Urban road ponding rapid monitoring method and system based on structure re-parameterization

Abstract

The invention provides an urban road ponding rapid monitoring method and system based on structure re-parameterization, and belongs to the technical field of intelligent monitoring and computer vision. The invention aims to solve the problem that edge equipment reasoning speed and complex water body feature recognition precision are difficult to consider in existing ponding monitoring. According to the method, a four-stage structural re-parameterized network is constructed, and a core module of the four-stage structural re-parameterized network comprises parallel 1 * 1, 3 * 3, cavity 3 * 3 and 5 * 5 convolution branches in a training stage so as to fully capture multi-scale and frequency domain characteristics of a ponding area; in the inference stage, the branch parameters are folded and fused into a single equivalent 5 * 5 convolution kernel for operation through mathematical transformation. According to the invention, low-delay and high-throughput real-time monitoring is realized while the detection precision of accumulated water reflection, weak texture and marginal areas is effectively improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and artificial intelligence, and in particular to a semantic segmentation technology based on deep learning, specifically applied in intelligent transportation and urban emergency management scenarios to achieve real-time monitoring and regional positioning of urban road flooding. Background Technology

[0002] Urban road flooding is a frequent urban disaster during heavy rain, seriously threatening traffic safety and urban operational efficiency. Traditional flood monitoring mainly relies on manual patrols or fixed-point water level sensors. The former has poor timeliness and narrow coverage, while the latter has high installation and maintenance costs and difficulty in determining the specific distribution range of floodwater. In recent years, with the development of deep learning technology, computer vision-based flood detection methods have gradually become mainstream. These methods typically use semantic segmentation networks (such as UNet, DeepLab series, or PSPNet) to perform pixel-level classification of surveillance videos or vehicle images. However, existing visual detection solutions face a severe "accuracy-speed" dilemma in practical implementation:

[0003] 1. Computational burden of high-precision models: In order to handle the complex visual features of waterlogged areas (such as specular reflection, reflections, and confusion with wet road surfaces), high-performance models usually use deep networks and complex attention mechanisms, resulting in a large number of model parameters and high inference latency, making it difficult to achieve real-time operation on edge devices with limited computing power (such as roadside computing units and vehicle-mounted chips).

[0004] 2. Insufficient feature extraction in lightweight models: Lightweight networks designed to achieve real-time performance (such as MobileNet and ShuffleNet) often sacrifice speed by drastically reducing the number of channels or simplifying the convolutional structure. This results in a limited receptive field, making it difficult to distinguish between "water reflections" and "real road surfaces," or to accurately define the edges of water accumulation, leading to a large number of false positives or false negatives.

[0005] 3. Limitations of existing reparameterization techniques: Although existing techniques (such as RepVGG and MobileOne) have introduced the concept of structural reparameterization to accelerate inference, their standard designs typically only include 1×1 and 3×3 branches. This general design lacks optical property optimization for the specific transparent medium of "water"—for example, it cannot effectively cross the water surface reflection area with a large receptive field, and it also lacks the ability to specifically suppress high-frequency textures on wet road surfaces. Summary of the Invention

[0006] This application provides a method and system for rapid monitoring of urban road water accumulation based on structural reparameterization, addressing the problem that existing road water accumulation monitoring methods cannot simultaneously achieve both real-time inference speed on edge devices and high-precision recognition of complex optical features of water accumulation (such as strong reflection, weak texture, and blurred edges).

[0007] The core technology of this invention is to construct a four-stage reparameterized neural network. The core is to design a training module that includes four specific parallel branches: 1×1, 3×3, dilated 3×3, and 5×5 large kernel convolution. During the inference stage, the training module is mathematically folded into a single 5×5 convolution kernel. This allows for accurate capture of water accumulation areas without increasing inference time, by utilizing a large receptive field and multi-frequency domain features.

[0008] In a first aspect, this application provides a rapid monitoring method for urban road waterlogging based on structure reparameterization, the method comprising the following steps: Acquire urban road perception data and construct a four-stage structured reparameterization network; The four-stage structure reparameterization network was trained using a training dataset containing annotations of waterlogged areas to obtain the trained network. The urban road perception data to be monitored is input into the trained network for inference, and the water accumulation detection result is output as a pixel-level water accumulation area mask. The four-stage structure reparameterization network consists of stacked structure reparameterization modules. Each structure reparameterization module contains four parallel convolutional branches during the training phase. These four parallel convolutional branches are as follows: 1×1 convolution branch, 3×3 convolution branch, 3×3 dilated convolution branch and 5×5 convolution branch; During the inference phase, the parameters of the four parallel convolutional branches of the structure reparameterization module are folded and fused into a single equivalent 5×5 convolutional kernel for inference using a single equivalent 5×5 convolutional kernel.

