Road surface disaster detection method and system, electronic equipment and storage medium

Abstract

The invention discloses a road surface disaster detection method and system, electronic equipment and a storage medium, and relates to the technical field of intelligent traffic management and computer vision, and the method comprises the steps: collecting image data containing a plurality of road surface disaster types, carrying out the data marking in each image data, and carrying out the data marking; a data sample is constructed through each piece of image data subjected to data annotation; training the improved YOLOv10 model by using the data sample until a loss function of the improved model reaches a preset end condition, and determining the improved model of which the loss function reaches the preset end condition as a disaster detection model; acquiring a to-be-detected image of a to-be-detected road surface, and inputting the to-be-detected image into the disaster detection model for processing to obtain a disaster detection result of the to-be-detected road surface, the disaster detection result including a disaster type and a disaster position; by combining the deep learning technology and optimization algorithm design, the problem that the existing method is insufficient in detection precision, robustness and applicability is solved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent traffic management and computer vision technology, and more specifically, to a road disaster detection method, system, electronic equipment and storage medium. Background Art

[0002] With the rapid development of modern transportation networks, real-time detection and processing of road hazards (such as cracks, potholes, water accumulation, oil pollution, etc.) has become an important part of ensuring road safety and improving traffic efficiency; however, current road hazard detection methods mainly rely on manual inspections or traditional camera monitoring, which have the following significant limitations: Manual inspection: time-consuming and labor-intensive, limited coverage, and subject to human subjective factors, which can easily lead to the omission of disasters; Traditional monitoring: Monitoring equipment based on fixed cameras has insufficient performance in detecting small or subtle disasters (such as cracks), especially under changing lighting or complex background conditions.

[0003] Existing automated detection methods mostly use deep learning models, but when faced with various types of road disasters, obvious scale changes, and complex environments (such as lighting changes and target occlusion), there are still problems such as insufficient detection accuracy and poor robustness. Summary of the invention

[0004] The purpose of the present invention is to provide a road hazard detection method, system, electronic device and storage medium, which combine deep learning technology with optimization algorithm design to solve the problems of insufficient detection accuracy, robustness and applicability of existing methods.

[0005] The above technical objectives of the present invention are achieved through the following technical solutions: In a first aspect, the present application provides a road hazard detection method, comprising the following specific steps: Collect image data covering various types of road disasters, annotate each image data, and construct data samples through each annotated image data; The improved YOLOv10 model is trained using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and the improved YOLOv10 model whose loss function reaches the preset end condition is determined as a disaster detection model, wherein: The improved YOLOv10 model includes FM module and EMA module, and the loss function is calculated by angle loss, distance loss, shape loss and IoU loss; The image to be tested of the road surface to be tested is obtained, and the image to be tested is input into the disaster detection model for processing to obtain the disaster detection result of the road surface to be tested, and the disaster detection result includes the disaster type and disaster location.

[0006] The beneficial effects of the present invention are as follows: in this scheme, firstly, image data covering various types of road disasters are collected, and data annotation is performed in each image data to construct a data sample; secondly, the improved YOLOv10 model is trained using the training set in the constructed data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, thereby determining the improved YOLOv10 model whose loss function reaches the preset end condition as a disaster detection model; finally, the acquired image to be tested is input into the disaster detection model for processing, and finally the disaster detection result of the road to be tested is obtained.

[0007] In this solution, various road disaster types such as cracks, potholes, water accumulation, and oil stains are detected, and it can also be expanded to other potential new disaster types; by covering different traffic scenarios (such as highways, urban roads) and environmental conditions (such as sunny days, rainy days, and insufficient light), the system's wide applicability is improved; at the same time, through the improved YOLOv10 model, the FM module, the EMA module, and the improved loss function are introduced to optimize the model, enhance the small target detection capability, and optimize the performance under complex backgrounds, thereby greatly reducing the missed detection rate and improving the overall detection performance, thus solving the current problems of insufficient detection accuracy and poor robustness when facing various road disaster types, obvious scale changes, and complex environments.

[0008] In this solution, multimodal feature fusion and dynamic modulation mechanisms are used to significantly enhance the stability and robustness of the system in complex traffic scenarios such as lighting changes and target occlusion, ensuring high performance in a variety of environments. At the same time, through lightweight modules and optimized algorithms, real-time operation capabilities are guaranteed while reducing hardware resource requirements, supporting efficient deployment on edge devices (such as vehicle-mounted terminals) to meet the needs of intelligent traffic management.

