CN117079277A - Traffic scene real-time semantic segmentation method based on deep learning - Google Patents

Abstract

The invention discloses a real-time semantic segmentation method of traffic scenes based on deep learning, which includes the following steps. Step 1: Obtain training images and preprocess the training images; Step 2: Construct a traffic scene semantic segmentation network, and the encoder during construction The MobileNetV2 backbone feature extraction network is used to form the DeeplabV3+ network model of the backbone network MobielNetV2; the training images are used to train the traffic scene semantic segmentation network to obtain the semantic segmentation network model; step 3, perform real-time semantic segmentation of the traffic scene based on the semantic segmentation network model. By improving the DeepLabv3+ decoder and loss function, the segmentation accuracy of the existing semantic segmentation model is improved, while the amount of calculation and parameters are reduced, so that it can be better obtained in vehicle-mounted embedded platforms with limited hardware storage and computing power. widely used.

Description

A method of real-time semantic segmentation of traffic scenes based on deep learning

Technical field

The invention belongs to the field of computer vision technology, and specifically belongs to a real-time semantic segmentation method of traffic scenes based on deep learning.

Background technique

The semantic segmentation task is a key and basic technology and belongs to the field of image segmentation. The core idea is to assign semantic information consisting of specific predefined categories to each pixel in the image through a certain method. Specifically, basic classification networks that have good performance in image classification tasks, such as AlexNet and VGG, are used as backbone networks to extract corresponding features, improve them according to different application scenarios, and then restore the features to The size of the original image is used to generate the segmented image. Semantic segmentation has a wide range of applications, such as remote sensing image analysis, scene analysis, etc.

In addition, in the field of vehicle autonomous driving environment perception, image semantic segmentation is a key technology that can classify target objects such as traffic participants, road edges, and obstacles ahead in traffic road scene images pixel by pixel. According to segmentation The results further identify target types, obtain information about obstacles in front of the vehicle, and plan the passable range of vehicles in the passable area, etc., thereby providing relevant environment perception results for the autonomous driving system.

However, on urban roads, the traffic scene is more complex and has diverse targets, including not only vehicles, roads, buildings, vegetation, and nearby pedestrians, which account for a large proportion of objects, but also traffic signs, street lights, small non-motorized vehicles, and Pedestrians, vehicles and other objects in the distance occupy a smaller proportion. Autonomous driving and vehicle-assisted driving systems are the main applications of traffic scene semantic segmentation tasks. Their purpose is to enable vehicles to realize automatic navigation, avoid vehicles and pedestrians, and identify obstacles and traffic signs without human operation or assisted operation. Card tasks. Therefore, the accuracy and real-time performance of semantic segmentation are crucial to the safety of drivers, road vehicles and pedestrians. As a pixel-level image recognition technology, image semantic segmentation is more accurate than other technologies, but it is also more difficult. The algorithm running time, segmentation accuracy, and hardware storage space need to be strictly controlled. It is not only necessary to identify and segment relatively large objects in the vicinity, but also to accurately segment distant or smaller objects to facilitate early judgment of road conditions. Moreover, the semantic segmentation algorithm takes a long time, and the collection and annotation of data volume and the complexity of the algorithm need to be continuously improved, so that it can achieve true autonomous driving and ensure the safety of people and vehicles. In addition, due to the complexity and cumbersomeness of the semantic segmentation network model, its large number of parameters and calculations lead to slow network reasoning. It is not suitable for deployment on vehicle-mounted chip platforms with limited computing resources and low time-consuming requirements. It is difficult to take into account real-time at the same time. performance and accuracy.

To sum up, the existing technology has problems such as inaccurate segmentation of small objects in traffic scene images, making it difficult to apply on vehicle-mounted chip platforms with limited computing resources and low time-consuming requirements.

Contents of the invention

In order to solve the problems existing in the existing technology, the present invention provides a real-time semantic segmentation method of traffic scenes based on deep learning, which is used to solve the problem of inaccurate segmentation of small objects in existing traffic scene images and difficulty in computing resources and limitations. Problems with in-vehicle chip platform applications with lower time-consuming requirements. By improving the DeepLabv3+ decoder and loss function, a traffic scene semantic segmentation method based on deep learning is proposed, which improves the segmentation accuracy of the existing semantic segmentation model, while reducing the amount of calculation and parameters, so that it can better It is widely used in vehicle embedded platforms with limited hardware storage and computing power.

