End-edge-cloud cooperative road disease detection method and system

Abstract

The invention discloses an end-edge-cloud cooperative road disease detection method and system. The method comprises the following steps: dynamically calculating a confidence coefficient threshold value and a detection interval according to a vehicle speed; a vehicle-mounted camera is adopted to collect road images; deploying a lightweight detection model based on a vehicle-mounted NPU / GPU and taking the lightweight detection model as an edge node, and performing ROI extraction and disease feature retrieval on the road image according to a confidence coefficient threshold to realize disease preliminary screening; transmitting the extracted ROI to a cloud server cluster, performing reinspection by a full-precision model, and outputting a disease fine inspection result; and the server cluster aggregates the model update parameters of the edge nodes through a federated learning algorithm, generates a global model and issues the global model to the edge nodes for dynamic optimization. According to the technical scheme, the contradiction between detection efficiency and precision in a high-speed scene is solved, the model generalization ability is improved, efficient detection of road diseases is achieved, and the problems that in the prior art, the data transmission amount is large, the real-time performance is poor, and the calculation power is wasted are solved.

Description

Technical Field

[0001] The present invention relates to the field of image analysis technology, and in particular to an end-edge-cloud collaborative road damage detection method and an end-edge-cloud collaborative road damage detection system. Background Art

[0002] Currently, the traditional approach to road defect detection involves inspecting vehicles and uploading full-resolution images (e.g., 12MP) captured by onboard cameras directly to the cloud for inspection. However, this involves enormous data transfer volumes (a single hour of inspection generates 1GB of data), high 5G data costs, and poor real-time performance (latency > 200ms). Furthermore, this approach is highly dependent on the cloud, and in complex road conditions (such as tunnels and weak signal areas), cloud disconnection can lead to interrupted detection, resulting in insufficient robustness.

[0003] In addition, there is widespread waste of computing power in the hardware computing resources of existing automotive embedded devices (such as Jetson AGX).

[0004] The existing technology discloses a road defect detection method based on an end-cloud collaborative architecture, but does not design dynamic ROI extraction and edge caching for road defect detection. The general architecture cannot meet the real-time requirements. Summary of the Invention

[0005] In response to the above problems, the present invention provides an end-edge-cloud collaborative road defect detection method and system. Through a three-layer collaborative architecture of end-side preprocessing-edge preliminary screening-cloud refinement, and through dynamic adjustment of the confidence threshold based on vehicle speed, task layered diversion processing is realized, which solves the contradiction between detection efficiency and accuracy in high-speed scenarios, and realizes rapid preliminary screening and retrieval of high-frequency defects through edge caching. Combined with federated learning, a detection-learning-optimization closed loop is formed, which improves the model generalization ability and realizes efficient detection of road defects. In addition, it provides hardware acceleration in combination with on-board heterogeneous hardware to solve the problems of large data transmission volume, poor real-time performance and waste of computing power in the existing technology.

[0006] To achieve the above objectives, the present invention provides an end-edge-cloud collaborative road damage detection method, comprising:

[0007] Obtain vehicle speed information and dynamically calculate confidence thresholds and detection intervals based on vehicle speed;

[0008] Using a vehicle-mounted camera to collect road images of the road section to be inspected, and preprocessing the road images;

[0009] Deploy lightweight detection models based on vehicle-mounted NPU / GPU and use them as edge nodes.

[0010] Performing ROI extraction and disease feature retrieval on the road image according to the confidence threshold using the lightweight detection model to achieve initial disease screening;

[0011] The extracted ROI is transmitted to the cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs the disease inspection results;

[0012] The server cluster aggregates the model update parameters of the edge nodes through a federated learning algorithm, generates a global model, and sends it to the edge nodes to dynamically optimize the lightweight detection model.

[0013] In the above technical solution, preferably, the specific process of dynamically calculating the confidence threshold and the detection interval based on the vehicle speed includes:

[0014] The confidence threshold is calculated based on the acquired vehicle speed information. The calculation formula is: confidence threshold = 0.8 + V / 100, where the unit of vehicle speed V is km / h;

[0015] Based on the acquired vehicle speed information, the detection interval i of the vehicle camera is calculated according to the following formula:

[0016]

[0017] Where d is the detection distance in meters, v is the current vehicle speed in km / h, and f is the video frame rate in fps.