[0009] Based on the above technical solution, the present invention can also be improved as follows.

[0010] Furthermore, the FM module includes a context aggregation unit and a feature modulation unit, wherein: A context aggregation unit is used to generate a multi-level feature set from each data in the training set or the image to be tested; The feature modulation unit is used to adjust the fine granularity of each feature in the feature set through a dynamic modulator.

[0011] The beneficial effect of adopting the above further scheme is that after replacing the original SPPF module in the model with the FM module, the model can significantly improve the modeling ability and detection performance of multi-scale features. Different from the fixed pooling operation of the SPPF module, the Focal Modulation (FM module) dynamically aggregates contextual information from local to global through multi-level deep convolution, which can more accurately capture the detailed features of small targets and strengthen the global semantic understanding under complex backgrounds. This design can show higher robustness for small target detection (such as fine cracks) and complex scenes (such as lighting changes or target occlusion); at the same time, the Focal Modulation module adopts a lightweight computing structure to avoid the high computational complexity problem that may be caused by SPPF, and optimizes computing efficiency through deep convolution and linear transformation, which not only reduces the hardware resource requirements, but also improves the inference speed; overall, Focal Modulation surpasses SPPF in detection accuracy, scene adaptability and efficiency, providing a more efficient and robust solution for road disaster detection.

[0012] Furthermore, the above EMA module includes a feature grouping unit, a parallel sub-network unit and a cross-space learning unit, wherein: A feature grouping unit, used to disperse the individual features in the fine-grained adjusted feature set into different subgroups; Parallel sub-network units are used to obtain the global spatial information and multi-scale local features of each feature in each sub-group through different multi-scale branch models. The global spatial information is extracted by 1D global average pooling operations in the horizontal and vertical directions, and the multi-scale local features are extracted by convolution operations. The cross-space learning unit is used to combine the global spatial information and multi-scale local features of each feature in each subgroup through a matrix dot product operation to obtain a weighted aggregate feature of each feature in each subgroup.

[0013] The beneficial effects of adopting the above further scheme are: the EMA module has significant advantages in multi-scale feature modeling and computational efficiency; the EMA module can capture local details and global context information at the same time by introducing multi-level context aggregation and dynamic modulation mechanism, making it perform better in small target detection and complex scenes; in addition, the EMA module uses group normalization and lightweight deep convolution to avoid the complexity of global attention calculation in PSA, while reducing the demand for computing resources and improving detection efficiency, which is more suitable for actual deployment; compared with the PSA module, the dynamic feature modulation mechanism of the EMA module enables it to have stronger adaptability and show higher robustness in lighting changes, occlusion or complex backgrounds. By enhancing the fine feature extraction and global semantic understanding of small targets, EMA improves the detection accuracy and recall rate of the model. These improvements provide YOLOv10 with stronger performance and higher application value in tasks such as road disaster detection.

[0014] Furthermore, the above loss function is specifically: ; In the formula, represents the loss function, represents the IoU loss, Indicates distance loss, and distance loss Obtained through angle loss calculation, Represents shape loss.

[0015] The beneficial effects of adopting the above further scheme are: the improved loss function enables the model to show a clearer optimization direction and significantly higher convergence efficiency in bounding box optimization; SIoU dynamically adjusts the gradient optimization path of the bounding box in the horizontal and vertical directions by introducing angle cost and shape alignment mechanism, effectively reducing unnecessary "wandering" phenomenon during training. This design enables the prediction box to approach the center of the target box more quickly, thereby accelerating training convergence, and significantly improving the model's adaptability to complex target shapes, especially in scenes with large aspect ratio differences or obvious target position deviations; compared with the CIoU loss function, the improved SIoU loss function shows higher positioning accuracy and recall rate in small target detection and complex backgrounds; its dynamic angle guidance strategy significantly reduces the conflict of multi-component optimization, and the shape cost effectively solves the shortcomings of aspect ratio matching between the prediction box and the real box. This multi-level optimization not only enables the model to capture the detailed features of tiny targets (such as road cracks) more accurately, but also enhances the robustness and generalization ability in multi-target detection scenarios; by optimizing training efficiency and improving final performance, SIoU becomes a more efficient and stable solution for target detection tasks.

