Track recognition method based on EfficientNet

Abstract

This invention relates to an orbit recognition method based on EfficientNet, comprising the following steps: constructing a network structure based on the EfficientNetB0 structure; obtaining an orbit recognition model based on an orbit dataset through a network training method; acquiring real-time orbit images as input to the orbit recognition model to obtain the coordinates of the recognized orbit.

Description

Technical Field

[0001] This invention relates to the field of rail transit safety. Background Technology

[0002] In modern rail transit systems, accurate identification and location of track structures are crucial for ensuring operational safety and improving efficiency. Currently, track identification is mainly achieved through traditional machine vision or general segmentation models. However, traditional machine vision involves complex parameter adjustments and often faces problems such as insufficient accuracy and low efficiency when dealing with complex environments and real-time monitoring; general segmentation models lack the precision for track selection and have a high false positive rate. EfficientNet, as a novel convolutional neural network architecture, has attracted widespread attention due to its efficient feature extraction capabilities and small model size. Summary of the Invention

[0003] Based on this, and in view of the shortcomings of the existing technology, this invention proposes a track recognition method based on EfficientNet.

[0004] To achieve the above objectives, the solution of the present invention is as follows: Step (1) Construct a network structure based on the EfficientNetB0 structure; Step (2) Obtain a track recognition model based on the track dataset through network training methods; Step (3) Collect real-time track images as input to the track recognition model to obtain the track coordinates.

[0005] Furthermore, step (1), which involves constructing a network structure based on the EfficientNetB0 architecture, includes:

[0006] The network structure described in step (11) is based on the EfficientNetB0 structure, including 49 convolutional layers, 17 pooling layers, 1 Flatten layer, 33 fully connected layers, and 17 dropout layers;

[0007] Step (12) involves the convolutional layer extracting image features; the pooling layer employs global average pooling or average pooling; the Flatten layer converts the input data into a one-dimensional array; the fully connected layer is used for feature recalibration or mapping of feature representations to data sample labels; the network structure consists of 9 modules, as follows:

[0008] Module 1 is a single convolutional layer with a kernel size of 3×3 and a stride of 2;

[0009] Module 2 is a single MBConv structure, and the kernel size used in the depthwise separable convolutional layer is 3×3.

[0010] Module 3 consists of two repeated MBConv structures. The depthwise separable convolutional layers use a 3×3 kernel with a stride of 2.

[0011] Module 4 consists of two repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 2.

[0012] Module 5 consists of three repeated MBConv structures. The depthwise separable convolutional layers use a 3×3 kernel with a stride of 2.

[0013] Module 6 consists of three repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 1.

[0014] Module 7 consists of four repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 2.

[0015] Module 8 is a single MBConv structure, and the depthwise separable convolutional layer uses a 3×3 kernel with a stride of 1.

[0016] Module 9 consists of one convolutional layer with a kernel size of 1×1; one average pooling layer; and one fully connected layer 1, with an output vector length of 65.

[0017] The MBConv structure includes: a convolutional layer with a kernel size of 1×1 and a stride of 1; a depthwise separable convolutional layer with a stride of 1 or 2; an SE attention mechanism module; a convolutional layer with a kernel size of 1×1 and a stride of 1; and a dropout layer.

[0018] The SE attention mechanism module includes: one global average pooling layer; one fully connected layer 2, the output vector length of which is one-quarter of the input vector length; and one fully connected layer 3, the output vector length of which is the same as the output vector length of the depthwise separable convolutional layer.

[0019] In step (12) of step (13), the fully connected layer 3 of the SE attention mechanism module uses the Sigmoid activation function, as shown in the following formula:

[0020]

[0021] Where x is the input vector; except for the fully connected layer 3 of the SE attention mechanism module, the Swish activation function is used for nonlinear transformation, as shown in the following formula:

[0022]

[0023] Furthermore, the network training method in step (2) includes:

[0024] Step (21) takes the orbit dataset as the input of the network structure, passes it through a convolutional layer or a pooling layer, and then calculates the feature map through an activation function until the final prediction result is obtained;

[0025] In the prediction result described in step (22), x l,i The x-coordinate value is the left-side horizontal axis value. r,i is the x-coordinate value of the right rail; T is the number of points after the single rail is truncated in the y-axis direction, and the value of i ranges from [1,T].

[0026] The loss function L for step (23) is as follows:

[0027] L = L l +α1L s +α2L m

[0028] Among them, L l For linear loss; L s For volatility loss; L m α1 and α2 are the scaling factors for volatility loss and cutoff loss, respectively.

