Vehicle overrunning monitoring method based on axle data and system thereof

Abstract

The application discloses a vehicle overloading driving monitoring method and system based on axle data. With the increase of vehicle load, the stress borne by the axle increases and the axle deforms to different degrees, and the deformation is continuous and weakly progressive. Therefore, the stress of the axle under different loads is judged based on the axle image features through artificial intelligence technology, and then whether the vehicle belongs to the overloading driving state is judged. In this way, the driving state of the vehicle can be monitored based on the visual sensor of the vehicle itself, and a pressure sensor does not need to be deployed on the road surface, so that real-time online monitoring of the driving state of the vehicle can be realized.

Description

Technical Field

[0001] This application relates to the field of vehicle monitoring, and more specifically, to a method and system for monitoring overloaded vehicle driving based on axle data. Technical Background

[0002] With economic development and the booming transportation industry, some vehicle owners, in their pursuit of economic gain, frequently engage in overloading and exceeding weight limits. Illegal overloading and exceeding weight limits not only severely damages highway and bridge infrastructure, easily leading to road traffic accidents and endangering the lives and property of the people, but also seriously disrupts the transportation market order.

[0003] Therefore, it is necessary to monitor whether vehicles are overloaded. Traditional methods for monitoring overloaded vehicles rely on comparing the vehicle pressure value collected by pressure sensors with a predetermined threshold. However, this requires installing overload detection devices at specific locations on the road, meaning that overload monitoring can only be conducted at those locations. Furthermore, overload detection devices deployed on roads are easily vandalized.

[0004] Therefore, an optimized vehicle overload monitoring solution is needed. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a method and system for monitoring vehicle overload driving based on axle data. It considers that as the vehicle load increases, the stress on the axle increases, resulting in varying degrees of deformation, and that this deformation is continuous and gradually progressive. Therefore, based on artificial intelligence technology, the stress on the axle under different loads is analyzed using axle image features to determine whether the vehicle is in an overload driving state. This allows for monitoring of the vehicle's driving status based on its own visual sensors, eliminating the need to deploy pressure sensors on the road surface, thereby enabling real-time online monitoring of the vehicle's driving status.

[0006] According to one aspect of this application, a method for monitoring vehicle overload based on axle data is provided, comprising:

[0007] Axle data acquisition steps: acquire the axle detection image and axle reference image of the vehicle to be tested, wherein the axle reference image is the axle image of a vehicle of the same model as the vehicle to be tested under overload driving conditions;

[0008] Twin detection step: The axle detection image and the axle reference image are processed through a twin network model containing a first image encoder and a second image encoder to obtain detection feature vector and reference feature vector;

[0009] Transfer evaluation steps: Calculate the transfer matrix of the detected feature vector relative to the reference feature vector; and

[0010] Monitoring result generation steps: The transition matrix is ​​passed through a classifier to obtain a classification result, which is used to indicate whether the vehicle to be detected is in an overloaded driving state.

[0011] In the above-mentioned vehicle overload monitoring method based on axle data, the twin detection step includes: a detection image encoding sub-step: using the first image encoder of the twin network model to perform multi-scale deep convolutional encoding on the axle detection image to obtain the detection feature vector; and a reference image encoding sub-step: using the second image encoder of the twin network model to perform multi-scale deep convolutional encoding on the axle reference image to obtain the reference feature vector.

[0012] In the above-mentioned vehicle overload monitoring method based on axle data, the first image encoder and the second image encoder have the same network structure.

[0013] In the above-mentioned vehicle overload monitoring method based on axle data, the first image encoder and the second image encoder include a convolutional neural network as a feature extractor and a hybrid convolutional layer cascaded with the convolutional neural network.

[0014] In the above-mentioned vehicle overload monitoring method based on axle data, the detection image encoding sub-step includes: a secondary sub-step for detection image feature extraction: using each layer of the convolutional neural network of the first image encoder as a feature extractor to perform convolution processing, pooling processing, and nonlinear activation processing on the input data respectively so that the last layer of the convolutional neural network as a feature extractor outputs a detection feature map; and a secondary sub-step for multi-scale perception of the detection feature map: using the hybrid convolutional layer of the first image encoder to perform multi-scale perception on the detection feature map to obtain the detection feature vector.

[0015] In the above-mentioned vehicle overload monitoring method based on axle data, the reference image encoding sub-step includes: a reference image feature extraction secondary sub-step: using each layer of the convolutional neural network of the second image encoder as a feature extractor to perform convolution processing, pooling processing, and nonlinear activation processing on the input data respectively so that the last layer of the convolutional neural network as a feature extractor outputs a reference feature map; and a reference feature map multi-scale perception secondary sub-step: using the hybrid convolutional layer of the second image encoder to perform multi-scale perception on the reference feature map to obtain the reference feature vector.

[0016] In the above-mentioned vehicle overload monitoring method based on axle data, the multi-scale perception secondary sub-step of the detection feature map includes: a first-scale convolutional encoding tertiary sub-step: performing two-dimensional convolutional encoding on the detection feature map using the first convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a first-scale perceived feature map, wherein the first convolutional layer uses a first convolutional kernel with a first size; a second-scale convolutional encoding tertiary sub-step: performing two-dimensional convolutional encoding on the detection feature map using the second convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a second-scale perceived feature map, wherein the second convolutional layer uses a first dilated convolutional kernel with a first dilation rate, and the size of the first dilated convolutional kernel is equal to the first size of the first convolutional kernel; and a third-scale convolutional encoding tertiary sub-step: performing two-dimensional convolutional encoding on the detection feature map using the third convolutional layer of the hybrid convolutional layer of the first image encoder. The process involves: obtaining a third-scale perceptual feature map, wherein the third convolutional layer uses a second dilated convolutional kernel with a second dilation rate, the size of which is equal to the first size of the first convolutional kernel; a third-scale convolutional encoding sub-step: performing two-dimensional convolutional encoding on the detection feature map using the fourth convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a fourth-scale perceptual feature map, wherein the fourth convolutional layer uses a third dilated convolutional kernel with a third dilation rate, the size of which is equal to the first size of the first convolutional kernel; a third-scale fusion sub-step: fusing the first-scale perceptual feature map, the second-scale perceptual feature map, the third-scale perceptual feature map, and the fourth-scale perceptual feature map to obtain a fused feature map; and a third-scale dimensionality reduction sub-step: performing global mean pooling on each feature matrix along the channel dimension of the fused feature map to obtain the detection feature vector.