[0016] Furthermore, the above distance loss is specifically: ,in: , ; ; In the formula, Represents the distance loss, It represents the X-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. It represents the Y-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. Represents the center point of the target box in the image data after data annotation x Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model x Coordinate values, Indicates the width range of the target box in the data annotation and the target box in the prediction result. Indicates the height range of the target box in the data annotation and the target box in the prediction result. represents the comprehensive weight factor, Represents the center point of the target box in the image data after data annotation y Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model y Coordinate value; represents the angle loss, and the angle loss is specifically: , ;in: ; ; In the formula, represents the normalized relative height in the angle loss formula, Indicates the height range of the target box in the data annotation and the target box in the prediction result. Represents the Euclidean distance between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model.

[0017] Furthermore, the above shape loss is specifically: ,in: , ; In the formula, represents the shape loss, represents the adjustment parameter in the shape loss function, Represents the width of the target box in the image data after data annotation. represents the width loss weight, represents the height loss weight, Indicates the width of the target box in the prediction result output by the improved YOLOv10 model. Indicates the height of the target box in the prediction result output by the improved YOLOv10 model. Represents the height of the target box in the image data after data annotation.

[0018] Furthermore, the above IoU loss is specifically: ; In the formula, Represents the border information of the target box in the image data after data annotation. Represents the bounding box information of the target box in the prediction result output by the improved YOLOv10 model. represents the IoU loss.

[0019] In a second aspect, the present application provides a road hazard detection method system, which is applied to a road hazard detection method in any one of the first aspects, including: A data set construction module is used to collect image data covering various types of road disasters, annotate each image data, and construct data samples through each annotated image data; The model optimization training module is used to train the improved YOLOv10 model using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and the improved YOLOv10 model whose loss function reaches the preset end condition is determined as a disaster detection model, wherein: The improved YOLOv10 model includes FM module and EMA module, and the loss function is calculated by angle loss, distance loss, shape loss and IoU loss; The disaster detection and assessment module is used to obtain the image to be tested of the road surface to be tested, and input the image to be tested into the disaster detection model for processing to obtain the disaster detection result of the road surface to be tested. The disaster detection result includes the disaster type and disaster location.

[0020] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, a road hazard detection method according to any one of the first aspects is implemented.

[0021] In a fourth aspect, the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute a road hazard detection method according to any one of the first aspects.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: In the present application, firstly, image data covering various types of road disasters are collected, and data annotation is performed in each image data to construct a data sample; secondly, the improved YOLOv10 model is trained using the training set in the constructed data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, thereby determining the improved YOLOv10 model whose loss function reaches the preset end condition as the disaster detection model; finally, the acquired image to be tested is input into the disaster detection model for processing, and finally the disaster detection result of the road to be tested is obtained.

[0023] In this application, various road disaster types such as cracks, potholes, water accumulation, and oil stains are detected, and it can also be expanded to other potential new disaster types; by covering different traffic scenarios (such as highways, urban roads) and environmental conditions (such as sunny days, rainy days, and insufficient light), the wide applicability of the system is improved; at the same time, through the improved YOLOv10 model, the FM module, the EMA module, and the improved loss function are introduced to optimize the model, enhance the small target detection capability, and optimize the performance under complex backgrounds, thereby greatly reducing the missed detection rate and improving the overall detection performance, thereby solving the current problems of insufficient detection accuracy and poor robustness when facing various road disaster types, obvious scale changes, and complex environments.

[0024] In this application, multimodal feature fusion and dynamic modulation mechanisms are utilized to significantly enhance the stability and robustness of the system in complex traffic scenarios such as illumination changes and target occlusion, ensuring high performance in a variety of environments. At the same time, lightweight modules and optimized algorithms are used to reduce hardware resource requirements while ensuring real-time operation capabilities, supporting efficient deployment on edge devices (such as vehicle-mounted terminals) to meet the needs of intelligent traffic management. During implementation, standardized interface design and modular architecture can be used to support multiple input data sources (such as drone video streams, vehicle-mounted cameras), and the system can quickly adapt to future new road disaster detection needs. It can also provide data support for intelligent traffic management and road maintenance, improve overall operational efficiency, and reduce labor costs.