[0018] In the above technical solution, preferably, the specific process of preprocessing the road image includes:

[0019] Using OpenCV to perform non-local mean filtering and distortion correction on the road image;

[0020] The road image is downsampled by a preset ratio.

[0021] In the above technical solution, preferably, the lightweight detection model is used to perform ROI extraction and disease feature retrieval on the road image according to the confidence threshold to achieve initial disease screening. The specific process includes:

[0022] Deploy the pruned YOLOv8n-pruned model on the vehicle's NPU / GPU as a lightweight detection model;

[0023] According to the confidence threshold obtained by dynamic calculation, the lightweight detection model performs ROI extraction on the road image;

[0024] According to the high-frequency disease feature library pre-stored in the edge node, the extracted ROI is subjected to feature retrieval in the high-frequency disease feature library according to the Hamming distance to achieve preliminary disease screening;

[0025] The ROI with high-frequency diseases obtained in the initial screening is output to the server cluster. If the initial screening process does not match, all the extracted ROIs are output to the server cluster.

[0026] In the above technical solution, preferably, during the ROI extraction process, the detection box is ⊕(width×20%, height×20%), where ⊕ represents a boundary extension operation, width represents the width of the ROI detection box, and height represents the height of the ROI detection box.

[0027] In the above technical solution, preferably, during the execution of the lightweight detection model, the YOLOv8n-pruned model file is loaded using the TensorRT engine, the optimization parameters are configured through the builder, the builder is called to generate the optimized TensorRT engine, and the engine is directly deployed in the vehicle NPU / GPU for operation;

[0028] The TensorRT engine initializes the YOLOv8n-pruned model and identifies the convolutional layer and the pooling layer;

[0029] The convolution layer is optimized using the CUDA programming model, and the convolution operation is decomposed into multiple parallel thread blocks. For the pooling layer, the pooling task is directed to the NPU for execution through the TensorRT engine plug-in mechanism or by directly calling the API of the on-board NPU.

[0030] In the above technical solution, preferably, the high-frequency disease feature library in the edge node is a SimHash fingerprint library formed by hashing feature vectors of typical road diseases that occur frequently into binary fingerprints.

[0031] In the above technical solution, preferably, the extracted ROI is transmitted to a cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs a precise disease inspection result. The specific process includes:

[0032] The server cluster runs the full-precision model of YOLOv8x to perform precise disease inspection on the ROI obtained from the initial screening;

[0033] According to the preset result template, the disease inspection results including disease type, size and severity level are output.

[0034] In the above technical solution, preferably, the server cluster aggregates the model update parameters of the edge nodes through a federated learning algorithm, generates a global model, and sends it to the edge nodes to dynamically optimize the lightweight detection model. The specific process includes:

[0035] The edge node transmits the parameters of the lightweight detection model to the server cluster, and the server cluster aggregates the model parameters of each edge node;

[0036] A global model is generated by performing weighted averaging on the model parameters, and the model parameters of the global model are sent down to the edge nodes, so as to dynamically optimize the lightweight detection model deployed on all edge nodes.

[0037] The present invention further proposes an end-edge-cloud collaborative road defect detection system, which applies the end-edge-cloud collaborative road defect detection method disclosed in any of the above technical solutions, including:

[0038] The vehicle speed dynamic detection module is used to obtain the vehicle speed information and calculate the confidence threshold and detection interval based on the vehicle speed dynamics;

[0039] A road image acquisition module is used to acquire road images of the road section to be inspected using a vehicle-mounted camera and to pre-process the road images;

[0040] A disease feature initial screening module is used to deploy a lightweight detection model based on the vehicle's NPU / GPU and serve as an edge node. The lightweight detection model is used to perform ROI extraction and disease feature retrieval on the road image according to the confidence threshold to achieve initial disease screening;

[0041] The cloud-based disease precision inspection module is used to transmit the extracted ROI to the cloud-based server cluster, where the full-precision model run by the server cluster re-inspects the ROI and outputs the disease precision inspection results;

[0042] A federated node update module is used for the server cluster to aggregate the model update parameters of the edge nodes through a federated learning algorithm, generate a global model and send it to the edge nodes to dynamically optimize the lightweight detection model.