[0017] In the above-mentioned vehicle overload monitoring method based on axle data, the fusion of three-level sub-steps includes: a fourth-level sub-step for determining feature map optimization factors: calculating the feature map optimization factors of the first-scale perceived feature map, the second-scale perceived feature map, the third-scale perceived feature map, and the fourth-scale perceived feature map respectively using the following formulas to obtain the first to fourth feature map optimization factors: wherein, the formulas are:

[0018]

[0019] Among them, f i,j,k The feature value at position (i,j,k) of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map, |f i,j,k | 2The square of the absolute value of each feature value is represented, and W, H and C are the width, height and number of channels of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map or the fourth scale-aware feature map, respectively. log represents the logarithm to the base 2. The optimization factor applies to the fourth sub-step: the first to fourth scale-aware feature maps are weighted by the optimization factors of the first to fourth feature maps as weighting coefficients to obtain the first to fourth weighted scale-aware feature maps, and the first to fourth weighted scale-aware feature maps are fused to obtain the fused feature map.

[0020] In the above-mentioned vehicle overload monitoring method based on axle data, the transfer evaluation step includes: calculating the transfer matrix of the detected feature vector relative to the reference feature vector using the following formula; wherein, the formula is:

[0021]

[0022] Where V a V represents the detection feature vector. b Let M represent the reference eigenvector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0023] In the above-mentioned vehicle overload monitoring method based on axle data, the monitoring result generation step includes: processing the transition matrix using the classifier with the following formula to obtain the classification result; wherein, the formula is: softmax{(W n B n ):…:(W1,B1)|Project(M)}, where Project(M) represents projecting the transition matrix into a vector, W1 to W n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias vector of each fully connected layer.

[0024] According to another aspect of this application, a vehicle overload monitoring system based on axle data is provided, comprising:

[0025] The axle data acquisition unit is used to acquire the axle detection image and axle reference image of the vehicle to be inspected. The axle reference image is the axle image of a vehicle of the same model as the vehicle to be inspected under overload driving conditions.

[0026] The twin detection unit is used to pass the axle detection image and the axle reference image through a twin network model containing a first image encoder and a second image encoder to obtain detection feature vectors and reference feature vectors;

[0027] A transfer evaluation unit is used to calculate the transfer matrix of the detected feature vector relative to the reference feature vector; and

[0028] The monitoring result generation unit is used to pass the transition matrix through a classifier to obtain a classification result, which is used to indicate whether the vehicle to be detected is in an overloaded driving state.

[0029] In the aforementioned vehicle overload monitoring system based on axle data, the twin detection unit includes: a detection image encoding subunit, used to perform multi-scale deep convolutional encoding on the axle detection image using a first image encoder of the twin network model to obtain the detection feature vector; and a reference image encoding subunit, used to perform multi-scale deep convolutional encoding on the axle reference image using a second image encoder of the twin network model to obtain the reference feature vector.

[0030] In the above-mentioned vehicle overload monitoring system based on axle data, the first image encoder and the second image encoder have the same network structure.

[0031] In the aforementioned vehicle overload monitoring system based on axle data, the first image encoder and the second image encoder include a convolutional neural network as a feature extractor and a hybrid convolutional layer cascaded with the convolutional neural network.

[0032] In the aforementioned vehicle overload monitoring system based on axle data, the detection image encoding subunit includes: a detection image feature extraction secondary subunit, used to perform convolution processing, pooling processing, and nonlinear activation processing on the input data using each layer of the convolutional neural network of the first image encoder as a feature extractor, so that the last layer of the convolutional neural network as a feature extractor outputs a detection feature map; and a detection feature map multi-scale perception secondary subunit, used to perform multi-scale perception on the detection feature map using the hybrid convolutional layer of the first image encoder to obtain the detection feature vector.

[0033] In the aforementioned vehicle overload monitoring system based on axle data, the reference image encoding subunit includes: a reference image feature extraction secondary subunit, used to perform convolution processing, pooling processing, and nonlinear activation processing on the input data using each layer of the convolutional neural network of the second image encoder as a feature extractor, so that the last layer of the convolutional neural network as a feature extractor outputs a reference feature map; and a reference feature map multi-scale perception secondary subunit, used to perform multi-scale perception on the reference feature map using the hybrid convolutional layer of the second image encoder to obtain the reference feature vector.

[0034] In the aforementioned vehicle overload monitoring system based on axle data, the multi-scale perception secondary subunit for detecting feature maps includes: a first-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the first convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a first-scale perception feature map, wherein the first convolutional layer uses a first convolutional kernel with a first size; a second-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the second convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a second-scale perception feature map, wherein the second convolutional layer uses a first dilated convolutional kernel with a first dilation rate, and the size of the first dilated convolutional kernel is equal to the first size of the first convolutional kernel; and a third-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the third convolutional layer of the hybrid convolutional layer of the first image encoder. The method comprises: obtaining a third-scale perceptual feature map, wherein the third convolutional layer uses a second dilated convolutional kernel with a second dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fourth-scale convolutional encoding three-level subunit, used to perform two-dimensional convolutional encoding on the detection feature map using the fourth convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a fourth-scale perceptual feature map, wherein the fourth convolutional layer uses a third dilated convolutional kernel with a third dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fusion three-level subunit, used to fuse the first-scale perceptual feature map, the second-scale perceptual feature map, the third-scale perceptual feature map, and the fourth-scale perceptual feature map to obtain a fused feature map; and a dimensionality reduction three-level subunit, used to perform global mean pooling on each feature matrix along the channel dimension of the fused feature map to obtain the detection feature vector.