[0025] In this application, after replacing the original SPPF module in the model with the FM module, the model can significantly improve the modeling ability and detection performance of multi-scale features. Different from the fixed pooling operation of the SPPF module, the Focal Modulation (FM module) dynamically aggregates contextual information from local to global through multi-level deep convolution, which can more accurately capture the detailed features of small targets and strengthen the global semantic understanding under complex backgrounds. This design can show higher robustness for small target detection (such as fine cracks) and complex scenes (such as lighting changes or target occlusion); at the same time, the FocalModulation module adopts a lightweight computing structure, avoiding the high computational complexity problem that may be caused by SPPF, and optimizing computing efficiency through deep convolution and linear transformation, which not only reduces the hardware resource requirements, but also improves the reasoning speed; overall, Focal Modulation surpasses SPPF in detection accuracy, scene adaptability and efficiency, providing a more efficient and robust solution for road disaster detection.

[0026] In this application, the EMA module has significant advantages in multi-scale feature modeling and computational efficiency; by introducing multi-level context aggregation and dynamic modulation mechanisms, the EMA module can simultaneously capture local details and global context information, making it perform better in small target detection and complex scenes; in addition, the EMA module uses group normalization and lightweight deep convolution to avoid the complexity of global attention calculation in PSA, while reducing computing resource requirements and improving detection efficiency, making it more suitable for actual deployment; compared with the PSA module, the dynamic feature modulation mechanism of the EMA module enables it to have stronger adaptability and show higher robustness in the face of lighting changes, occlusion or complex backgrounds. By enhancing the fine feature extraction and global semantic understanding of small targets, EMA improves the detection accuracy and recall rate of the model. These improvements provide YOLOv10 with stronger performance and higher application value in tasks such as road disaster detection.

[0027] In this application, the improved loss function enables the model to show a clearer optimization direction and significantly higher convergence efficiency in bounding box optimization; SIoU introduces angle cost and shape alignment mechanism to dynamically adjust the gradient optimization path of the bounding box in the horizontal and vertical directions, effectively reducing unnecessary "wandering" phenomenon during training. This design enables the prediction box to approach the center of the target box more quickly, thereby accelerating training convergence, and significantly improving the model's adaptability to complex target shapes, especially in scenes with large aspect ratio differences or obvious deviations in target positions; Compared with the CIoU loss function, the improved SIoU loss function shows higher positioning accuracy and recall rate in small target detection and complex backgrounds; its dynamic angle guidance strategy significantly reduces the conflict of multi-component optimization, while the shape cost effectively solves the shortcomings of the predicted box and the real box in aspect ratio matching. This multi-level optimization not only enables the model to more accurately capture the detailed features of small targets (such as road cracks), but also enhances the robustness and generalization ability in multi-target detection scenarios; by optimizing training efficiency and improving final performance, SIoU becomes a more efficient and stable solution for target detection tasks. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings: Figure 1 A method flow chart of the detection method in an embodiment of the present invention; Figure 2 A connection diagram of a detection system in an embodiment of the present invention; Figure 3 Schematic diagram of the connection of electronic equipment in an embodiment of the present invention. DETAILED DESCRIPTION

[0029] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0030] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0031] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0032] In the description of the embodiments of the present invention, "plurality" means at least 2.

[0033] Embodiment 1: In order to solve the problems of insufficient detection accuracy and poor robustness when facing various types of road disasters, obvious scale changes and complex environments, this embodiment provides a road disaster detection method, such as Figure 1 As shown, the following specific steps are included: S1, collect image data covering various types of road disasters, perform data annotation in each image data, and construct data samples through each data annotated image data.

[0034] Among them, during the data collection stage, image data covering various road disaster types such as cracks, potholes, water accumulation, oil pollution, etc. can be collected from a variety of data sources (such as drone aerial photography, road monitoring cameras, and vehicle-mounted cameras); these collection scenes cover different lighting conditions (such as day and night), weather environments (such as sunny days and rainy days), and road types (such as highways and city streets) to ensure the diversity and representativeness of the data.

[0035] Furthermore, in the data labeling stage, professional labeling tools are used to annotate the bounding boxes of the collected images, including the disaster type and location. After the labeling is completed, the dataset is divided into training and test sets, and converted into a standard format suitable for the YOLOv10 model to provide high-quality input for subsequent model training.