[0009] Furthermore, the four-stage reparameterization network consists of a first stage, a second stage, a third stage, and a fourth stage connected in sequence; the first stage is used for low-level feature extraction and resolution preservation; the second stage is used for multi-scale context aggregation; the third stage is used for detail magnification and difference enhancement, using 3×3 dilated convolution branches to distinguish between wet road surfaces and waterlogged areas; and the fourth stage is used for semantic integration and dense prediction.

[0010] Furthermore, the parameter folding and fusion process in the inference stage includes a unified mapping step for convolutional kernels: The kernels of the 1×1 convolutional branches are mapped to a 5×5 size with center alignment, and the remaining positions are filled with zero values; the kernels of the 3×3 convolutional branches are mapped to a 5×5 size with center alignment, and the remaining positions are filled with zero values; the kernels of the 3×3 dilated convolutional branches are mapped to a 5×5 size by inserting zero values ​​between the weight values ​​according to the dilation interval; the kernels of the 5×5 convolutional branches remain unchanged.

[0011] Furthermore, the parameter folding and fusion process in the inference stage also includes a batch normalization layer fusion step: The convolutional layers in each convolutional branch are fused with their subsequent batch normalization layers. Based on the mean, variance, scaling factor, and offset parameters of the batch normalization layers, the kernel weights and biases of the corresponding convolutional layers are updated to obtain the fused convolutional parameters with bias terms.

[0012] Furthermore, the parameter folding and fusion process in the inference phase also includes a final kernel fusion step: The convolutional kernels of the four convolutional branches, after being processed by the unified mapping step and the batch normalization layer fusion step, are summed element-wise to obtain a single equivalent 5×5 convolutional kernel.

[0013] Furthermore, the four-stage reparameterized network also includes a pre-sampling module and a pyramid pooling module; the pre-sampling module is set before the first stage and consists of two consecutive 3×3 convolutional layers, used to downsample the urban road perception data by 2 times; the pyramid pooling module is set after the fourth stage and is used to perform multi-scale semantic aggregation on the output features of the fourth stage. The four-stage reparameterized network has lightweight downsampling or upsampling operations between each stage.

[0014] Furthermore, in the step of training the four-stage structure reparameterization network using a training dataset containing annotations of waterlogged areas: End-to-end training is performed using a composite loss function that combines cross-entropy loss and IoU loss.

[0015] Furthermore, before inputting the urban road perception data to be monitored into the trained network, a preprocessing step is also included: The input urban road perception data is normalized and its resolution is adjusted. Urban road perception data includes street view camera images, dashcam images, or drone overhead views.

[0016] Secondly, this application provides an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the above-described rapid monitoring method for urban road waterlogging.

[0017] Thirdly, this application provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the above-described rapid monitoring method for urban road waterlogging.

[0018] The main contributions and innovations of this invention are as follows: 1. Breakthrough in the "accuracy-speed" performance bottleneck (extremely fast inference): By utilizing structural reparameterization technology, this invention successfully decouples the training architecture from the inference architecture. During the inference phase, the complex four-branch structure is losslessly compressed into a single 5×5 convolutional layer, eliminating the memory access cost caused by multiple branches. This allows the model to maintain its high-complexity feature extraction capability while running on low-computing-power edge devices with extremely high throughput.

[0019] 2. Solves the problem of missed detection caused by "specular reflection" of water accumulation (cross-reflection perception): Addressing the issue that the center of water accumulation often presents a "false road surface" characteristic due to reflections of the sky or vehicles, this invention's unique 5×5 convolutional branch provides a larger effective receptive field than the traditional 3×3 convolution. This allows the network to transcend the central reflection area and directly associate contextual information of the water accumulation boundary, thereby accurately determining the integrity of the water accumulation area and avoiding the "hollow" phenomenon in water accumulation detection.

[0020] 3. Significantly reduced false alarms caused by "wet road surface" (frequency domain feature differentiation): To address the problem of wet road surfaces being easily misjudged as standing water after rain, this invention introduces a 3×3 dilated convolution branch. Utilizing the sparse sampling characteristics of dilated convolution, it responds weakly to high-frequency textures (such as asphalt particles and cracks on wet road surfaces) (due to large differences between adjacent sampling points causing cancellation), but responds strongly to low-frequency smooth regions (such as standing water). This mechanism effectively distinguishes wet road surfaces from actual standing water from a frequency domain perspective, significantly reducing the false alarm rate.