[0035] In the aforementioned vehicle overload monitoring system based on axle data, the fusion three-level subunit includes a feature map optimization factor determination four-level subunit, used to calculate the feature map optimization factors of the first-scale sensing feature map, the second-scale sensing feature map, the third-scale sensing feature map, and the fourth-scale sensing feature map respectively using the following formulas to obtain the first to fourth feature map optimization factors:

[0036] The formula is as follows:

[0037]

[0038] Among them, f i,j,k The feature value at position (i,j,k) of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map, |f i,j,k | 2The square of the absolute value of each feature value is represented, and W, H and C are the width, height and number of channels of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map or the fourth scale-aware feature map, respectively. log represents the logarithm to the base 2. The optimization factor acts on the fourth-level sub-unit to weight the first to fourth scale-aware feature maps respectively with the optimization factor of the first to fourth feature maps as weighting coefficients to obtain the first to fourth weighted scale-aware feature maps, and then fuses the first to fourth weighted scale-aware feature maps to obtain the fused feature map.

[0039] In the aforementioned vehicle overload monitoring system based on axle data, the transfer evaluation unit is used to: calculate the transfer matrix of the detected feature vector relative to the reference feature vector using the following formula; wherein, the formula is:

[0040]

[0041] Where V a V represents the detection feature vector. b Let M represent the reference eigenvector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0042] In the aforementioned vehicle overload monitoring system based on axle data, the monitoring result generation unit is configured to: process the transition matrix using the classifier according to the following formula to obtain the classification result; wherein, the formula is:

[0043] softmax{(W n B n ):…:(W1,B1)|Project(M)}

[0044] Where Project(M) represents projecting the transition matrix into a vector, W1 to W... n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias vector of each fully connected layer.

[0045] Compared with existing technologies, the vehicle overload monitoring method and system based on axle data provided in this application takes into account that as the vehicle load increases, the stress on the axle will increase and cause different degrees of deformation, and the deformation is continuous and slightly progressive. Therefore, based on artificial intelligence technology, the axle image features are used to determine whether the vehicle is in an overload state by analyzing the stress on the axle under different loads. This allows the vehicle's own visual sensors to monitor its driving status without the need to deploy pressure sensors on the road surface, thereby enabling real-time online monitoring of the vehicle's driving status. Attached Figure Description

[0046] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0047] Figure 1 This is an application scenario diagram of the vehicle overload monitoring method based on axle data according to an embodiment of this application.

[0048] Figure 2 This is a flowchart of a vehicle overload monitoring method based on axle data according to an embodiment of this application.

[0049] Figure 3 This is a schematic diagram of the architecture of a vehicle overload monitoring method based on axle data according to an embodiment of this application.

[0050] Figure 4 This is a flowchart of the second-level sub-step of multi-scale perception of feature maps in the vehicle overload monitoring method based on axle data according to an embodiment of this application.

[0051] Figure 5 This is a block diagram of a vehicle overload monitoring system based on axle data according to an embodiment of this application. Detailed Implementation

[0052] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0053] Application Overview

[0054] As mentioned above, it is necessary to monitor whether vehicles are overloaded. Traditional methods for monitoring overloaded vehicles rely on comparing vehicle pressure values ​​collected by pressure sensors with predetermined thresholds. However, this requires installing overload detectors at specific locations on the road, meaning monitoring can only be conducted at those locations. Furthermore, overload detectors deployed on roads are easily vandalized. Therefore, an optimized overload monitoring solution is desired.

[0055] Accordingly, in the technical solution of this application, the applicant considers that as the vehicle load increases, the stress on the axle will increase and cause varying degrees of deformation. Therefore, overloaded driving can be monitored through axle images. That is, the stress on the axle under different loads can be determined through image features to determine whether the vehicle is in an overloaded driving state. This allows the vehicle's own visual sensors to monitor its driving status without the need to deploy pressure sensors on the road surface, thereby enabling real-time online monitoring of the vehicle's driving status.

[0056] More specifically, in the technical solution of this application, considering that the deformation of the axle is continuous and slightly progressive with changes in load, that is, the deformation features of the axle are not significant features in its image. Therefore, in the technical solution of this application, the axle detection image is mapped into a high-dimensional feature space and compared with the feature representation of the axle reference image in the high-dimensional feature space to analyze and judge the vehicle's driving state. It should be understood that although the axle detection image and the axle reference image may not differ much in the image source domain, after feature extraction and high-dimensional mapping, the differences in axle features between the axle detection image and the reference vehicle image will be amplified. In this way, the accuracy of axle state detection can be improved, thereby improving the accuracy of vehicle overload driving state detection.

[0057] Accordingly, in the technical solution of this application, an axle detection image and an axle reference image of the vehicle to be detected are obtained. The axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under overload driving conditions. Then, the axle detection image and the axle reference image are processed through a Siamese network model containing a first image encoder and a second image encoder to obtain a detection feature vector and a reference feature vector. Specifically, in the technical solution of this application, the Siamese network model contains two first image encoders and a second image encoder with the same network structure. The first image encoder and the second image encoder are used to extract features from the axle detection image and the axle reference image to map the axle detection image and the axle reference image into a high-dimensional feature space.

[0058] In the technical solution of this application, considering the characteristics of the axle image, the network structure of the first image encoder and the second image encoder includes: a deep convolutional neural network model as a feature extractor + a hybrid convolutional layer. The deep convolutional neural network model as the feature extractor is a standard convolutional structure. The hybrid convolutional layer includes four parallel branches, consisting of a regular convolutional layer with a kernel size of 3×3 and three dilated convolutional layers with a kernel size of 3×3, which operate on the input feature map respectively. The dilation rates of the three branches of the dilated convolution are set to 2, 3, and 4 respectively. By setting different dilation rates, image information of different receptive fields can be obtained, thus obtaining feature maps of different scales. While expanding the receptive field, downsampling loss of information is avoided. Then, the feature maps of the four branches are fused to make the sampling more dense, which has both high-level features and does not increase the number of additional parameters.