In order to achieve the above objects, the present invention provides the following technical solutions:

A method of real-time semantic segmentation of traffic scenes based on deep learning, including the following steps,

Step 1: Obtain training images and preprocess the training images;

Step 2: Construct a traffic scene semantic segmentation network. The encoder during construction uses the MobileNetV2 backbone feature extraction network to form the DeeplabV3+ network model of the backbone network MobielNetV2; use training images to train the traffic scene semantic segmentation network to obtain a semantic segmentation network model;

Step 3: Perform real-time semantic segmentation of traffic scenes based on the semantic segmentation network model.

Preferably, in step 1, the preprocessing process specifically includes converting the training image into a tensor format, adjusting the image size for normalization, and adding noise for data enhancement through random flipping.

Preferably, step 2 specifically includes the following steps:

Step 2.1, the encoder when constructing the traffic scene semantic segmentation network uses MobileNetV2 lightweight network architecture to extract features, and improves Xception's DeeplabV3+ network into MobielNetV2's DeeplabV3+ network; MobielNetV2's DeeplabV3+ network obtains advanced semantic features and Low-level semantic features;

Input high-level semantic features into the ASPP module, which introduces deep convolutional neural network, CBAM attention mechanism module and SE attention mechanism module;

Step 2.2, input the training image into the DeeplabV3+ network model and use the Focal Loss loss function for training; after entering the deep convolutional neural network, the corresponding low-level semantic features are obtained;

Step 2.3, perform a 1×1 convolution operation through the ASPP module in step 2.1 to obtain multi-scale context information;

Step 2.4, through a 3×3 convolution operation and 4 times bilinear upsampling, the output segmented image is obtained;

Step 2.5, perform decoding operations through the Decoder module to obtain the results after semantic segmentation.

Further, in step 2.1, the input and output of the residual part of the MobileNetV2 lightweight network are directly connected.

Further, in step 2.1, the DeeplabV3+ network of MobielNetV2 first uses 1×1 convolution to increase the dimension, then uses 3×3 depth-separable convolution for feature extraction, and finally uses 1×1 convolution to reduce the dimension;

After completing the feature extraction of MobielNetV2, two effective feature layers are obtained. One of the effective feature layers is the result of compressing the height and width of the input image twice, and the other effective feature layer is the result of compressing the height and width of the input image four times.

Further, in step 2.1, the formula of the depth-separable convolution parameter in the ASPP module is,

h·w·C _in ·C _out +C _in ·C _out

In the formula, the depth of the input tensor is C _in , the depth of the output tensor is C _out , h and w are the width and height of the image respectively.

Further, in step 2.1, the formula of the CBAM attention mechanism module in the ASPP module is,

Among them, M _c (F) is the channel attention module, M _s (F) is the spatial attention module, the input function is represented by F; the Sigmoid activation function is represented by σ, W ₀ and W ₁ represent the parameters of the fully connected layer, and/> They are the features after average pooling and maximum pooling respectively; f ^7×7 represents a convolution operation with a convolution kernel size of 7×7. The spatial attention module is more concerned about the feature map at the spatial level.

Preferably, in step 2, the SE attention mechanism module formula in the ASPP module is,

Among them, u represents the convolved feature map, the number of channels of u is C, and W×H is the spatial dimension of u.

Preferably, in step 2.2, the Focal Loss loss function formula is,

CE(p,y)＝CE(p _t )＝-log(

Among them, p refers to the probability that the predicted sample is a positive sample; y is the real label, and p _t represents the possibility that the sample belongs to the real category.