[0043] Compared with the existing technology, the beneficial effects of the present invention are: through the three-layer collaborative architecture of end-side preprocessing-edge preliminary screening-cloud-based refinement, and through the dynamic adjustment of the confidence threshold based on vehicle speed, task hierarchical diversion processing is realized, which solves the contradiction between detection efficiency and accuracy in high-speed scenarios, and realizes rapid preliminary screening and retrieval of high-frequency diseases through edge caching. Combined with federated learning, a detection-learning-optimization closed loop is formed, which improves the model generalization ability and realizes efficient detection of road diseases. In addition, combined with on-board heterogeneous hardware to provide hardware acceleration, it solves the problems of large data transmission volume, poor real-time performance and waste of computing power in the existing technology. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic diagram of the system architecture of a device-edge-cloud collaborative road defect detection method disclosed in one embodiment of the present invention;

[0045] Figure 2 A schematic diagram of dynamic task offloading for a device-edge-cloud collaborative road defect detection method disclosed in an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of edge node hardware acceleration for an end-edge-cloud collaborative road defect detection method disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying 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. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0048] The present invention is described in further detail below with reference to the accompanying drawings:

[0049] like Figure 1 As shown, a device-edge-cloud collaborative road defect detection method provided by the present invention includes:

[0050] Obtain vehicle speed information and dynamically calculate confidence thresholds and detection intervals based on vehicle speed;

[0051] Use the vehicle-mounted camera to collect the road image of the road section to be inspected and pre-process the road image;

[0052] Deploy lightweight detection models based on vehicle-mounted NPU / GPU and use them as edge nodes.

[0053] A lightweight detection model is used to extract ROIs and retrieve disease features from road images according to confidence thresholds to achieve initial disease screening.

[0054] The extracted ROI is transmitted to the cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs the disease inspection results;

[0055] The server cluster aggregates the model update parameters of the edge nodes through the federated learning algorithm, generates a global model and sends it to the edge nodes to dynamically optimize the lightweight detection model.

[0056] In this implementation, a three-layer collaborative architecture of end-side preprocessing, edge screening, and cloud-based refinement is used, as well as dynamic adjustment of the confidence threshold based on vehicle speed to achieve hierarchical task diversion processing, thereby resolving the contradiction between detection efficiency and accuracy in high-speed scenarios. Edge caching is used to achieve rapid initial screening and retrieval of high-frequency defects, and combined with federated learning to form a detection-learning-optimization closed loop, the model's generalization capability is improved, and efficient detection of road defects is achieved. In addition, hardware acceleration is provided by combining on-board heterogeneous hardware, solving the problems of large data transmission volume, poor real-time performance, and waste of computing power in existing technologies.

[0057] Specifically, the end-side vehicle camera collects high-definition road images, capturing one frame every 50ms, for example, with a resolution of 12MP, supporting external triggering and a frame rate of 30FPS. The on-board NPU / GPU serves as the edge node. After the initial disease screening through the lightweight detection model, the cloud-based server cluster performs a detailed and re-inspection of the disease, realizing a three-layer collaborative architecture of end-side preprocessing, edge initial screening, and cloud-based fine-tuning.

[0058] like Figure 2 As shown, in the above embodiment, preferably, the specific process of dynamically calculating the confidence threshold and the detection interval according to the vehicle speed includes:

[0059] The confidence threshold is calculated based on the acquired vehicle speed information. The calculation formula is: confidence threshold = 0.8 + V / 100, where the unit of vehicle speed V is km / h;

[0060] Based on the acquired vehicle speed information, the detection interval i of the vehicle camera is calculated according to the following formula:

[0061]

[0062] Where d is the detection distance in meters, v is the current vehicle speed in km / h, and f is the video frame rate in fps.

[0063] During implementation, the confidence threshold is dynamically adjusted according to vehicle speed (for example, the threshold is increased to 90% at 60km / h and reduced to 80% at 30km / h) to balance the detection accuracy in high-speed scenarios.

[0064] Among them, the detection interval is automatically adjusted according to the vehicle speed. The interval is increased at low speed (repeated recognition) and at high speed (good road quality). The optimal interval is maintained at intermediate speeds to balance detection coverage and detection efficiency.

[0065] Specifically, during implementation, the detection distance is dynamically calculated based on vehicle speed, a minimum speed threshold is set, and the maximum detection interval is limited to avoid excessive frame intervals. When the vehicle speed is below the minimum speed threshold, the minimum speed is used to calculate the frame interval, which can achieve the maximum detection interval.