[0059] Specifically, in the technical solution of this application, the hybrid convolutional layer in the first image encoder / second image encoder fuses feature maps obtained from multiple branches. However, since the dilated convolution of each branch has a different dilation rate, the feature maps obtained from each branch have different amounts of expressive information. Thus, directly fusing feature maps by means such as dot-matrix addition will affect the information expression effect of the fused feature maps.

[0060] Therefore, for the feature map obtained from each branch, the energy aggregation factor of its wavelet function family is calculated, and expressed as:

[0061]

[0062] f i,j,k is the feature value at position (i,j,k) of the feature map, where W, H, and C are the width, height, and number of channels of the feature map, respectively, and log represents the logarithm to the base 2.

[0063] Specifically, the applicant of this application considers that for high-dimensional manifolds, since information representation tends to concentrate on high-frequency components, information tends to be distributed at the edges of the manifold. Wavelet-like function families, as separable transformations used to separate the edges in the feature distribution dimension, can convert the latent states of high-dimensional features into frequency components and express the information content in a wavelet-like energy manner. Thus, by using these functions as weighting coefficients to weight the feature maps of each branch before fusion, the degree of information aggregation in the fused feature map within its feature representation space can be improved, thereby enhancing the information representation effect of the fused feature map.

[0064] After obtaining the detection feature vector of the axle detection image and the reference feature vector of the axle reference image through the Siamese network model, a transition matrix is ​​calculated between the detection feature vector and the reference feature vector. That is, the transition matrix represents the difference between the feature representation of the axle detection image and the feature representation of the axle reference image in a high-dimensional feature space. Next, a classifier processes the transition matrix to obtain a classification result indicating whether the vehicle to be detected is in an overloaded driving state.

[0065] In summary, by using image features to analyze the stress on the axle under different loads, it is possible to determine whether the vehicle is overloaded. This allows the vehicle's own visual sensors to monitor its driving status without the need to deploy pressure sensors on the road surface, thus enabling real-time online monitoring of the vehicle's driving status.

[0066] Based on this, this application proposes a method for monitoring vehicle overload driving based on axle data, comprising: an axle data acquisition step: acquiring an axle detection image and an axle reference image of the vehicle to be detected, wherein the axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under overload driving conditions; a twin detection step: passing the axle detection image and the axle reference image through a twin network model including a first image encoder and a second image encoder to obtain a detection feature vector and a reference feature vector; a transfer evaluation step: calculating a transfer matrix of the detection feature vector relative to the reference feature vector; and a monitoring result generation step: passing the transfer matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the vehicle to be detected is under overload driving conditions.

[0067] Figure 1 This diagram illustrates an application scenario of the vehicle overload monitoring method based on axle data according to an embodiment of this application. Figure 1 As shown, in this application scenario, it is first based on the vehicle (e.g., such as...). Figure 1 The vision sensor (e.g., Se1) shown in the diagram) itself Figure 1 Se1, as shown, obtains the axle of the vehicle to be detected (e.g., as shown in the figure). Figure 1 The image shown is A). Then, the axle detection image of the vehicle to be detected is input to a server deployed with a vehicle overload monitoring algorithm based on axle data (e.g., such as...). Figure 1 In the illustrated S), the server is able to process the axle detection image of the vehicle to be detected based on the vehicle overload driving monitoring algorithm based on axle data, so as to obtain a classification result indicating whether the vehicle to be detected is in an overload driving state.

[0068] After introducing the basic principles of this application, various non-limiting embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0069] Exemplary methods

[0070] Figure 2 This is a flowchart of a vehicle overload monitoring method based on axle data according to an embodiment of this application. Figure 2 As shown, the vehicle overload monitoring method based on axle data according to an embodiment of this application includes: S110, axle data acquisition step: acquiring an axle detection image and an axle reference image of the vehicle to be detected, wherein the axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under an overload driving state; S120, twin detection step: passing the axle detection image and the axle reference image through a twin network model including a first image encoder and a second image encoder to obtain a detection feature vector and a reference feature vector; S130, transfer evaluation step: calculating a transfer matrix of the detection feature vector relative to the reference feature vector; and S140, monitoring result generation step: passing the transfer matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the vehicle to be detected is under an overload driving state.

[0071] Figure 3 This is a schematic diagram of the architecture of a vehicle overload monitoring method based on axle data according to an embodiment of this application. Figure 3 As shown, in this network architecture, firstly, axle detection images and axle reference images of the vehicle to be detected are acquired. The axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under an overloaded driving condition. Next, the axle detection images and the axle reference images are passed through a Siamese network model containing a first image encoder and a second image encoder to obtain detection feature vectors and reference feature vectors. Then, the transition matrix of the detection feature vectors relative to the reference feature vectors is calculated. Finally, the transition matrix is ​​passed through a classifier to obtain a classification result, which indicates whether the vehicle to be detected is under an overloaded driving condition.

[0072] In the axle data acquisition step S110, an axle detection image and an axle reference image of the vehicle to be detected are acquired. The axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under overload driving conditions. As mentioned above, it is necessary to monitor whether a vehicle is overloaded. Traditional methods for monitoring overloaded vehicles rely on comparing the vehicle pressure value collected by a pressure sensor with a predetermined threshold. However, this requires setting up overload detection devices at specific locations on the road, meaning that overload monitoring can only be performed at specific locations on the road. Furthermore, overload detection devices deployed on the road are easily damaged. Therefore, an optimized overload monitoring scheme is desired.

[0073] Accordingly, in the technical solution of this application, the applicant considers that as the vehicle load increases, the stress on the axle will increase and cause varying degrees of deformation. Therefore, overloaded driving can be monitored through axle images. That is, the stress on the axle under different loads can be determined through image features to determine whether the vehicle is in an overloaded driving state. This allows the vehicle's own visual sensors to monitor its driving status without the need to deploy pressure sensors on the road surface, thereby enabling real-time online monitoring of the vehicle's driving status.

