the local structure of the bridge or tunnel in the vertical, horizontal or longitudinal directions. The vibration response data can refer to the small periodic motion characteristics of the local area of the bridge or tunnel measured by the radar detection device during continuous sampling, including but not limited to data such as the main vibration frequency, amplitude change, and vibration mode characteristics.

[0185] In specific implementation, first, according to the structural characteristics of bridges and tunnels, fixed continuous-wave radar or frequency-modulated continuous-wave radar detection devices are arranged at the main girders, bearings and key connection nodes of bridges, as well as in areas such as the walls, vaults and side walls of tunnels. The radar detection devices emit electromagnetic signals at a set frequency and receive the reflected signals from the target surface in real time. Subsequently, by analyzing the phase change of the reflected signals, the displacement change amount of the monitored target within the sampling period is calculated. The displacement change amount can be converted based on the wavelength of the transmitted signal and the phase change amount to obtain the displacement data of the local structure of the bridge or tunnel in the vertical, horizontal or longitudinal directions. Then, the vibration frequency extraction and displacement trend analysis are performed on the displacement change data collected within a continuous time period to identify characteristic parameters such as the maximum displacement amplitude, the change of the main vibration frequency, and the change of the vibration mode, and the characteristic parameters are compared with the corresponding safety thresholds of the bridge or tunnel structure. Finally, based on the comparison result, it is judged whether there are abnormal changes in the target bridge area or the target tunnel area, including situations such as the displacement amount exceeding the set safety threshold, abnormal vibration frequency, and abnormal change of the vibration mode, and the monitoring status of the bridge structure and the tunnel structure is updated based on the judgment result to trigger the subsequent early warning processing process. It should be noted that although the steps of the method in the present disclosure are described in a specific order in the drawings, this does not require or imply that these steps must be executed in this specific order, or that all the shown steps must be executed to achieve the desired result. Additionally or alternatively, some steps can be omitted, multiple steps can be combined into one step for execution, and / or one step can be decomposed into multiple steps for execution, etc.

[0186] In addition, in the present exemplary embodiment, an intelligent early warning device for dangerous sections based on image acquisition and recognition is also provided. Referring to Figure 5 as shown, the intelligent early warning device 500 for dangerous sections based on image acquisition and recognition includes: an image acquisition module 510, an anomaly recognition module 520, an early warning sending module 530, and a drone guidance module 540. Among them:

[0187] The image acquisition module 510 is configured to acquire road images, bridge images, and tunnel images acquired by image acquisition devices and drones;

[0188] The anomaly recognition module 520 is configured to input the road images, bridge images, and tunnel images into corresponding anomaly recognition models respectively based on the image source and image structure characteristics to generate anomaly recognition results;

[0189] The early warning sending module 530 is configured to generate early warning information based on the anomaly recognition results and automatically send the early warning information to display terminals, user terminals, and management terminals along the road;

[0190] The UAV guidance module 540 is configured to control the UAV to go to the corresponding abnormal area in response to the abnormal recognition result meeting a preset condition, and use the UAV to perform on-site audio warning and traffic guidance.

[0191] The specific details of each module of the above intelligent warning device for dangerous sections based on image acquisition and recognition have been described in detail in the corresponding intelligent warning method for dangerous sections based on image acquisition and recognition, so they will not be elaborated here.

[0192] It should be noted that although several modules or units of the intelligent warning device for dangerous sections based on image acquisition and recognition are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of the two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided and embodied by multiple modules or units.

[0193] In addition, in the exemplary embodiments of the present disclosure, an electronic device capable of implementing the above intelligent warning method for dangerous sections based on image acquisition and recognition is also provided.

[0194] Those skilled in the art can understand that various aspects of the present disclosure can be implemented as a system, method, or program product. Therefore, various aspects of the present disclosure can be specifically implemented in the following forms, namely: a complete hardware embodiment, a complete software embodiment (including firmware, microcode, etc.), or an embodiment combining hardware and software aspects, which can be collectively referred to as "circuit", "module", or "system" here.

[0195] Next, refer to Figure 6 to describe the electronic device 600 according to this embodiment of the present disclosure. Figure 6 The shown electronic device 600 is only an example and should not bring any limitation to the functions and usage scope of the embodiments of the present disclosure.

[0196] As Figure 6 shown, the electronic device 600 is presented in the form of a general-purpose computing device. The components of the electronic device 600 may include, but are not limited to: at least one of the above processing units 610, at least one of the above storage units 620, a bus 630 connecting different system components (including the storage unit 620 and the processing unit 610), and a display unit 640.