[0066] In the above embodiment, preferably, the specific process of preprocessing the road image includes:

[0067] Use OpenCV to perform non-local mean filtering and distortion correction on road images;

[0068] The road image is downsampled to a preset ratio.

[0069] Among them, downsampling can reduce the road image to 1 / 4 resolution (320×240), reducing the computational workload of edge nodes.

[0070] like Figure 3 As shown, in the above embodiment, preferably, a lightweight detection model is used to extract ROI and retrieve disease features from the road image according to the confidence threshold to achieve initial disease screening. The specific process includes:

[0071] Deploy the pruned YOLOv8n-pruned model on the vehicle's NPU / GPU as a lightweight detection model;

[0072] According to the confidence threshold calculated dynamically, the lightweight detection model extracts ROI from the road image;

[0073] Based on the high-frequency disease feature library pre-stored in the edge node, such as the SimHash fingerprints of 1,000 typical cracks and pits, the extracted ROI is searched for features in the high-frequency disease feature library according to the Hamming distance. The search time is less than 1ms, achieving initial disease screening.

[0074] The ROI with high-frequency diseases obtained in the initial screening is output to the server cluster. If the initial screening process does not match, all the extracted ROIs are output to the server cluster.

[0075] Among them, the edge cache module stores the SimHash fingerprint library of high-frequency diseases, and gives priority to local rapid retrieval during detection, reducing dependence on the cloud and data transmission volume.

[0076] In the above embodiment, preferably, during the ROI extraction process, the detection box is ⊕(width×20%, height×20%), where ⊕ represents a boundary expansion operation, width represents the width of the ROI detection box, and height represents the height of the ROI detection box, that is, the detection box is expanded by 20% pixels.

[0077] In the above embodiment, preferably, during the execution of the lightweight detection model, the Python or C++ API of the TensorRT engine is used to load the YOLOv8n-pruned model file, such as the ONNX format file, and the optimization parameters are configured through the Builder, including specifying the target platform (JetsonAGX), selecting the precision mode (INT8 priority), and setting the dynamic tensor shape (adapting to different resolution inputs).

[0078] The builder is called to generate an optimized TensorRT engine, which is then deployed directly on the vehicle's NPU / GPU. Actual tests show that when processing 320×240 resolution images, using TensorRT acceleration reduces single-frame inference time from 100ms in the native PyTorch framework to less than 45ms, increasing inference speed by over 55%. Meanwhile, the model's mean average precision (mAP) drops by less than 1%, achieving a balance between efficiency and accuracy.

[0079] The TensorRT engine initializes the YOLOv8n-pruned model and uses TensorRT or a custom scheduling algorithm to analyze the network structure of the YOLOv8n-pruned model and identify the convolutional layers (such as the C3 module) and pooling layers (such as the SPPF module).

[0080] The convolution layer is optimized using the CUDA programming model, decomposing the convolution operation into multiple parallel thread blocks to fully utilize the GPU's large-scale parallel computing capabilities. At the same time, the CUDA kernel function is customized based on the characteristics of the road disease detection model. For example, the Winograd algorithm is used to accelerate the 3×3 convolution operation, which improves the computational efficiency of the convolution layer by 40%. For the pooling layer, the pooling task is assigned to the NPU for execution through the TensorRT engine's plug-in mechanism or by directly calling the on-board NPU API. The NPU uses its built-in tensor processing unit to quickly complete the pooling operation, and the processing speed is 2 times faster than that executed on the GPU. Through this refined task allocation and hardware optimization, the computing power utilization of the entire model has been increased from 60% under the traditional unified scheduling method to 95%, effectively avoiding idle hardware resources or competition.

[0081] In the above implementation, the high-frequency disease feature library in the edge node preferably consists of a SimHash fingerprint library, which is constructed by hashing the feature vectors of frequently occurring typical road diseases into binary fingerprints. During implementation, every time an edge node detects 100 new diseases, it extracts features, generates SimHash fingerprints, and caches them. Simultaneously, it uploads the model fine-tuning parameters to the cloud.