[0021] 4. Flexible edge deployment adaptability: The four-stage network architecture design supports adjusting the number of modules in each stage as needed, and the single-path structure during inference is extremely hardware-friendly (no need for complex custom operator support), which greatly reduces the deployment threshold of the algorithm on various terminals such as urban cameras, drones and patrol cars.

[0022] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description

[0023] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a rapid monitoring method for urban road waterlogging based on structure reparameterization according to an embodiment of this application; Figure 2 This is a block diagram of a four-stage structure reparameterization network according to an embodiment of this application; Figure 3 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with one or more embodiments of this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of one or more embodiments of this specification as detailed in the appended claims.

[0025] It should be noted that the steps of the corresponding methods are not necessarily performed in the order shown and described in this specification in other embodiments. In some other embodiments, the methods may include more or fewer steps than described in this specification. Furthermore, a single step described in this specification may be broken down into multiple steps in other embodiments; and multiple steps described in this specification may be combined into a single step in other embodiments.

[0026] Example 1 This embodiment provides a rapid monitoring method for urban road water accumulation based on structural reparameterization. This method aims to address the challenge of existing deep learning models struggling to balance inference speed and water accumulation detection accuracy on edge devices (especially considering interference from water surface reflection and wet road surfaces).

[0027] like Figure 1 As shown, the method flow of this embodiment mainly includes the following steps: Step S1: Data Acquisition and Preprocessing Collect multi-source sensing data of urban roads, including but not limited to images from roadside surveillance cameras, vehicle dashcams, or drone aerial views. Manually annotate the collected images, marking pixel-level masks of waterlogged areas, and construct a dataset containing training, validation, and test sets.

[0028] Furthermore, to improve the model's adaptability to different environments, the dataset should cover multiple weather conditions (sunny days, rainy days), multiple lighting conditions (day and night, shadows), and multiple viewpoints. Therefore, before inputting the data into the network, preprocessing is performed, including image normalization (such as subtracting the mean and dividing by the standard deviation) and resolution adjustment (resize), and data augmentation operations such as random cropping and flipping can be performed.

[0029] Step S2: Construct a four-stage structure reparameterized network Build as Figure 2 The neural network shown is composed of three parts: a pre-sampling layer, a four-stage feature extraction network, and a pyramid pooling head.

[0030] 1. Pre-sampling layer (Stem Block): like Figure 2 As shown, the network's front end features two consecutive 3×3 convolutional operations. This structure serves as the backbone guiding layer, performing preliminary feature transformations on the original high-resolution input image and achieving a 2x downsampling. This step reduces the feature map resolution while preserving key spatial information, effectively decreasing the computational burden of subsequent stages.

[0031] 2. Four-stage feature extraction network: The network's overall architecture is divided into four stages, designed to simulate the human visual process from texture perception to semantic understanding. Each stage consists of several RepModules (structural reparameterization modules). The RepModule training phase includes parallel 1×1, 3×3, dilated 3×3, and 5×5 convolutional branches, followed by Batch Normalization (BN) and ReLU, and residual connections are used. Stages are connected using lightweight downsampling (e.g., depthwise separable convolutions) or upsampling (e.g., bilinear interpolation + 1×1 convolutions) to maintain efficiency. Details are as follows: Stage 1 (Low-level feature extraction): Contains 4 RepModules, maintaining high resolution, used to extract detailed features such as road surface texture and water ripples.

[0032] Phase 2 (Multi-scale Context Aggregation): Contains 4 RepModules, which further reduce the resolution and expand the receptive field through downsampling.

[0033] Phase 3 (Differentiation Amplification and Detail Enhancement): Contains 5 RepModules. This is the core phase for distinguishing between waterlogged and non-waterlogged key areas (such as wet pavement).

[0034] Phase 4 (Semantic Integration and Dense Prediction): Contains 5 RepModules for abstracting high-level semantic features.

[0035] In summary, the configuration of the number of RepModules in each stage can be referenced in Table 1 below: Table 1. Configuration table of embodiments of structure reparameterization networks

[0036] Inter-stage connections: Lightweight downsampling operations (such as depthwise separable convolutions with a stride of 2) or upsampling operations are set between stages to adjust the resolution.