[0036] S2, training the improved YOLOv10 model using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and determining the improved YOLOv10 model whose loss function reaches the preset end condition as a disaster detection model, wherein: The improved YOLOv10 model includes FM module and EMA module, and the loss function is calculated by angle loss, distance loss, shape loss and IoU loss.

[0037] Specifically, after replacing the original SPPF module in the model with the FM module, the model can significantly improve the modeling ability and detection performance of multi-scale features. Different from the fixed pooling operation of the SPPF module, the Focal Modulation (FM module) dynamically aggregates contextual information from local to global through multi-level deep convolution, which can more accurately capture the detailed features of small targets and enhance the global semantic understanding under complex backgrounds. This design can show higher robustness for small target detection (such as subtle cracks) and complex scenes (such as lighting changes or target occlusion).

[0038] Optionally, the FM module includes a context aggregation unit and a feature modulation unit, wherein: A context aggregation unit is used to generate a multi-level feature set from each data in the training set or the image to be tested; The feature modulation unit is used to adjust the fine granularity of each feature in the feature set through a dynamic modulator.

[0039] Specifically, in the context aggregation unit, the input features X Performing multi-layer context aggregation can be expressed as: , ,in, represents a linear transformation, used for feature space projection, It is a deep convolution module used to extract contextual information of different scales. L The depth of the representation layer, Indicates The feature output of the layer; the last layer of features is globally averaged pooled to obtain the global features: ; Finally, a multi-level feature set is generated: .

[0040] Furthermore, in the feature modulation unit, the gating mechanism is used to dynamically adjust the multi-level features and aggregate them into a modulator M , modulator M , through linear transformation Obtained, can be expressed as: , inject the modulator into each feature, i.e.: , where represents the final output feature, represents the linear projection of the query features, represents element-by-element multiplication; in the above formula, , ; Represents a linear transformation used to generate gating weights .

[0041] The above process extracts multi-level features through context aggregation and combines dynamic modulators to achieve fine-grained adjustment of features, which is suitable for efficient feature processing and enhancement in various visual tasks.

[0042] Furthermore, compared with the PSA module, the dynamic feature modulation mechanism of the EMA module enables it to have stronger adaptability and show higher robustness in the face of lighting changes, occlusion or complex backgrounds. By enhancing the fine feature extraction and global semantic understanding of small targets, EMA improves the detection accuracy and recall rate of the model. These improvements provide YOLOv10 with stronger performance and higher application value in tasks such as road hazard detection.

[0043] Optionally, the EMA module includes a feature grouping unit, a parallel sub-network unit and a cross-space learning unit, wherein: A feature grouping unit, used to disperse the individual features in the fine-grained adjusted feature set into different subgroups; Parallel sub-network units are used to obtain the global spatial information and multi-scale local features of each feature in each sub-group through different multi-scale branch models. The global spatial information is extracted by 1D global average pooling operations in the horizontal and vertical directions, and the multi-scale local features are extracted by convolution operations. The cross-space learning unit is used to combine the global spatial information and multi-scale local features of each feature in each subgroup through a matrix dot product operation to obtain a weighted aggregate feature of each feature in each subgroup.

[0044] Among them, the EMA input feature map can be expressed as: , G Here it is represented as the number of groups. The feature grouping method can disperse the channel features into different subgroups, so that each group of sub-features has independent semantic information, while reducing the computational complexity and providing a solid foundation for subsequent attention modeling.

[0045] Furthermore, after feature grouping, EMA uses a parallel subnetwork structure to model features at multiple scales, including a 1×1 branch and a 3×3 branch. In the 1×1 branch, global spatial information is extracted through 1D global average pooling operations in the horizontal and vertical directions, defined as follows: ; ; In the formula, represents the features aggregated horizontally in the cth channel, represents the features aggregated along the vertical direction in the cth channel. These aggregated features are further processed by a shared 1×1 convolution to capture local interaction information across channels.

[0046] On the other hand, the 3×3 branch uses a 3×3 convolution operation to extract multi-scale local features, thereby enhancing the receptive field of the model and improving the expressiveness of spatial features. Finally, the features generated by the two branches are integrated into one through a fusion operation, achieving comprehensive modeling of short-range and long-range dependencies.