Compared with the existing technology, the present invention has the following beneficial technical effects:

The present invention provides a real-time semantic segmentation method of traffic scenes based on deep learning, which uses MobileNetV2 lightweight network architecture to extract features and reduce the amount of calculation. In the ASPP module, depth-separable convolution is used instead of ordinary convolution to reduce the number of parameters. And CBAM and channel attention mechanism SE-attention are added to solve the problem of less correlation between different features when merging features at multiple scales, and improve the prediction accuracy of the model. And replace the cross-entropy loss function with the Focal Loss loss function to solve the problem of imbalance of positive and negative samples and imbalance of difficult-to-classify samples. The present invention uses the PASCAL VOC data set to train and test the improved network model. The results show that the improved DeepLabV3+ network model has an average intersection-over-union ratio (mIoU) of 80.04%, an accuracy of 92.8%, and a parameter amount reduced to 6.13M. Compared with the original DeepLabV3+, the mIoU and accuracy rates are increased by 1.08% and 2.1% respectively, the parameter amount is reduced by 48.58M, and the calculation amount is reduced by 10.82GFlop. The network proposed by the present invention can effectively improve the accuracy of image segmentation, accelerate the speed of semantic segmentation, and reduce the computing load of the device.

Description of the drawings

Figure 1 is the reverse residual block.

Figure 2 is the MobileNetV2 convolution block structure.

Figure 3 is an architecture diagram of the lightweight traffic scene semantic segmentation network improved by DeepLabV3+.

Figure 4 is the mIoU curve of the original model and the improved model.

Figure 5 is the mPA curve of the original model and the improved model.

Figure 6 is the loss curve for a total of 30,000 iterations.

Figure 7 is a partial visualization.

Detailed ways

The present invention will be further described in detail below with reference to specific examples, which are explanations rather than limitations of the present invention.

The invention provides a traffic scene semantic segmentation method based on deep learning, which includes the following steps:

Step 1: Obtain the file path of the training image and the corresponding label, read the image, and convert the image into tensor format. The purpose is to create a higher-dimensional matrix, resize the image size and perform normalization processing. Unification processing is to limit the preprocessed data to a certain range [0,1] or [-1,1]. Data enhancement is then performed by randomly flipping (horizontally, up and down) the image, adding noise, etc. Data enhancement can increase the amount of training data and improve the generalization ability of the model; increase noise data and improve the robustness of the model.

Step 2: Construct a traffic scene semantic segmentation network and improve it on the basis of the benchmark model. Use the Cityscapes data set to train the traffic scene semantic segmentation network to obtain the semantic segmentation network model proposed by the present invention. The specific steps are as follows:

Step 2.1: The DeepLabV3+ network first extracts input image features through a deep convolutional neural network (DeepConvolutional Neural Network, DCNN), thereby obtaining high-level and low-level semantic features. The advanced semantic features are input to the ASPP module, which includes 4 dilated convolution layers with different dilation rates and 1 pooling layer. After convolution of the dilated convolution layer and the pooling operation of the pooling layer, 5 feature map to obtain multi-scale information.

Step 2.2: Input the images in the Cityscapes dataset into the improved DeepLabV3+ network, and obtain the corresponding low-level semantic features after entering the DCNN. In order to reduce the number of parameters in the convolution operation, depth-separable convolution is introduced into the ASPP module to further compress the model and improve the network inference speed.

Step 2.3: In order to improve the accuracy of the model and highlight the target features in a large amount of background information, the present invention adds an attention mechanism module so that the network can learn important features on the image without much increase in time consumption and training parameters. In order to achieve a balance between accuracy and model calculation cost. Based on the original four branches of the ASPP module, in order to weight different areas and gather important information to generate output, the present invention adds a fifth branch, which introduces SE attention in the pooling layer. The mechanism, focusing on channel information, refines the feature maps extracted from the backbone network in the channel dimension, enabling the network to perform in-depth learning from information-rich channels, thereby effectively capturing the feature relationships between feature channels. After SE attention, the CBAM attention mechanism is introduced to enable the network to learn more effective features when the position information of the input feature has higher importance in both channel and spatial position.

Step 2.4: Use the above-mentioned improved ASPP module and then perform a 1×1 convolution operation to obtain multi-scale context information. By using multiple scales, more comprehensive information is extracted, including both global overall information and local detailed information.

Step 2.5: Finally, through a 3×3 convolution operation and 4 times bilinear upsampling, the output segmented image is obtained.

In some difficult sample images, image edge details often cannot be segmented well, and some irregularly shaped target objects are difficult to classify correctly, which will lead to a large number of negative samples. That is to say, the sample is easy to classify, so it dominates the gradient update direction, the network's ability to learn useful information decreases, and it cannot accurately segment the target object. The present invention replaces the cross-entropy loss function in the benchmark model with the Focal Loss loss function to solve the above problems. Through the Focal Loss loss function, the dominant role of simple samples in the gradient update direction is weakened, and the network is prevented from learning a large amount of useless information. At the same time, it can prevent the model from shifting to categories with more samples and alleviate the problem of category imbalance.