[0037] Structure Reparameterization Module (RepModule) Design: This is the core component of this invention. During the training phase, the RepModule contains four parallel convolutional branches. The outputs of each branch are summed and then passed through a batch normalization (BN) layer and an activation function (ReLU). The specific configuration and function of the four branches are as follows: 1×1 convolutional branch: used for linear combination of channel information and feature projection.

[0038] 3×3 convolutional branch: used to capture regular local spatial features, i.e. local receptive fields.

[0039] 3×3 Dilated Convolution Branch: Sets the dilation rate (e.g., dilation=2). Its function is to differentiate between "wet road surface" and "water accumulation" by utilizing frequency domain differences, thus expanding the context without significantly increasing parameters. Wet road surfaces typically retain the high-frequency texture of asphalt (graininess, cracks). Due to the skipped sampling points, dilated convolution has a weaker response in high-frequency regions with large differences between adjacent pixels (they cancel each other out); while water accumulation regions mostly have low-frequency smooth signals and a stronger response. This branch effectively reduces false alarms about wet road surfaces after rain.

[0040] 5×5 convolutional branch: A larger receptive field to capture broad-scale water surface features. Its role is "cross-reflection perception." Urban puddles often exhibit specular reflection (reflecting the sky, vehicles, trees), leading to missing textures in the center of the puddles or misjudging them as artifacts. The large 5×5 kernel can cross the central reflection area, directly associating with the contextual information of the puddles' edges, ensuring that the detected puddles area is complete and not hollow.

[0041] 3. Pyramid pooling head: like Figure 2 As shown, a pyramid pooling module is connected after stage 4. This module aggregates global context information through pooling operations at different scales to address the problem of large scale variations in water accumulation areas (from tiny puddles to large-scale flooding). Finally, the feature map processed by this module is used to output a predicted image of the water accumulation area through a convolutional layer.

[0042] Step S3, Model Training The constructed network is trained end-to-end using the dataset prepared in step S1.

[0043] Loss function: A composite loss function of cross-entropy loss and IoU loss is used to simultaneously optimize pixel classification accuracy and region overlap.

[0044] Optimization strategy: Use SGD or Adam optimizers, set an initial learning rate (e.g., 0.01), and adjust the learning rate using a cosine annealing strategy. During training, each parallel branch updates its parameters independently to obtain the maximum feature representation capability (i.e., the benefit of "overparameterization").

[0045] Step S4: Structural reparameterization folding (inference model generation) After training, before entering the inference phase, the network needs to undergo "structural reparameterization," which losslessly compresses the multi-branch structure from training into a single convolutional kernel. This process leverages the linear additivity of convolution operations. Specifically, after model convergence, the convolutional kernels and BN parameters of each branch are folded into a single equivalent convolutional kernel (and corresponding biases) according to mathematical rules, replacing the multi-branch structure for inference, thereby significantly reducing the number of operators and memory accesses. The reparameterization folding steps include unified mapping of convolutional kernels, fusion of convolutional layers and BN layers, and the final kernel addition and fusion process. The specific folding steps are as follows: 1. Fusion of convolutional layers and batch normalization (BN) layers: For each branch (1×1, 3×3, 3×3 with holes, 5×5), its convolutional layer (weights (kernel) W, bias B) is first merged with its subsequent BN layer (mean µ, variance σ, scaling factor γ, offset β). The formulas for calculating the merged weights (kernel) W' and bias B' are as follows:

[0046]

[0047] in To prevent division by zero of small constants.

[0048] 2. Unified mapping of convolution kernels: In order to add kernels of different sizes, they need to be uniformly mapped to the largest kernel size (i.e., 5×5): 1×1 convolution: The weights are placed at the center of a 5×5 matrix (e.g., coordinates (2,2), in other words, they are mapped to the center of the 5×5 kernel in a center-aligned manner), and the rest are filled with 0.

[0049] 3×3 convolution: Place 3×3 weights in the center region of a 5×5 matrix (in other words, map them to the center of the 5×5 kernel in a center-aligned manner), and pad them with 0s.

[0050] 3×3 dilated convolution: Based on the dilation rate (e.g., 2), the 9 values ​​of the 3×3 weights are discretely distributed in the corresponding positions of the 5×5 matrix (e.g., (0,0), (0,2), (0,4)...), and the remaining positions (gaps) are filled with 0.

[0051] 5×5 convolution: Remains unchanged.

[0052] After the above processing, all convolutional kernels are mapped to a uniform size of 5×5.