[0047] Finally, the EMA module fuses the outputs of the two branches through a cross-space learning method to further improve the quality of feature representation. First, 2D global average pooling is performed on the outputs of the 1×1 branch and the 3×3 branch to encode global spatial information: ; Then, the results of the two branches are combined through the matrix dot product operation (dot-product) to generate cross-space attention weights. These weights are normalized by the Sigmoid function and finally used for weighted aggregation of the grouped features. The calculation formula is as follows: ; In the formula, is the Sigmoid function, Represents an element-by-element weighted operation; in this way, EMA establishes strong dependencies in the feature space and channel dimensions, and maintains the same size output as the input feature. The multi-scale design of EMA enables it to capture global semantic information and local detail information at the same time, thereby significantly improving model performance while maintaining computational efficiency, and is suitable for a variety of computer vision tasks.

[0048] Furthermore, compared with the CIoU loss function, the improved SIoU loss function shows higher positioning accuracy and recall rate in small target detection and complex backgrounds; its dynamic angle guidance strategy significantly reduces the conflict of multi-component optimization, while the shape cost effectively solves the lack of aspect ratio matching between the predicted box and the real box. This multi-level optimization not only enables the model to more accurately capture the detailed features of small targets (such as road cracks), but also enhances the robustness and generalization ability in multi-target detection scenarios; by optimizing training efficiency and improving final performance, SIoU becomes a more efficient and stable solution for target detection tasks.

[0049] Optionally, the above loss function is specifically: ; In the formula, represents the loss function, represents the IoU loss, Indicates distance loss, and distance loss Obtained through angle loss calculation, Represents shape loss.

[0050] Among them, the SIoU loss function consists of four cost functions: angle loss, distance loss, shape loss, and IoU loss.

[0051] Optionally, the above distance loss is specifically: ,in: , ; ; In the formula, Represents the distance loss, It represents the X-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. It represents the Y-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. Represents the center point of the target box in the image data after data annotation x Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model x Coordinate values, Indicates the width range of the target box in the data annotation and the target box in the prediction result. Indicates the height range of the target box in the data annotation and the target box in the prediction result. represents the comprehensive weight factor, Represents the center point of the target box in the image data after data annotation y Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model y Coordinate value; represents the angle loss, and the angle loss is specifically: , ;in: ; ; In the formula, represents the normalized relative height in the angle loss formula, Indicates the height range of the target box in the data annotation and the target box in the prediction result. Represents the Euclidean distance between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model.

[0052] Optionally, the above shape loss is specifically: ,in: , ; In the formula, represents the shape loss, Represents the adjustment parameter in the shape loss function, which is used to control the weight of the shape loss, that is, the influence of the deviation in shape between the prediction box and the annotation box on the total loss. The recommended value of is usually between 2 and 6, with the optimal value close to 4. Represents the width of the target box in the image data after data annotation. Represents the width loss weight, which is used to measure the difference in width between the predicted box and the labeled box. Represents the height loss weight, which is used to measure the difference in height between the predicted box and the labeled box. Indicates the width of the target box in the prediction result output by the improved YOLOv10 model. Indicates the height of the target box in the prediction result output by the improved YOLOv10 model. Represents the height of the target box in the image data after data annotation.

[0053] Optionally, the above IoU loss is specifically: ; In the formula, Represents the border information of the target box in the image data after data annotation. Represents the bounding box information of the target box in the prediction result output by the improved YOLOv10 model. represents the IoU loss.

[0054] The improved YOLOv10 model was trained with a high-quality dataset, using data augmentation techniques (such as rotation, scaling, and flipping) to improve the generalization of the model, and optimizing hyperparameters (such as learning rate and batch size) to ensure efficient training. The model was validated on the test set, and its detection accuracy and robustness were fully analyzed through evaluation indicators (such as IoU and mAP) to ensure reliable performance in various scenarios.

[0055] S3, obtaining an image of the road to be detected, and inputting the image to be detected into a disaster detection model for processing to obtain a disaster detection result of the road to be detected, wherein the disaster detection result includes a disaster type and a disaster location.

[0056] Among them, the optimized YOLOv10 model is deployed to edge devices (such as roadside units or vehicle-mounted terminals) to realize real-time detection of road disasters. The system can quickly identify and locate various types of disasters, generate alarm information and push it to the management platform; the detected disaster data is classified and stored to support maintenance decisions, and the application effect of the system in the actual road environment is evaluated through periodic analysis of the detection results.