In step 2 of the present invention, the MobileNetV2 lightweight network architecture is used in the encoder to extract features and reduce the amount of calculation. MobileNetV2 introduces the ResNet residual idea, combining high-dimensional features with a linear rectification (Rectified Linear Unit, ReLU) function. The core idea is to design a reverse residual block, as shown in Figure 1. First, the number of channels of the input low-dimensional features is expanded by 6 times to obtain high-dimensional features, and then depth-separable convolution processing is performed. Then the ReLU activation layer is removed, and then replaced with a mapping layer, and finally the low-dimensional features are output.

The MobileNetV2 convolution block structure is shown in Figure 2. When stride=1, the dimensions of the input features and the output features are the same, and a residual structure is introduced while the block with stride=2 does not have a residual structure. The reverse residual structure of the MobileNetV2 network improves the efficiency of memory usage. Compared with the MobileNetV1 network, the reverse residual structure can also extract more features.

In step 2 of the present invention, depth-separable convolution is introduced in the ASPP module. For a certain convolution layer, if the depth of the input tensor is C _in and the depth of the output tensor is C _out , h and w are respectively The width, height and depth of the image are separable convolution parameters as shown in formula (1):

h·w·C _in ·C _out +C _in ·C _out (1)

The parameter ratio between depthwise separable convolution and standard convolution is shown in formula (2):

It can be seen that using depth-separable convolution under the same conditions can effectively reduce the number of model parameters. Therefore, the depth-separable convolution method is used in this invention to improve the semantic segmentation model architecture.

In step 2 of the present invention, a CBAM attention mechanism module (Convolutional BlockAt-tentionModule) is added, as shown in formula (3) and formula (4):

Among them, M _c (F) is the channel attention module, M _s (F) is the spatial attention module, the input function is represented by F; the Sigmoid activation function is represented by σ, W ₀ and W ₁ represent the parameters of the fully connected layer, and/> They are the features after average pooling and max pooling respectively. f ^7×7 represents a convolution operation with a convolution kernel size of 7×7. The spatial attention module is more concerned about the feature map at the spatial level.

In step 2 of the present invention, the SE attention mechanism module (Squeeze-and-ExcitationModule, SEModule) is added, as shown in formula (5):

Among them, u represents the convolved feature map, the number of channels of u is C, and W×H is the spatial dimension of u.

Next, two fully connected layers (FC) and excitation operations are used to obtain the information of the compressed real sequence, thereby improving the module nonlinearity. As shown in formula (6):

s＝σ[W ₂ δ(W ₁ z)] (6)

Among them, W ₁ and W ₂ are the parameters of the two fully connected layers, δ is the nonlinear activation function ReLU, and σ is the Sigmoid activation function.

Finally, the original function is multiplied channel by channel by the channel weight coefficient obtained by the excitation operation to obtain the weighted features. As shown in formula (7):

In step 3 of the present invention, the Focal Loss loss function is used to solve the problem of imbalance between positive and negative samples. Formula (8) is a binary cross-entropy loss function. Among them, p refers to the probability that the predicted sample is a positive sample; y is the true label. For the convenience of expression, p _t is used to express the possibility that the sample belongs to the real category, and the expression method is as follows: formula (9). From formula (8) and formula (9), we can know that the expression form of the CE function is as formula (10). According to formula (10), the expression form of Focal Loss loss function is obtained, as shown in formulas (11) and (12).

CE(p,y)＝CE(p _t )＝-log(p _t ) (10)

FL(p _t )＝-α _t (1-p _t ) ^γ log(p (12)

This invention uses MobileNetV2 as the backbone feature extraction network, reduces the amount of parameters by 48.58M, reduces the amount of calculation by 10.82GFlops, and applies it to the ASPP module. Secondly, the attention mechanism is used to improve the segmentation accuracy of the image. On this basis, the Focal Loss loss function is used to solve the problem of imbalance between positive and negative samples. The network proposed by the present invention can effectively improve the accuracy of image segmentation, accelerate the speed of semantic segmentation, and reduce the computing load of the device.