[0074] More specifically, in the technical solution of this application, considering that the deformation of the axle is continuous and slightly progressive with changes in load, that is, the deformation features of the axle are not significant features in its image. Therefore, in the technical solution of this application, the axle detection image is mapped into a high-dimensional feature space and compared with the feature representation of the axle reference image in the high-dimensional feature space to analyze and judge the vehicle's driving state. It should be understood that although the axle detection image and the axle reference image may not differ much in the image source domain, after feature extraction and high-dimensional mapping, the differences in axle features between the axle detection image and the reference vehicle image will be amplified. In this way, the accuracy of axle state detection can be improved, thereby improving the accuracy of vehicle overload driving state detection.

[0075] Accordingly, in the technical solution of this application, an axle detection image and an axle reference image of the vehicle to be tested are obtained. The axle reference image is an axle image of a vehicle of the same model as the vehicle to be tested under overload driving conditions. The axle detection image of the vehicle to be tested can be obtained based on the vehicle's own vision sensor, without the need to deploy pressure sensors on the road surface.

[0076] In the twin detection step S120, the axle detection image and the axle reference image are processed through a twin network model comprising a first image encoder and a second image encoder to obtain detection feature vectors and reference feature vectors. Specifically, in the technical solution of this application, the twin network model comprises two image encoders with the same network structure: a first image encoder and a second image encoder. The first image encoder and the second image encoder are used to extract features from the axle detection image and the axle reference image to map them into a high-dimensional feature space.

[0077] Specifically, in this embodiment, the twin detection step S120 includes: a detection image encoding sub-step: performing multi-scale deep convolutional encoding on the axle detection image using a first image encoder of the twin network model to obtain the detection feature vector; and a reference image encoding sub-step: performing multi-scale deep convolutional encoding on the axle reference image using a second image encoder of the twin network model to obtain the reference feature vector. The first image encoder and the second image encoder have the same network structure.

[0078] In the technical solution of this application, considering the characteristics of the axle image, the network structure of the first image encoder and the second image encoder includes: a deep convolutional neural network model as a feature extractor + a hybrid convolutional layer.

[0079] Specifically, in this embodiment, the detection image encoding sub-step includes: a secondary sub-step for detecting image features: using each layer of the convolutional neural network of the first image encoder as a feature extractor to perform convolution processing, pooling processing, and nonlinear activation processing on the input data respectively so that the last layer of the convolutional neural network as a feature extractor outputs a detection feature map; and a secondary sub-step for detecting feature map multi-scale perception: using the hybrid convolutional layer of the first image encoder to perform multi-scale perception on the detection feature map to obtain the detection feature vector.

[0080] The deep convolutional neural network model used as the feature extractor is a standard convolutional structure. The hybrid convolutional layer includes four parallel branches, consisting of a regular convolutional layer with a kernel size of 3×3 and three dilated convolutional layers with a kernel size of 3×3. These branches operate on the input feature map. The dilation rates of the three dilated convolutional branches are set to 2, 3, and 4, respectively. By setting different dilation rates, image information of different receptive fields can be obtained, thus obtaining feature maps of different scales. This expands the receptive field while avoiding information loss due to downsampling.

[0081] Specifically, in this embodiment, the reference image encoding sub-step includes: a secondary sub-step for reference image feature extraction: using each layer of the convolutional neural network of the second image encoder as a feature extractor to perform convolution processing, pooling processing, and nonlinear activation processing on the input data respectively so that the last layer of the convolutional neural network as a feature extractor outputs a reference feature map; and a secondary sub-step for multi-scale perception of the reference feature map: using the hybrid convolutional layer of the second image encoder to perform multi-scale perception on the reference feature map to obtain the reference feature vector.

[0082] Figure 4This is a flowchart of the second-level sub-step of multi-scale perception of feature maps in the vehicle overload monitoring method based on axle data according to an embodiment of this application. Figure 4 The multi-scale perception secondary sub-step of the detection feature map includes: a first-scale convolutional encoding tertiary sub-step S210, in which the detection feature map is subjected to two-dimensional convolutional encoding using the first convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a first-scale perceptual feature map, wherein the first convolutional layer uses a first convolutional kernel with a first size; a second-scale convolutional encoding tertiary sub-step S220, in which the detection feature map is subjected to two-dimensional convolutional encoding using the second convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a second-scale perceptual feature map, wherein the second convolutional layer uses a first dilated convolutional kernel with a first dilation rate, and the size of the first dilated convolutional kernel is equal to the first size of the first convolutional kernel; and a third-scale convolutional encoding tertiary sub-step S230, in which the detection feature map is subjected to two-dimensional convolutional encoding using the third convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a third-scale perceptual feature map. The feature map is further subdivided into three levels: a third convolutional layer using a second dilated convolutional kernel with a second dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fourth-scale convolutional encoding sub-step S240, where the detection feature map is encoden using a fourth convolutional layer of the hybrid convolutional layer of the first image encoder in two dimensions to obtain a fourth-scale perceptual feature map, wherein the fourth convolutional layer uses a third dilated convolutional kernel with a third dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fusion sub-step S250, where the first-scale perceptual feature map, the second-scale perceptual feature map, the third-scale perceptual feature map, and the fourth-scale perceptual feature map are fused to obtain a fused feature map; and a dimensionality reduction sub-step S260, where global mean pooling is performed on each feature matrix along the channel dimension of the fused feature map to obtain the detection feature vector.

[0083] Specifically, in this embodiment, the fusion three-level sub-step S250 includes: a four-level sub-step for determining feature map optimization factors: calculating the feature map optimization factors of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, and the fourth scale-aware feature map respectively using the following formula to obtain the first to fourth feature map optimization factors: wherein, the formula is:

[0084]

[0085] Among them, f i,j,k The feature value at position (i,j,k) of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map, |fi,j,k | 2 The expression represents the square of the absolute value of each feature, where W, H, and C are the width, height, and number of channels of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map, respectively, and log represents the logarithm to base 2. The optimization factor operates in four sub-steps: the first to fourth scale-aware feature maps are weighted using the optimization factors of the first to fourth feature maps as weighting coefficients to obtain the first to fourth weighted scale-aware feature maps, and then these weighted feature maps are merged to obtain the fused feature map. In other words, the four branch feature maps are merged, resulting in denser sampling, which provides high-level features without adding additional parameters.