[0029] Linear loss L l The formula is as follows:

[0030]

[0031] Among them, I l,i ,I r,i U represents the intersection of the horizontal expansion of the i-th predicted point and the actual point on the left and right rails, respectively; l,i U r,i Let be the union of the horizontal expansions of the i-th predicted point and the actual point on the left and right rails, respectively. The calculation formula is as follows:

[0032]

[0033] Where the superscript ^ represents the actual value; w is the horizontal expansion width;

[0034] Volatility Loss L s The formula is as follows:

[0035]

[0036] Where the superscript ^ represents the true value; r is the proportion of the number of sign changes of the first-order difference of the x-coordinates of all points on the track to the total number of points; β s The smoothing point for the β-proportional loss is:

[0037]

[0038] Cut-off loss L m The formula is as follows:

[0039]

[0040] Where the superscript ^ represents the true value; β m The proportional smoothing point for truncating the loss;

[0041] Step (24): Using the Newton accelerated gradient descent method with adaptive momentum, continuously update the network parameters according to the latest calculated loss function value until the loss function value converges or the preset number of iterations is reached, and obtain the track recognition model.

[0042] Further, step (3) of obtaining the identified orbit coordinates includes:

[0043] Step (31) The real-time image of the track is input into the track recognition model. After passing through the convolutional layer or pooling layer, the feature map is calculated by the activation function until the fully connected layer obtains the final detection result.

[0044] The detection result in step (32) is a vector of length 65, where the 1st to 32nd values ​​are the predicted values ​​of the left horizontal axis, i.e., x. l,i The 33rd to 64th values ​​are the predicted x-coordinate values ​​for the right track. r,i The 65th value is the number of ordinates of the single track after truncation, i.e., T; the range of i is [1, T].

[0045] In step (33), a total of 64 orbital points were obtained in step (32). The left orbital coordinate sequence is (x l,i ,y i The right-hand orbital coordinate sequence is (x r,i ,y i ), y-axis Where H is the image height, and i ranges from [1, 32]; combined with the number of vertical coordinate points T of the truncated single track, the range of i after truncation is [1, T].

[0046] This invention offers the following advantages: It proposes a track recognition method based on EfficientNet, constructing a network structure based on EfficientNetB0. Using a track dataset as input, a high-precision track recognition model is trained. Real-time track images are acquired, and the coordinates of the recognized track are obtained through the high-precision model. By extracting depth features from the track images and introducing a unique and efficient loss function, the model performance is optimized. This significantly reduces computational resource consumption and improves real-time processing capabilities while enhancing model accuracy. Achieving accurate and real-time recognition further meets the high-precision and high-efficiency requirements of rail transit systems, providing strong support for the safe operation of rail transit. Attached Figure Description

[0047] Figure 1 This is a flowchart of the detection process of the present invention.

[0048] Figure 2 This is a data flow diagram of the network of this invention.

[0049] Figure 3 This is a schematic diagram of the prediction results of the embodiment. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] To facilitate understanding by those skilled in the art, an embodiment is provided to illustrate an orbit recognition method based on EfficientNet, comprising the following steps:

[0052] Step (1) Build a network structure based on EfficientNetB0; Step (2) Obtain a track recognition model based on the track dataset through network training methods; Step (3) Collect real-time track images as input to the track recognition model to obtain the track coordinates.

[0053] Specifically, the network structure built in step (1) is based on the EfficientNetB0 structure as the backbone network.

[0054] Specifically, in the network structure built in step (1), the MBConv structure multiplier factor of module 2 in step (12) is 1, the MBConv structure multiplier factor of the other modules is 6, and the dropout rate is 0.2.

[0055] Specifically, in the network training method of step (2), the orbital dataset used in step (21) consists of images and label data. The image shooting conditions include different weather (sunny, rainy, foggy), different times (daytime, nighttime), and different scenes (elevated, tunnel, ground). The label data consists of left and right track coordinates. In this embodiment, there are 62,500 images and labels, with a training set ratio of 70% and a test set ratio of 30%.

[0056] Specifically, in the network training method of step (2), the loss function scaling coefficient α1 in step (23) is 0.2, α2 is 0.5, the horizontal expansion width w is 15, and the scaling smoothing point β of the fluctuation loss is... s The proportional smoothing point β with a value of 0.05 and a cutoff loss. m The value is 0.015.

[0057] Specifically, in the network training method of step (2), step (24) uses the LBFGS optimizer to train the network with a learning rate of 0.0005 and a predetermined number of iterations of 1000 steps.

[0058] Specifically, the results of the embodiments are as follows: Figure 3 As shown: The image size is 512*512, and the model output results have a total of 64 coordinate points, with 32 points each for the left and right tracks; after truncation, the number of vertical coordinates for a single track is T=22; the final track coordinate results are the coordinate points below the dotted line i=T in the figure, with 22 points each for the left and right tracks.

Claims

1. A trajectory recognition method based on EfficientNet, characterized in that, include: Step 1: Construct a network structure based on the EfficientNetB0 architecture; Step 2: Based on the track dataset, obtain a track recognition model through network training methods; Step 3: Acquire real-time images of the track as input to the track recognition model to obtain the coordinates of the track.