[0057] Specifically, when performing pavement crack detection, the following steps can be followed: 1. The system collects a real-time video stream of a section of highway, inputs the collected video frames into the YOLOv10 model, and detects the location and size of pavement cracks. 2. The system generates alarm information based on the detection results, including the location coordinates and size of the cracks. 3. The system sends the detection results to the maintenance management platform through the wireless communication module, generates maintenance suggestions, and prompts relevant departments to make timely repairs. 4. The management platform stores the detection data for subsequent trend analysis and evaluation.

[0058] Furthermore, when conducting multi-disaster type detection, the following steps can be followed: 1. In urban road scenarios, the system collects video data from drone aerial photography and inputs the optimized YOLOv10 model to detect potholes, water accumulation and oil pollution on the road; 2. The system sorts the candidate detection results and selects the optimal detection results based on the different detection priorities of each disaster; 3. The system generates a report, which includes the specific distribution, area and severity of each disaster type, and simultaneously pushes it to the urban traffic management department; 4. The detection data is used to plan emergency response plans and resource allocation.

[0059] Furthermore, when conducting all-weather road surface inspections, the following steps can be followed: 1. When lighting conditions change (such as from day to night), the system continuously collects road surface videos through the on-board camera and inputs the data into the YOLOv10 model; 2. The system automatically adjusts parameters to adapt to changes in lighting conditions to ensure that detection accuracy is not affected by the environment; 3. The detection results are directly fed back to the driver and uploaded to the cloud platform for real-time monitoring and analysis by relevant departments; 4. The system uses historical detection data to compare with current results, analyzes the changing trends of road hazards, and provides long-term maintenance recommendations.

[0060] In summary, this application proposes an efficient and accurate road hazard detection system based on the optimization of the YOLOv10 model; by introducing the FocalModulation module and the EMA module, the system significantly improves the small target detection capability and performance in complex scenes. The use of the SIoU loss function further optimizes the positioning accuracy of the bounding box and the training convergence efficiency.

[0061] Embodiment 2: This embodiment of the present application provides a road hazard detection system, which is applied to a road hazard detection method in any one of Embodiment 1, such as Figure 2 As shown, including: A data set construction module is used to collect image data covering various types of road disasters, annotate each image data, and construct data samples through each annotated image data; The model optimization training module is used to train the improved YOLOv10 model using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and the improved YOLOv10 model whose loss function reaches the preset end condition is determined as a disaster detection model, wherein: The improved YOLOv10 model includes FM module and EMA module, and the loss function is calculated by angle loss, distance loss, shape loss and IoU loss; The disaster detection and assessment module is used to obtain the image to be tested of the road surface to be tested, and input the image to be tested into the disaster detection model for processing to obtain the disaster detection result of the road surface to be tested. The disaster detection result includes the disaster type and disaster location.

[0062] Specifically, it adopts a modular design, including a dataset construction module, a model optimization training module, and a disaster detection and assessment module; through standardized interfaces, the system supports multiple data source inputs (such as drone video streams and vehicle-mounted cameras), and can be quickly expanded to adapt to new disaster types and detection needs; in practical applications, through lightweight model design and optimization algorithms, real-time detection and efficient deployment are achieved, suitable for running on edge devices, such as vehicle-mounted terminals and roadside units. At the same time, the system shows strong adaptability and robustness to complex environments such as lighting changes and target occlusion; at the same time, the detection results generated by the system not only support alarm information and real-time feedback, but also provide long-term trend analysis, providing decision support for intelligent traffic management, road maintenance and infrastructure detection, and have important practical value.

[0063] Embodiment 3: This embodiment of the present application provides an electronic device, such as Figure 3 As shown, it includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a road hazard detection method of any one of the embodiments 1 is implemented.

[0064] Embodiment 4: The embodiment of the present application provides a non-transitory computer-readable storage medium, which stores computer instructions, and the computer instructions enable a computer to execute a road hazard detection method according to any one of Embodiment 1.

[0065] The above specific implementation methods further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific implementation methods of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A road disaster detection method, characterized in that: The specific steps include: Collect image data covering various types of road disasters, perform data annotation in each of the image data, and construct data samples through each of the data-annotated image data; The improved YOLOv10 model is trained using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and the improved YOLOv10 model whose loss function reaches the preset end condition is determined as a disaster detection model, wherein: The improved YOLOv10 model includes an FM module and an EMA module, and the loss function is obtained by calculating angle loss, distance loss, shape loss and IoU loss; An image of the road to be detected is obtained, and the image is input into the disaster detection model for processing to obtain a disaster detection result of the road to be detected, wherein the disaster detection result includes a disaster type and a disaster location.