[0082] In the above embodiment, preferably, the extracted ROI is transmitted to a cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs the disease inspection result. The specific process includes:

[0083] The server cluster runs the full-precision YOLOv8x model to perform precise disease inspection on the ROI obtained from the initial screening;

[0084] According to the preset result template, the disease inspection results including disease type, size and severity level are output.

[0085] In the above embodiment, preferably, the server cluster aggregates the model update parameters of the edge nodes through the federated learning algorithm, generates a global model, and sends it to the edge nodes to dynamically optimize the lightweight detection model. The specific process includes:

[0086] The edge nodes transmit the parameters of the lightweight detection model to the server cluster, and the server cluster aggregates the model parameters of each edge node;

[0087] A global model is generated by weighted averaging the model parameters, and the model parameters of the global model are sent to the edge nodes, so as to dynamically optimize the lightweight detection models deployed on all edge nodes.

[0088] In this implementation, the cloud collects gradient updates from edge nodes and generates a global model through weighted averaging. Through federated learning parameter aggregation, the present invention achieves a closed loop of "detection-learning-optimization." Edge nodes continuously accumulate new disease data locally and fine-tune the model, while the cloud regularly aggregates and updates, allowing the global model to continuously adapt to emerging disease characteristics (such as new crack morphologies). This eliminates the need to recollect and upload large amounts of raw data, significantly improving the model's dynamic optimization efficiency and detection accuracy, ensuring the system maintains high performance over the long term.

[0089] During the implementation process, the cloud collects updated parameters of 10 edge nodes every time, triggering federated learning aggregation, generating a new global model and sending it down.

[0090] The present invention further provides an end-edge-cloud collaborative road defect detection system, which applies the end-edge-cloud collaborative road defect detection method disclosed in any of the above embodiments, including:

[0091] The vehicle speed dynamic detection module is used to obtain the vehicle speed information and calculate the confidence threshold and detection interval based on the vehicle speed dynamics;

[0092] The road image acquisition module is used to use the vehicle-mounted camera to collect the road image of the road section to be inspected and pre-process the road image;

[0093] The disease feature initial screening module is used to deploy a lightweight detection model based on the vehicle's NPU / GPU and serve as an edge node. The lightweight detection model extracts ROIs and retrieves disease features from road images according to confidence thresholds to achieve initial disease screening.

[0094] The cloud-based disease precision inspection module is used to transmit the extracted ROI to the cloud server cluster. The full-precision model run by the server cluster re-inspects the ROI and outputs the disease precision inspection results.

[0095] The federated node update module is used by the server cluster to aggregate the model update parameters of the edge nodes through the federated learning algorithm, generate a global model and send it to the edge nodes to dynamically optimize the lightweight detection model.

[0096] According to the end-edge-cloud collaborative road defect detection system disclosed in the above embodiment, the functions to be implemented by each module thereof correspond to the steps of the end-edge-cloud collaborative road defect detection method disclosed in the above embodiment. During the implementation process, operations are performed with reference to the above embodiment, which will not be repeated here.

[0097] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art will readily appreciate that various modifications and variations of the present invention are possible. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A road damage detection method based on device-edge-cloud collaboration, characterized in that: include: Obtain vehicle speed information and dynamically calculate confidence thresholds and detection intervals based on vehicle speed; Using a vehicle-mounted camera to collect road images of the road section to be inspected, and preprocessing the road images; Deploy lightweight detection models based on vehicle-mounted NPU / GPU and use them as edge nodes. Performing ROI extraction and disease feature retrieval on the road image according to the confidence threshold using the lightweight detection model to achieve initial disease screening; The extracted ROI is transmitted to the cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs the disease detection results; The server cluster aggregates the model update parameters of the edge nodes through a federated learning algorithm, generates a global model, and sends it to the edge nodes to dynamically optimize the lightweight detection model.

2. The device-edge-cloud collaborative road damage detection method according to claim 1, characterized in that: The specific process of dynamically calculating the confidence threshold and the detection interval based on the vehicle speed includes: The confidence threshold is calculated based on the acquired vehicle speed information. The calculation formula is: confidence threshold = 0.8 + V / 100; Based on the acquired vehicle speed information, the detection interval i of the vehicle camera is calculated according to the following formula: Where d is the detection distance in meters, v is the current vehicle speed in km / h, and f is the video frame rate in fps.