2. The method as described in claim 1, characterized in that, The network structure based on the EfficientNetB0 architecture described in step 1 includes: Step 11: The network structure is based on the EfficientNetB0 structure, including 49 convolutional layers, 17 pooling layers, 1 Flatten layer, 33 fully connected layers, and 17 dropout layers. Step 12: The convolutional layer extracts image features; the pooling layer uses global average pooling or average pooling; the Flatten layer converts the input data into a one-dimensional array; the fully connected layer is used for feature recalibration or mapping of feature representations to data sample labels; the network structure consists of 9 modules, as follows: Module 1 is a single convolutional layer with a kernel size of 3×3 and a stride of 2; Module 2 is a single MBConv structure, and the kernel size used in the depthwise separable convolutional layer is 3×3. Module 3 consists of two repeated MBConv structures. The depthwise separable convolutional layers use a 3×3 kernel with a stride of 2. Module 4 consists of two repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 2. Module 5 consists of three repeated MBConv structures. The depthwise separable convolutional layers use a 3×3 kernel with a stride of 2. Module 6 consists of three repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 1. Module 7 consists of four repeated MBConv structures. The depthwise separable convolutional layers use a 5×5 kernel with a stride of 2. Module 8 is a single MBConv structure, and the depthwise separable convolutional layer uses a 3×3 kernel with a stride of 1. Module 9 consists of one convolutional layer with a kernel size of 1×1; one average pooling layer; and one fully connected layer 1, with an output vector length of 65. The MBConv structure includes: a convolutional layer with a kernel size of 1×1 and a stride of 1; a depthwise separable convolutional layer with a stride of 1 or 2; an SE attention mechanism module; a convolutional layer with a kernel size of 1×1 and a stride of 1; and a dropout layer. The SE attention mechanism module includes: one global average pooling layer; one fully connected layer 2, the output vector length of which is one-quarter of the input vector length; and one fully connected layer 3, the output vector length of which is the same as the output vector length of the depthwise separable convolutional layer. Step 13: In step 12, the fully connected layer 3 of the SE attention mechanism module uses the Sigmoid activation function, as shown in the following formula: Where x is the input vector; except for the fully connected layer 3 of the SE attention mechanism module, the Swish activation function is used for nonlinear transformation, as shown in the following formula:

3. The method as described in claim 1, characterized in that, The network training method described in step 2 includes: Step 21: Use the track dataset as input to the network structure. After passing through convolutional or pooling layers, the feature map is calculated using the activation function until the final prediction result is obtained. Step 22, in the prediction result, x l,i The x-coordinate value is the left-side horizontal axis value. r,i is the x-coordinate value of the right rail; T is the number of points after the single rail is truncated in the y-axis direction, and the value of i ranges from [1,T]. Step 23, the loss function L is as follows: L=L l +α1L s +α2L m Among them, L l For linear loss; L s For volatility loss; L m α1 and α2 are the scaling factors for volatility loss and cutoff loss, respectively. Linear loss L l The formula is as follows: Among them, I l,i ,I r,i U represents the intersection of the horizontal expansion of the i-th predicted point and the actual point on the left and right rails, respectively; l,i U r,i Let be the union of the horizontal expansions of the i-th predicted point and the actual point on the left and right rails, respectively. The calculation formula is as follows: Where the superscript ^ represents the actual value; w is the horizontal expansion width; Volatility Loss L s The formula is as follows: Where the superscript ^ represents the true value; r is the proportion of the number of sign changes of the first-order difference of the x-coordinates of all points on the track to the total number of points; β s The smoothing point for the β-proportional loss is: Cut-off loss L m The formula is as follows: Where the superscript ^ represents the true value; β m The proportional smoothing point for truncating the loss; Step 24: Using the Newton-Range Accelerated Gradient Descent method with adaptive momentum, continuously update the network parameters based on the latest calculated loss function value until the loss function value converges or the preset number of iterations is reached, thus obtaining the track recognition model.

4. The method as described in claim 1, characterized in that, Step 3: Obtain the coordinates of the identified orbit, including: Step 31: Input the real-time track image into the track recognition model. After passing through the convolutional layer or pooling layer, the feature map is calculated by the activation function until the fully connected layer is reached to obtain the final detection result. Step 32: The detection result is a vector of length 65, where the 1st to 32nd values ​​are the predicted values ​​of the left horizontal axis, i.e., x. l,i The 33rd to 64th values ​​are the predicted x-coordinate values ​​for the right track. r,i The 65th value is the number of points after the single track is truncated, i.e., T; the range of i is [1, T]. Step 33: In step 32, a total of 64 orbital points were obtained. The left orbital coordinate sequence is (x... l,i ,y i The right-hand orbital coordinate sequence is (x r,i ,y i ), y-axis Where H is the image height, and i ranges from [1, 32]; combined with the number of points T after the single-track truncation, the range of i after truncation is [1, T].

Images