[0053] 3. Final nuclear fusion: The four 5×5 matrices obtained after fusion and mapping are added element-wise, and the corresponding bias terms are also added to obtain a single equivalent 5×5 convolution kernel W. final and bias B final .

[0054] 4. Model replacement: In the inference model structure, the original RepModule multi-branch structure is replaced with the above single 5×5 convolutional layer.

[0055] Step S5, Inference and Output The image to be monitored is input into the single-path network processed in step S4. Since there is no longer a multi-branch structure in the network, the data stream only needs to undergo a series of efficient 5×5 convolution operations to quickly output the water accumulation area mask.

[0056] Example 2 This embodiment also provides an electronic device, see reference. Figure 3 It includes a memory 404 and a processor 402, wherein the memory 404 stores a computer program and the processor 402 is configured to run the computer program to perform the steps in any of the above method embodiments.

[0057] Specifically, the processor 402 may include a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0058] Memory 404 may include a mass storage device for data or instructions. For example, and not limitingly, memory 404 may include a hard disk drive (HDD), a floppy disk drive, a solid-state drive (SSD), flash memory, an optical disk drive, a magneto-optical disk drive, magnetic tape, or a Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 404 may include removable or non-removable (or fixed) media. Where appropriate, memory 404 may be internal or external to a data processing device. In a particular embodiment, memory 404 is non-volatile memory. In a particular embodiment, memory 404 includes read-only memory (ROM) and random access memory (RAM). Where appropriate, the ROM may be a mask-programmed ROM, a programmable read-only memory (PROM), an erasable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), an electrically alterable read-only memory (EAROM), or flash memory, or a combination of two or more of these. Where appropriate, the RAM can be Static Random-Access Memory (SRAM) or Dynamic Random-Access Memory (DRAM). DRAM can be Fast Page Mode Dynamic Random-Access Memory (FPMDRAM), Extended Data Out Dynamic Random-Access Memory (EDODRAM), Synchronous Dynamic Random-Access Memory (SDRAM), etc.

[0059] The memory 404 can be used to store or cache various data files that need to be processed and / or communicated, as well as possible computer program instructions executed by the processor 402.

[0060] The processor 402 reads and executes computer program instructions stored in the memory 404 to implement any of the urban road waterlogging rapid monitoring methods based on structure reparameterization in the above embodiments.

[0061] Optionally, the electronic device may further include a transmission device 406 and an input / output device 408, wherein the transmission device 406 is connected to the processor 402, and the input / output device 408 is connected to the processor 402.

[0062] The transmission device 406 can be used to receive or send data via a network. Specific examples of the network described above may include wired or wireless networks provided by the communication provider of the electronic device. In one example, the transmission device includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 406 may be a Radio Frequency (RF) module used for wireless communication with the Internet.

[0063] Input / output device 408 is used to input or output information.

[0064] Example 3 This embodiment also provides a readable storage medium storing a computer program, the computer program including program code for controlling a process to execute the process, the process including the rapid monitoring method for urban road waterlogging based on structure reparameterization according to Embodiment 1.

[0065] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated here.

[0066] Generally, various embodiments can be implemented in hardware or dedicated circuitry, software, logic, or any combination thereof. Some aspects of the invention can be implemented in hardware, while others can be implemented by firmware or software executed by a controller, microprocessor, or other computing device, but the invention is not limited thereto. Although various aspects of the invention may be shown and described as block diagrams, flowcharts, or using some other graphical representation, it should be understood that, by way of non-limiting example, these blocks, apparatuses, systems, techniques, or methods described herein can be implemented in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0067] Embodiments of the present invention can be implemented by computer software, which may be executable by a data processor of a mobile device, such as a processor entity, or by hardware, or by a combination of software and hardware. Computer software or programs (also referred to as program products) including software routines, applets, and / or macros can be stored in any device-readable data storage medium, and they include program instructions for performing specific tasks. The computer program product may include one or more computer-executable components configured to perform the embodiments when the program is run. The one or more computer-executable components may be at least one piece of software code or a portion thereof. Additionally, it should be noted in this respect that, as Figure 1 Any box in the logical flow can represent a program step, or interconnected logic circuits, boxes and functions, or a combination of program steps and logic circuits, boxes and functions. Software can be stored on physical media such as memory chips or blocks of storage implemented within a processor, magnetic media such as hard disks or floppy disks, and optical media such as DVDs and their data variants, CDs, etc. The physical medium is a non-transient medium.