3. The device-edge-cloud collaborative road damage detection method according to claim 2, characterized in that: The specific process of preprocessing the road image includes: Using OpenCV to perform non-local mean filtering and distortion correction on the road image; The road image is downsampled by a preset ratio.

4. The device-edge-cloud collaborative road damage detection method according to claim 3, characterized in that: The lightweight detection model is used to extract ROI and retrieve disease features from the road image according to the confidence threshold to achieve initial disease screening. The specific process includes: Deploy the pruned YOLOv8n-pruned model on the vehicle's NPU / GPU as a lightweight detection model; According to the confidence threshold obtained by dynamic calculation, the lightweight detection model performs ROI extraction on the road image; According to the high-frequency disease feature library pre-stored in the edge node, the extracted ROI is subjected to feature retrieval in the high-frequency disease feature library according to the Hamming distance to achieve preliminary disease screening; The ROI with high-frequency diseases obtained in the initial screening is output to the server cluster. If the initial screening process does not match, all the extracted ROIs are output to the server cluster.

5. The device-edge-cloud collaborative road damage detection method according to claim 4 is characterized in that: During the ROI extraction process, the detection box is ⊕(width×20%, height×20%), where ⊕ represents the boundary extension operation, width represents the width of the ROI detection box, and height represents the height of the ROI detection box.

6. The device-edge-cloud collaborative road damage detection method according to claim 4, characterized in that: During the execution of the lightweight detection model, the YOLOv8n-pruned model file is loaded using the TensorRT engine, the optimization parameters are configured through the builder, the builder is called to generate the optimized TensorRT engine, and it is directly deployed and run in the vehicle's NPU / GPU; The TensorRT engine initializes the YOLOv8n-pruned model and identifies the convolutional layer and the pooling layer; The convolution layer is optimized using the CUDA programming model, and the convolution operation is decomposed into multiple parallel thread blocks. For the pooling layer, the pooling task is directed to the NPU for execution through the TensorRT engine plug-in mechanism or by directly calling the API of the on-board NPU.

7. The device-edge-cloud collaborative road damage detection method according to claim 4, characterized in that: The high-frequency disease feature library in the edge node is a SimHash fingerprint library formed by hashing the feature vectors of typical road diseases that occur frequently into binary fingerprints.

8. The device-edge-cloud collaborative road damage detection method according to claim 4, characterized in that: The extracted ROI is transmitted to the cloud server cluster, and the full-precision model run by the server cluster re-inspects the ROI and outputs the disease inspection results. The specific process includes: The server cluster runs the full-precision model of YOLOv8x to perform precise disease inspection on the ROI obtained from the initial screening; According to the preset result template, the disease inspection results including disease type, size and severity level are output.

9. The device-edge-cloud collaborative road damage detection method according to claim 8, characterized in that: The server cluster aggregates the model update parameters of the edge nodes through a federated learning algorithm, generates a global model, and sends it to the edge nodes to dynamically optimize the lightweight detection model. The specific process includes: The edge node transmits the parameters of the lightweight detection model to the server cluster, and the server cluster aggregates the model parameters of each edge node; A global model is generated by performing weighted averaging on the model parameters, and the model parameters of the global model are sent down to the edge nodes, so as to dynamically optimize the lightweight detection model deployed on all edge nodes.

10. A device-edge-cloud collaborative road disease detection system, characterized in that: The method for detecting road defects using the device-edge-cloud collaboration according to any one of claims 1 to 9 comprises: The vehicle speed dynamic detection module is used to obtain the vehicle speed information and calculate the confidence threshold and detection interval based on the vehicle speed dynamics; A road image acquisition module is used to acquire road images of the road section to be inspected using a vehicle-mounted camera and to pre-process the road images; A disease feature initial screening module is used to deploy a lightweight detection model based on the vehicle's NPU / GPU and serve as an edge node. The lightweight detection model is used to perform ROI extraction and disease feature retrieval on the road image according to the confidence threshold to achieve initial disease screening; The cloud-based disease precision inspection module is used to transmit the extracted ROI to the cloud-based server cluster, where the full-precision model run by the server cluster re-inspects the ROI and outputs the disease precision inspection results; A federated node update module is used for the server cluster to aggregate the model update parameters of the edge nodes through a federated learning algorithm, generate a global model and send it to the edge nodes to dynamically optimize the lightweight detection model.

Images