[0197] Among them, the storage unit stores program codes, and the program codes can be executed by the processing unit 610, so that the processing unit 610 executes the steps according to various exemplary embodiments of the present disclosure described in the above "exemplary method" section of this specification. For example, the processing unit 610 can execute asFigure 1 In step S110 shown in Figure 1 , road images, bridge images, and tunnel images collected by an image acquisition device and a drone are obtained; in step S120, the road images, bridge images, and tunnel images are respectively input into corresponding anomaly recognition models to generate anomaly recognition results, where the anomaly recognition models include a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model; in step S130, a warning message is generated based on the anomaly recognition results and sent to display terminals and user terminals along the road; in step S140, in response to the anomaly recognition results meeting a preset condition, the drone is controlled to fly to the corresponding anomaly area, and on-site audio warnings and traffic guidance are carried out using the drone.

[0198] The storage unit 620 may include a readable medium in the form of a volatile storage unit, such as a random access storage unit (RAM) 621 and / or a cache storage unit 622, and may further include a read-only storage unit (ROM) 623.

[0199] The storage unit 620 may also include a program / utility 624 having a set (at least one) of program modules 625. Such program modules 625 include, but are not limited to: an operating system, one or more application programs, other program modules, and program data. Each or some combination of these examples may include an implementation of a network environment.

[0200] The bus 630 may represent one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0201] The electronic device 600 may also communicate with one or more external devices 670 (such as a keyboard, a pointing device, a Bluetooth device, etc.), may also communicate with one or more devices that enable a user to interact with the electronic device 600, and / or may communicate with any device that enables the electronic device 600 to communicate with one or more other computing devices (such as a router, a modem, etc.). Such communication may be carried out through an input / output (I / O) interface 650. And the electronic device 600 may also communicate with one or more networks (such as a local area network (LAN), a wide area network (WAN), and / or a public network, such as the Internet) through a network adapter 6100. As shown in the figure, the network adapter 6100 communicates with other modules of the electronic device 600 through the bus 630. It should be understood that although not shown in the figure, other hardware and / or software modules may be used in conjunction with the electronic device 600, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, etc.

[0202] From the description of the above embodiments, it is easy for those skilled in the art to understand that the example embodiments described herein can be implemented by software or in the form of software combined with necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a portable hard drive, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0203] In an exemplary embodiment of the present disclosure, there is also provided a computer-readable storage medium, on which there is a program product capable of implementing the above method of the present specification. In some possible embodiments, various aspects of the present disclosure can also be implemented in the form of a program product, which includes program code. When the program product runs on a terminal device, the program code is used to enable the terminal device to execute the steps according to various exemplary embodiments of the present disclosure described in the above "Exemplary Method" section of the present specification.

[0204] Reference Figure 7 As shown, a program product 700 for implementing the above intelligent warning method for dangerous sections based on image acquisition and recognition according to an embodiment of the present disclosure is described. It can be a portable compact disc read-only memory (CD-ROM) and includes program code, and can run on a terminal device, such as a personal computer. However, the program product of the present disclosure is not limited thereto. In this document, the readable storage medium can be any tangible medium that contains or stores a program, and this program can be used by or in combination with an instruction execution system, apparatus, or device.

[0205] The program product can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples (non-exhaustive list) of the readable storage medium include: an electrical connection having one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0206] A computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which readable program code is carried. Such a propagated data signal may take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. The readable signal medium may also be any readable medium other than a readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device.

[0207] The program code contained on the readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RF, and the like, or any suitable combination of the foregoing.

[0208] The program code for performing the operations of the present disclosure may be written in any combination of one or more programming languages, including object-oriented programming languages such as Java, C++, etc., and also including conventional procedural programming languages such as the "C" language or similar programming languages. The program code may be executed entirely on the user's computing device, partially on the user's device, executed as a stand-alone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In the case of a remote computing device, the remote computing device may be connected to the user's computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computing device (e.g., through the Internet using an Internet service provider).

[0209] In addition, the above drawings are only schematic illustrations of the processes included in the method according to the exemplary embodiments of the present disclosure, and are not for limiting purposes. It is easy to understand that the processes shown in the above drawings do not indicate or limit the chronological order of these processes. Additionally, it is also easy to understand that these processes may be executed synchronously or asynchronously, for example, in multiple modules.