[0068] Those skilled in the art should understand that the technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments have been described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0069] The above embodiments are merely illustrative of several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A rapid monitoring method for urban road waterlogging based on structure reparameterization, characterized in that, Includes the following steps: Acquire urban road perception data and construct a four-stage structured reparameterization network; The four-stage structure reparameterization network was trained using a training dataset containing annotations of waterlogged areas to obtain a trained network. The urban road perception data to be monitored is input into the trained network for inference, and the water accumulation detection result is output as a pixel-level water accumulation area mask. The four-stage structure reparameterization network consists of stacked structure reparameterization modules. Each structure reparameterization module includes four parallel convolutional branches during the training phase. These four parallel convolutional branches are as follows: 1×1 convolution branch, 3×3 convolution branch, 3×3 dilated convolution branch and 5×5 convolution branch; During the inference phase, the parameters of the four parallel convolutional branches of the structure reparameterization module are folded and fused into a single equivalent 5×5 convolutional kernel for inference.

2. The rapid monitoring method for urban road waterlogging as described in claim 1, characterized in that, The four-stage reparameterization network comprises a first stage, a second stage, a third stage, and a fourth stage connected in sequence; the first stage is used for low-level feature extraction and resolution preservation; the second stage is used for multi-scale context aggregation; the third stage is used for detail magnification and difference enhancement, using the 3×3 dilated convolution branch to distinguish between wet road surfaces and waterlogged areas; and the fourth stage is used for semantic integration and dense prediction.

3. The rapid monitoring method for urban road waterlogging as described in claim 1, characterized in that, The parameter folding and fusion process in the inference phase includes a unified mapping step for convolutional kernels: The kernels of the 1×1 convolutional branches are mapped to a 5×5 size with center alignment, and the remaining positions are filled with zero values; the kernels of the 3×3 convolutional branches are mapped to a 5×5 size with center alignment, and the remaining positions are filled with zero values; the kernels of the 3×3 dilated convolutional branches are mapped to a 5×5 size by inserting zero values ​​between the weight values ​​according to the dilation interval; the kernels of the 5×5 convolutional branches retain their original size.

4. The rapid monitoring method for urban road waterlogging as described in claim 3, characterized in that, The parameter folding and fusion process in the inference phase also includes a batch normalization layer fusion step: The convolutional layers in each convolutional branch are fused with their subsequent batch normalization layers. Based on the mean, variance, scaling factor, and offset parameters of the batch normalization layers, the kernel weights and biases of the corresponding convolutional layers are updated to obtain the fused convolutional parameters with bias terms.

5. The rapid monitoring method for urban road waterlogging as described in claim 4, characterized in that, The parameter folding fusion process in the inference phase also includes a final kernel fusion step: The convolutional kernels of the four convolutional branches, after being processed by the unified mapping step and the batch normalization layer fusion step, are summed element-wise to obtain the single equivalent 5×5 convolutional kernel.

6. The rapid monitoring method for urban road waterlogging as described in claim 2, characterized in that, The four-stage reparameterized network further includes a pre-sampling module and a pyramid pooling module; the pre-sampling module is set before the first stage and consists of two consecutive 3×3 convolutional layers, used to downsample the urban road perception data by 2 times; the pyramid pooling module is set after the fourth stage and used to perform multi-scale semantic aggregation on the output features of the fourth stage. The four-stage reparameterized network has lightweight downsampling or upsampling operations between each stage.

7. The rapid monitoring method for urban road waterlogging as described in claim 1, characterized in that, In the step of training the four-stage structure reparameterization network using a training dataset containing annotations of waterlogged areas: End-to-end training is performed using a composite loss function that combines cross-entropy loss and IoU loss.

8. The rapid monitoring method for urban road waterlogging as described in claim 1, characterized in that, Before inputting the urban road perception data to be monitored into the trained network, a preprocessing step is also included: The input urban road perception data is normalized and its resolution is adjusted. The urban road perception data includes street view camera images, dashcam images, or drone overhead views.

9. An electronic device comprising a memory and a processor, characterized in that, The memory stores a computer program, and the processor is configured to run the computer program to perform the rapid monitoring method for urban road waterlogging as described in any one of claims 1 to 8.

10. A readable storage medium, characterized in that, The readable storage medium stores a computer program, the computer program including program code for controlling a process to execute the process, the process including the rapid monitoring method for urban road waterlogging according to any one of claims 1 to 8.

Images