[0086] Specifically, in the technical solution of this application, the hybrid convolutional layer in the first image encoder / second image encoder fuses feature maps obtained from multiple branches. However, since the dilated convolution of each branch has a different dilation rate, the feature maps obtained from each branch have different amounts of expressive information. Therefore, directly fusing feature maps using methods such as dot-matrix addition will affect the information representation effect of the fused feature maps. Therefore, for each feature map obtained from each branch, its wavelet function family-like energy aggregation factor is calculated, expressed as:

[0087]

[0088] f i,j,k is the feature value at position (i,j,k) of the feature map, where W, H, and C are the width, height, and number of channels of the feature map, respectively, and log represents the logarithm to the base 2.

[0089] Specifically, the applicant of this application considers that for high-dimensional manifolds, since information representation tends to concentrate on high-frequency components, information tends to be distributed at the edges of the manifold. Wavelet-like function families, as separable transformations used to separate the edges in the feature distribution dimension, can convert the latent states of high-dimensional features into frequency components and express the information content in a wavelet-like energy manner. Thus, by using these functions as weighting coefficients to weight the feature maps of each branch before fusion, the degree of information aggregation in the fused feature map within its feature representation space can be improved, thereby enhancing the information representation effect of the fused feature map.

[0090] In the transfer evaluation step S130, a transfer matrix is ​​calculated between the detected feature vector and the reference feature vector. That is, the transfer matrix represents the difference between the feature representation of the axle detection image and the feature representation of the axle reference image in the high-dimensional feature space.

[0091] Specifically, in this embodiment, the transfer evaluation step S130 includes: calculating the transfer matrix of the detected feature vector relative to the reference feature vector using the following formula; wherein, the formula is:

[0092]

[0093] Where V a V represents the detection feature vector. b Let M represent the reference eigenvector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0094] In the monitoring result generation step S140, the transition matrix is ​​processed by a classifier to obtain a classification result, which indicates whether the vehicle to be detected is in an overloaded driving state. Thus, by using image features to assess the stress on the axles under different loads, it is possible to determine whether the vehicle is in an overloaded driving state.

[0095] Specifically, in this embodiment of the application, the monitoring result generation step S140 includes: processing the transition matrix using the classifier with the following formula to obtain the classification result; wherein, the formula is:

[0096] softmax{(W n B n ):…:(W1,B1)|Project(M)}

[0097] Where Project(M) represents projecting the transition matrix into a vector, W1 to W... n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias vector of each fully connected layer.

[0098] In summary, the vehicle overload monitoring method based on axle data according to the embodiments of this application is explained. It takes into account that as the vehicle load increases, the stress on the axle will increase and cause different degrees of deformation, and the deformation is continuous and slightly progressive. Therefore, based on artificial intelligence technology, the axle image features are used to determine the stress on the axle under different loads and then to determine whether the vehicle is in an overload driving state. In this way, the vehicle's own visual sensors can be used to monitor its own driving state without deploying pressure sensors on the road surface, thereby realizing real-time online monitoring of the vehicle's driving state.

[0099] Exemplary System

[0100] Figure 5 This is a block diagram of a vehicle overload monitoring system based on axle data according to an embodiment of this application. Figure 5As shown, a vehicle overload monitoring system 100 based on axle data according to an embodiment of this application includes: an axle data acquisition unit 110, used to acquire an axle detection image and an axle reference image of a vehicle to be detected, wherein the axle reference image is an axle image of a vehicle of the same model as the vehicle to be detected under an overload driving state; a twin detection unit 120, used to pass the axle detection image and the axle reference image through a twin network model including a first image encoder and a second image encoder to obtain a detection feature vector and a reference feature vector; a transfer evaluation unit 130, used to calculate a transfer matrix of the detection feature vector relative to the reference feature vector; and a monitoring result generation unit 140, used to pass the transfer matrix through a classifier to obtain a classification result, wherein the classification result is used to indicate whether the vehicle to be detected is under an overload driving state.

[0101] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the twin detection unit includes: a detection image encoding subunit, used to perform multi-scale deep convolutional encoding on the axle detection image using a first image encoder of the twin network model to obtain the detection feature vector; and a reference image encoding subunit, used to perform multi-scale deep convolutional encoding on the axle reference image using a second image encoder of the twin network model to obtain the reference feature vector.

[0102] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the first image encoder and the second image encoder have the same network structure.

[0103] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the first image encoder and the second image encoder include a convolutional neural network as a feature extractor and a hybrid convolutional layer cascaded with the convolutional neural network.

[0104] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the detection image encoding subunit includes: a detection image feature extraction secondary subunit, used to perform convolution processing, pooling processing, and nonlinear activation processing on the input data using each layer of the convolutional neural network of the first image encoder as a feature extractor, so that the last layer of the convolutional neural network as a feature extractor outputs a detection feature map; and a detection feature map multi-scale perception secondary subunit, used to perform multi-scale perception on the detection feature map using the hybrid convolutional layer of the first image encoder to obtain the detection feature vector.

[0105] In one example, in the aforementioned vehicle overload monitoring system based on axle data, the reference image encoding subunit includes: a reference image feature extraction secondary subunit, used to perform convolution processing, pooling processing, and nonlinear activation processing on the input data using each layer of the convolutional neural network of the second image encoder as a feature extractor, so that the last layer of the convolutional neural network as a feature extractor outputs a reference feature map; and a reference feature map multi-scale perception secondary subunit, used to perform multi-scale perception on the reference feature map using the hybrid convolutional layer of the second image encoder to obtain the reference feature vector.