2. A road hazard detection method according to claim 1, characterized in that: The FM module includes a context aggregation unit and a feature modulation unit, wherein: The context aggregation unit is used to generate a multi-level feature set from each data in the training set or the image to be tested; The feature modulation unit is used to adjust the fine granularity of each feature in the feature set through a dynamic modulator.

3. A road hazard detection method according to claim 2, characterized in that: The EMA module includes a feature grouping unit, a parallel sub-network unit and a cross-space learning unit, wherein: The feature grouping unit is used to distribute each feature in the feature set that has been fine-grained adjusted into different subgroups; The parallel sub-network unit is used to respectively obtain the global spatial information and multi-scale local features of each feature in each sub-group through different multi-scale branch models, wherein the global spatial information is extracted through 1D global average pooling operations in the horizontal direction and the vertical direction, and the multi-scale local features are extracted through convolution operations; The cross-space learning unit is used to combine the global spatial information and multi-scale local features of each feature in each subgroup by means of a matrix dot product operation to obtain a weighted aggregation feature of each feature in each subgroup.

4. A road hazard detection method according to any one of claims 1 to 3, characterized in that: The loss function is specifically: ; In the formula, represents the loss function, represents the IoU loss, Indicates distance loss, and distance loss Obtained through angle loss calculation, Represents shape loss.

5. A road hazard detection method according to claim 4, characterized in that: The distance loss is specifically: ,in: , ; ; In the formula, Represents the distance loss, It represents the X-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. It represents the Y-axis distance component between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model. Represents the center point of the target box in the image data after data annotation x Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model x Coordinate values, Indicates the width range of the target box in the data annotation and the target box in the prediction result. Indicates the height range of the target box in the data annotation and the target box in the prediction result. represents the comprehensive weight factor, Represents the center point of the target box in the image data after data annotation y Coordinate values, Represents the center point of the target box in the prediction result output by the improved YOLOv10 model y Coordinate value; represents the angle loss, and the angle loss is specifically: , ;in: ； ； In the formula, represents the normalized relative height in the angle loss formula, Indicates the height range of the target box in the data annotation and the target box in the prediction result. Represents the Euclidean distance between the center point of the target box in the image data after data annotation and the center point of the target box in the prediction result output by the improved YOLOv10 model.

6. A road hazard detection method according to claim 4, characterized in that: The shape loss is specifically: ,in: , ; In the formula, represents the shape loss, represents the adjustment parameter in the shape loss function, Represents the width of the target box in the image data after data annotation. represents the width loss weight, represents the height loss weight, Indicates the width of the target box in the prediction result output by the improved YOLOv10 model. Indicates the height of the target box in the prediction result output by the improved YOLOv10 model. Represents the height of the target box in the image data after data annotation.

7. A road hazard detection method according to claim 4, characterized in that: The IoU loss is specifically: ; In the formula, Represents the border information of the target box in the image data after data annotation. Represents the bounding box information of the target box in the prediction result output by the improved YOLOv10 model. represents the IoU loss.

8. A road hazard detection system, applied to a road hazard detection method according to any one of claims 1 to 7, characterized in that: include: A data set construction module is used to collect image data covering various types of road disasters, perform data annotation in each of the image data, and construct data samples through each of the data-annotated image data; The model optimization training module is used to train the improved YOLOv10 model using the training set in the data sample until the loss function of the improved YOLOv10 model reaches a preset end condition, and the improved YOLOv10 model whose loss function reaches the preset end condition is determined as a disaster detection model, wherein: The improved YOLOv10 model includes an FM module and an EMA module, and the loss function is obtained by calculating angle loss, distance loss, shape loss and IoU loss; The disaster detection and assessment module is used to obtain the image to be tested of the road surface to be tested, and input the image to be tested into the disaster detection model for processing to obtain the disaster detection result of the road surface to be tested, and the disaster detection result includes the disaster type and disaster location.

9. An electronic device, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, a road hazard detection method according to any one of claims 1 to 7 is implemented.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, and the computer instructions enable a computer to execute a road hazard detection method as described in any one of claims 1-7.

Images