[0210] Through the description of the above embodiments, those skilled in the art can easily understand that the example embodiments described herein can be implemented by software or by a combination of software and necessary hardware. Therefore, the technical solutions according to the embodiments of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which may be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which may be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the embodiments of the present disclosure.

[0211] Other embodiments of the present disclosure will be readily apparent to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. The present disclosure is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include known common general knowledge or conventional technical means in the technical field not disclosed in the present disclosure. The specification and examples are only to be considered as exemplary, and the true scope and spirit of the present disclosure are pointed out by the claims.

[0212] It should be understood that the present disclosure is not limited to the exact structures described above and shown in the drawings, and various modifications and changes can be made without departing from its scope. The scope of the present disclosure is only limited by the appended claims.

Claims

1. An intelligent warning method for dangerous road sections based on image acquisition and recognition, characterized in that, Including: Obtaining road images, bridge images, and tunnel images collected by an image acquisition device and a drone; Respectively inputting the road images, bridge images, and tunnel images into corresponding anomaly recognition models to generate anomaly recognition results, where the anomaly recognition models include a road anomaly recognition model, a bridge anomaly recognition model, and a tunnel anomaly model; Generating a warning message based on the anomaly recognition results and automatically sending the warning message to display terminals, user terminals, and management terminals along the road; In response to the anomaly recognition results meeting a preset condition, controlling the drone to fly to the corresponding anomaly area and using the drone for on-site audio warning and traffic guidance.

2. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 1, characterized in that The obtaining of road images, bridge images, and tunnel images collected by an image acquisition device and a drone includes: Collecting bridge images through an image acquisition device pre-set on the bridge structure, where the bridge images include any one or more of images of the bridge substructure, bridge flexible structure, bridge deck, main girder, vehicle, person, and bearing; Collecting tunnel images through an image acquisition device pre-set on the tunnel structure, where the tunnel images include any one or more of images of the tunnel wall, vault, sidewall, tunnel pavement, cable trench, fan, vehicle, lighting equipment, and drainage structure; Obtaining road images through the drone at preset time intervals and inspection paths, where the road images include any one or more of images of the road surface, slope, drainage facilities, and guardrail; When the image acquisition device fails to complete the image acquisition of the corresponding area, collecting the bridge image or tunnel image of that area through the drone.

3. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 1, wherein The respectively inputting the road images, bridge images, and tunnel images into corresponding anomaly recognition models to generate anomaly recognition results includes: Inputting the road images into a road anomaly recognition model constructed by a lightweight convolutional neural network and a spatial pyramid pooling structure to generate road anomaly recognition results; Inputting the bridge images into a bridge anomaly recognition model constructed by a multi-layer residual structure to generate bridge anomaly recognition results; Inputting the tunnel images into a tunnel anomaly recognition model constructed by an encoder-decoder semantic segmentation network to generate tunnel anomaly recognition results; In response to the bridge anomaly recognition result being cracks, deformation, and displacement or the tunnel anomaly recognition result being cracks and water seepage, constructing anomaly time series data based on the historical image data and the current recognition results of the corresponding anomaly area; Inputting the anomaly time series data into a pre-trained prediction model for trend analysis to monitor the change trends of crack propagation, structural deformation quantity, and displacement trajectory.

4. The intelligent early warning method for dangerous sections based on image acquisition and recognition according to claim 3, characterized in that, The inputting the road images into a road anomaly recognition model constructed by a lightweight convolutional neural network and a spatial pyramid pooling structure to generate road anomaly recognition results includes: Adjust the size of the road image and standardize the pixel values to obtain preprocessed image data, and input the preprocessed image data into a convolutional network structure composed of multiple depthwise separable convolutional units to extract corresponding multi-scale image features layer by layer, generating a first feature data set; Perform multi-scale pooling operations on the first feature data set according to a set spatial pyramid structure, extract local statistical features at different scales, and generate a second feature data set; Concatenate the first feature data set and the second feature data set along the channel dimension to generate an enhanced feature data set, and the enhanced feature data set has the same spatial size as the first feature data set; Input the enhanced feature data set into an identification sub-network with a fully connected structure and a classification function, perform feature flattening and class classification processing, and output an abnormal class label corresponding to the road image.

5. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 3, characterized in that, The input of the bridge image into a bridge anomaly recognition model constructed by a multi-layer residual structure to generate a bridge anomaly recognition result includes: Perform region separation on the bridge image through a segmentation algorithm based on edge gradient and spatial texture features, and extract structure region images including the bridge substructure, bridge flexible structure, main girder, bridge deck, and bearings; Input the structure region image into a feature extraction network with a multi-layer residual structure, extract local edge features, joint features, and deformation contour features in different channels respectively, and generate multi-channel residual feature data; Perform cross-channel feature comparison operations on the multi-channel residual feature data, obtain differential feature data between different structure channels through feature difference calculation, and fuse the differential feature data with the original channel features in the multi-channel residual feature data to generate enhanced differential features; Input the enhanced differential features into an identification sub-network including a fully connected structure and a pattern matching classification function, and output an abnormal class label corresponding to the bridge image.

6. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 3, wherein The input of the tunnel image into a tunnel anomaly recognition model constructed by an encoder-decoder semantic segmentation network to generate a tunnel anomaly recognition result includes: Input the tunnel image into a semantic segmentation network including a multi-level encoding layer and a decoding layer structure, perform pixel-level semantic segmentation on each region of the tunnel image, and generate a segmentation label image including a wall region, a vault region, and a pipeline region; Based on the tunnel wall region and vault region marked in the segmentation label image, extract the segmentation probability distribution of the corresponding pixels, and perform significance analysis on the segmentation probability distribution to obtain an abnormal response image; Extract the regions in the abnormal response image with significance higher than a set threshold as candidate abnormal regions, and extract contour curvature and gray gradient features for each candidate abnormal region to obtain a region feature vector set; Input the region feature vector set into a classification network including a feature aggregation structure and a discriminant function, and output an abnormal class label corresponding to the tunnel image.

7. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 1, characterized in that, Generating a warning message based on the anomaly recognition result and automatically sending the warning message to display terminals, user terminals, and management terminals along the road, including: Perform data parsing on the abnormal recognition result, extract the abnormal category, abnormal location, and structure type fields, and generate a set of original warning contents; Integrate the set of original warning contents with road numbers, road locations, timestamps, and image recognition sources to construct a warning information data structure with positioning fields and structure fields; Based on the location field in the warning information data structure, match the display terminals along the road corresponding to the abnormal location, and identify user terminals that meet the conditions in combination with the user subscription strategy to generate a set of terminal push objects; Package the warning information data structure according to the configuration rules of the set of terminal push objects, and send it to the display terminal and the user terminal respectively for visual display.

8. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 1, characterized in that, Responding to the abnormal recognition result meeting the preset conditions and controlling the drone to go to the corresponding abnormal area includes: Parse the abnormal recognition result, extract the abnormal type, occurrence location, and image source, and compare the abnormal type with the preset conditions to determine whether the trigger condition is met; In response to the abnormal type meeting the trigger condition, calculate the three-dimensional navigation path of the target abnormal area based on the occurrence location and the current drone stay location, and generate a set of trajectory data; Send the set of trajectory data to the drone flight control system, configure the flight parameters and start the flight mission, and control the drone to fly to the target abnormal area.

9. The intelligent warning method for dangerous road sections based on image acquisition and recognition according to claim 1, characterized in that, It also includes: Obtain road infrared images, bridge infrared images, and tunnel infrared images and generate a set of infrared image data; Input the set of infrared image data into the infrared abnormal recognition model, extract the characteristics of the calorific value distribution of each pixel in the image, and generate a corresponding set of infrared feature data; Perform local temperature difference analysis on the set of infrared feature data, identify the image areas with abnormal calorific value distribution, and perform result annotation in combination with the structure type identifier of the image to generate a set of thermal abnormal targets; Perform classification operations and temperature threshold judgment operations on the set of thermal abnormal targets to generate infrared abnormal recognition results corresponding to the road infrared images, bridge infrared images, and tunnel infrared images; Collect displacement change data and vibration response data of the target area through radar detection devices pre-set on the bridge structure and tunnel structure, and determine whether the bridge structure and tunnel structure are abnormal according to the displacement change data and vibration response data.

10. An intelligent warning device for dangerous road sections based on image acquisition and recognition, characterized in that, It includes: An image acquisition module for obtaining road images, bridge images, and tunnel images collected by image acquisition devices and drones; An abnormal recognition module for inputting the road images, bridge images, and tunnel images into the corresponding abnormal recognition models respectively based on the image source and image structure characteristics to generate abnormal recognition results; A warning sending module for generating warning information based on the abnormal recognition result and automatically sending the warning information to the display terminals, user terminals, and management terminals along the road; A drone guidance module for controlling the drone to go to the corresponding abnormal area in response to the abnormal recognition result meeting the preset conditions, and using the drone for on-site audio warning and traffic guidance.

Images