[0106] In one example, in the aforementioned vehicle overload monitoring system based on axle data, the detection feature map multi-scale perception secondary subunit includes: a first-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the first convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a first-scale perception feature map, wherein the first convolutional layer uses a first convolutional kernel with a first size; a second-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the second convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a second-scale perception feature map, wherein the second convolutional layer uses a first dilated convolutional kernel with a first dilation rate, the size of the first dilated convolutional kernel being equal to the first size of the first convolutional kernel; and a third-scale convolutional coding tertiary subunit, used to perform two-dimensional convolutional coding on the detection feature map using the third convolutional layer of the hybrid convolutional layer of the first image encoder. A convolutional encoding is performed to obtain a third-scale perceptual feature map, wherein the third convolutional layer uses a second dilated convolutional kernel with a second dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fourth-scale convolutional encoding three-level subunit is used to perform two-dimensional convolutional encoding on the detection feature map using a fourth convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a fourth-scale perceptual feature map, wherein the fourth convolutional layer uses a third dilated convolutional kernel with a third dilation rate, the size of which is equal to the first size of the first convolutional kernel; a fusion three-level subunit is used to fuse the first-scale perceptual feature map, the second-scale perceptual feature map, the third-scale perceptual feature map, and the fourth-scale perceptual feature map to obtain a fused feature map; and a dimensionality reduction three-level subunit is used to perform global mean pooling on each feature matrix along the channel dimension of the fused feature map to obtain the detection feature vector.

[0107] In one example, in the aforementioned vehicle overload monitoring system based on axle data, the fusion three-level subunit includes: a feature map optimization factor determination four-level subunit, used to calculate the feature map optimization factors of the first-scale perceived feature map, the second-scale perceived feature map, the third-scale perceived feature map, and the fourth-scale perceived feature map respectively using the following formulas to obtain the first to fourth feature map optimization factors:

[0108] The formula is as follows:

[0109]

[0110] Among them, f i,j,k The feature value at position (i,j,k) of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map, |f i,j,k | 2 The square of the absolute value of each feature value is represented, and W, H and C are the width, height and number of channels of the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map or the fourth scale-aware feature map, respectively. log represents the logarithm to the base 2. The optimization factor acts on the fourth-level sub-unit to weight the first to fourth scale-aware feature maps respectively with the optimization factor of the first to fourth feature maps as weighting coefficients to obtain the first to fourth weighted scale-aware feature maps, and then fuses the first to fourth weighted scale-aware feature maps to obtain the fused feature map.

[0111] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the transfer evaluation unit is used to: calculate the transfer matrix of the detected feature vector relative to the reference feature vector using the following formula; wherein, the formula is:

[0112]

[0113] Where V a V represents the detection feature vector. b Let M represent the reference eigenvector, and let M represent the transition matrix. This represents the multiplication of a matrix and a vector.

[0114] In one example, in the above-mentioned vehicle overload monitoring system based on axle data, the monitoring result generation unit is configured to: process the transition matrix using the classifier according to the following formula to obtain the classification result; wherein, the formula is:

[0115] softmax{(W n B n):…:(W1,B1)|Project(M)}

[0116] Where Project(M) represents projecting the transition matrix into a vector, W1 to W... n Here are the weight matrices for each fully connected layer, B1 to B... n This represents the bias vector of each fully connected layer.

[0117] Here, those skilled in the art will understand that the specific functions and operations of each unit and module in the above-described vehicle overload monitoring system 100 based on axle data have been referenced above. Figures 1 to 4 The description of the vehicle overload monitoring method based on axle data has been detailed, and therefore, its repeated description will be omitted.

[0118] As described above, the vehicle overload monitoring system 100 based on axle data according to embodiments of this application can be implemented in various terminal devices, such as servers for monitoring vehicle overload based on axle data. In one example, the vehicle overload monitoring system 100 based on axle data according to embodiments of this application can be integrated into a terminal device as a software module and / or a hardware module. For example, the vehicle overload monitoring system 100 based on axle data can be a software module in the operating system of the terminal device, or it can be an application developed for the terminal device; of course, the vehicle overload monitoring system 100 based on axle data can also be one of many hardware modules of the terminal device.

[0119] Alternatively, in another example, the vehicle overload monitoring system 100 based on axle data and the terminal device can also be separate devices, and the vehicle overload monitoring system 100 based on axle data can be connected to the terminal device via wired and / or wireless networks, and transmit interactive information in accordance with an agreed data format.

[0120] The basic principles of this application have been described above with reference to specific embodiments. However, it should be noted that the advantages, benefits, and effects mentioned in this application are merely examples and not limitations, and should not be considered as essential features of each embodiment of this application. Furthermore, the specific details disclosed above are for illustrative and facilitative purposes only, and are not limitations. These details do not limit the application to the necessity of employing the aforementioned specific details for implementation.

[0121] The block diagrams of devices, apparatuses, devices, and systems involved in this application are merely illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, apparatuses, devices, and systems can be connected, arranged, and configured in any manner. Words such as “comprising,” “including,” “having,” etc., are open-ended terms meaning “including but not limited to,” and are used interchangeably with them. The terms “or” and “and” as used herein refer to the terms “and / or,” and are used interchangeably with them unless the context clearly indicates otherwise. The term “such as” as used herein refers to the phrase “such as but not limited to,” and is used interchangeably with it.

[0122] It should also be noted that in the apparatus, equipment, and methods of this application, the components or steps can be disassembled and / or recombined. These disassemblies and / or recombinations should be considered as equivalent solutions of this application.

[0123] The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use this application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein can be applied to other aspects without departing from the scope of this application. Therefore, this application is not intended to be limited to the aspects shown herein, but rather to be accorded the widest scope consistent with the principles and novel features disclosed herein.

[0124] The above description has been given for purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of this application to the forms disclosed herein. Although numerous exemplary aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for monitoring vehicle overload based on axle data, characterized in that, include: Axle data acquisition steps: acquire the axle detection image and axle reference image of the vehicle to be tested, wherein the axle reference image is the axle image of a vehicle of the same model as the vehicle to be tested under overload driving conditions; Twin detection steps: The axle detection image and the axle reference image are processed through a twin network model containing a first image encoder and a second image encoder to obtain detection feature vectors and reference feature vectors. The first image encoder and the second image encoder include a convolutional neural network as a feature extractor and a hybrid convolutional layer cascaded with the convolutional neural network. The hybrid convolutional layer includes four parallel branches, consisting of a regular convolutional layer with a kernel size of 3×3 and three dilated convolutional layers with a kernel size of 3×3, which operate on the input feature map respectively. The dilation rates of the three branches of the dilated convolution are set to 2, 3, and 4 respectively. The transfer evaluation step involves calculating the transfer matrix of the detected feature vector relative to the reference feature vector, specifically including: calculating the transfer matrix of the detected feature vector relative to the reference feature vector using the following formula; wherein the formula is: , in This represents the detection feature vector. Represents the reference feature vector, Denotes the transition matrix, Represents matrix-vector multiplication ;as well as Monitoring result generation steps: The transition matrix is ​​passed through a classifier to obtain a classification result, which is used to indicate whether the vehicle to be detected is in an overloaded driving state.

2. The vehicle overload monitoring method based on axle data according to claim 1, characterized in that, The twin detection steps include: The detection image encoding sub-step involves performing multi-scale deep convolutional encoding on the axle detection image using the first image encoder of the Siamese network model to obtain the detection feature vector; and Reference image encoding sub-step: The second image encoder of the Siamese network model is used to perform multi-scale deep convolutional encoding on the axle reference image to obtain the reference feature vector.

3. The vehicle overload monitoring method based on axle data according to claim 2, characterized in that, The first image encoder and the second image encoder have the same network structure.

4. The vehicle overload monitoring method based on axle data according to claim 1, characterized in that, The image encoding sub-step for detection includes: The second-level sub-step for detecting image feature extraction involves: using each layer of the convolutional neural network (CNN) of the first image encoder, which acts as a feature extractor, to perform convolution, pooling, and non-linear activation processing on the input data, so that the last layer of the CNN outputs a detection feature map; and The second-level sub-step of multi-scale perception of the detection feature map is as follows: the detection feature map is subjected to multi-scale perception using the hybrid convolutional layer of the first image encoder to obtain the detection feature vector.

5. The vehicle overload monitoring method based on axle data according to claim 4, characterized in that, The reference image encoding sub-step includes: The reference image feature extraction secondary sub-step involves using each layer of the convolutional neural network (CNN) acting as a feature extractor in the second image encoder to perform convolution, pooling, and non-linear activation processing on the input data, so that the last layer of the CNN outputs a reference feature map; and The second-level sub-step of multi-scale perception of the reference feature map is as follows: the reference feature map is subjected to multi-scale perception using the hybrid convolutional layer of the second image encoder to obtain the reference feature vector.

6. The vehicle overload monitoring method based on axle data according to claim 4, characterized in that, The multi-scale sensing secondary sub-step for detecting feature maps includes: First-scale convolutional coding three-level sub-step: The detection feature map is subjected to two-dimensional convolutional coding using the first convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a first-scale perceptual feature map, wherein the first convolutional layer uses a first convolutional kernel with a first size; The second-scale convolutional coding three-level sub-step: The detection feature map is subjected to two-dimensional convolutional coding using the second convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a second-scale perceptual feature map, wherein the second convolutional layer uses a first dilated convolutional kernel with a first dilation rate, and the size of the first dilated convolutional kernel is equal to the first size of the first convolutional kernel; The third-scale convolutional coding three-level sub-step: The detection feature map is subjected to two-dimensional convolutional coding using the third convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a third-scale perceptual feature map, wherein the third convolutional layer uses a second dilated convolutional kernel with a second dilation rate, and the size of the second dilated convolutional kernel is equal to the first size of the first convolutional kernel; The fourth-scale convolutional coding three-level sub-step: the detection feature map is subjected to two-dimensional convolutional coding using the fourth convolutional layer of the hybrid convolutional layer of the first image encoder to obtain a fourth-scale perceptual feature map, wherein the fourth convolutional layer uses a third dilated convolutional kernel with a third dilation rate, and the size of the third dilated convolutional kernel is equal to the first size of the first convolutional kernel; The three-level fusion sub-step involves fusing the first-scale perceived feature map, the second-scale perceived feature map, the third-scale perceived feature map, and the fourth-scale perceived feature map to obtain a fused feature map; and The third sub-step of dimensionality reduction is to perform global mean pooling on each feature matrix along the channel dimension of the fused feature map to obtain the detection feature vector.

7. The vehicle overload monitoring method based on axle data according to claim 6, characterized in that, The fusion three-level sub-steps include: The feature map optimization factor determination involves four sub-steps: calculating the feature map optimization factors for the first-scale perceived feature map, the second-scale perceived feature map, the third-scale perceived feature map, and the fourth-scale perceived feature map using the following formulas to obtain the first to fourth feature map optimization factors: The formula is as follows: , in, This represents the first scale-aware feature map, the second scale-aware feature map, the third scale-aware feature map, or the fourth scale-aware feature map. Location feature value Let represent the square of the absolute value of each eigenvalue, and , and These are the width, height, and number of channels of the first-scale perceived feature map, the second-scale perceived feature map, the third-scale perceived feature map, or the fourth-scale perceived feature map, respectively. Represents the logarithm to base 2; and The fourth sub-step of the optimization factor is as follows: the first to fourth scale-aware feature maps are weighted by the first to fourth feature map optimization factors as weighting coefficients to obtain the first to fourth weighted scale-aware feature maps, and the first to fourth weighted scale-aware feature maps are fused to obtain the fused feature map.

8. The vehicle overload monitoring method based on axle data according to claim 1, characterized in that, The monitoring result generation step includes: processing the transition matrix using the classifier with the following formula to obtain the classification result; The formula is as follows: ,in This indicates that the transition matrix is ​​projected as a vector. to Here are the weight matrices for each fully connected layer. to This represents the bias vector of each fully connected layer.

Images