The following is explained through a specific embodiment.

The data set used for training the method of the present invention is a high-precision annotated data set of the semantic segmentation part in Citscapes. The data set contains a training set, a verification set and a test set. Each image in the data set corresponds to annotations of the image, including strength segmentation annotations and semantic segmentation annotations. This invention only uses semantic segmentation annotation. There are 2975 pictures in the training set and 500 pictures in the verification set.

This invention uses Pytorch as the deep learning development framework, the programming language is python3.7, and is based on the Windows 10 system. The effectiveness of the method is verified through experiments. The hardware conditions of the experiment are Intel Xeon(R) Gold6226R with a main frequency of 2.90GHz and 32GB of memory. The GPU is Nvidia GRIDRTX8000P-24Q and the video memory is 24GB. During the network training process, the improved network uses the same hyperparameters as the original network. The initial learning rate is 7e-3, the weight decay is 1e-4, and the epoch is 100. The sgd optimization algorithm is used for learning, the momentum is set to 0.9, and the loss function is the Focal Loss function.

Specifically, the following steps are used to achieve this:

Step 1: Change the benchmark model (the backbone network is Xception's DeeplabV3+ network) to the DeeplabV3+ network of MobielNetV2. The backbone part of MobielNetV2 first uses 1×1 convolution for dimensionality enhancement, and then uses 3×3 depth separable convolution for feature extraction. Extraction, and finally using 1×1 convolution to reduce dimensionality. The input and output of the residual part of MobielNetV2 are directly connected. After completing the feature extraction of MobielNetV2, two effective feature layers are obtained. One of the effective feature layers is the result of compressing the height and width of the input image twice, and the other effective feature layer is the result of compressing the height and width of the input image four times.

Step 2: Based on step 1, in order to reduce the amount of parameters and calculations of ordinary convolution operations, depth-separable convolution is introduced in the first four branches of the ASPP module. First, use the same number of convolution kernels as the number of channels to perform 2D convolution on the feature maps of different channels. Several extracted features are independent in the channel direction. Then, a 3D convolution with a convolution kernel size of 1×1 is performed through point-by-point convolution. Product operation is used to extract features in the channel direction on the feature map obtained by depth convolution.

Step 3: On the basis of step 2, in order to weight different areas, important information can be gathered to produce output, a fifth branch is added based on the original 4 branches, and the fifth branch is introduced on the pooling layer The SE attention mechanism, focusing on channel information, refines the feature maps extracted from the backbone network in the channel dimension, enabling the network to perform in-depth learning from information-rich channels, thereby effectively capturing the feature relationships between feature channels. After SE attention, the CBAM attention mechanism is introduced to enable the network to learn more effective features when the position information of the input feature has higher importance in both channel and spatial position.

Step 4: Due to the small number of training samples in the Cityscapes data set, and the number of each category and the 21 target categories in the neural network learning data set cannot be uniformly distributed in complex scenarios. The Focal Loss loss function is used to replace the Cross Entropy Loss Function in DeepLabv3+ to balance the imbalance of each classification in the Cityscapes data set.

Figures 1 and 2 are the MobileNetV2 reverse residual block and convolution block structures respectively.

Figure 3 is the improved DeepLabv3+ network. The image is input into the network, and the corresponding low-level semantic features are obtained after entering the DCNN. In order to reduce the number of parameters in the convolution operation, depthwise separable convolution is introduced into the ASPP module. In addition, in order to improve the accuracy of the model and highlight the target features in a large amount of background information, the present invention adds an attention mechanism module so that the network can learn important features on the image without much increase in time consumption and training parameters. features, thereby achieving a balance between accuracy and model calculation cost. Through the above-mentioned improved ASPP module, a 1×1 convolution operation is performed to obtain multi-scale context information. Finally, the Decoder module performs the decoding operation and obtains the result after semantic segmentation. The present invention uses the Focal Loss function to solve the imbalance problem between positive and negative samples, thereby improving the segmentation accuracy of the network.

Figure 4 is the average intersection-over-union (mIoU) curve of the original model and the improved model, which increased from 78.96% to 80.04%. The present invention trains and verifies the improved DeepLabV3+ network on the Cityscapes data set to obtain the mIoU curve. The final model improves mIoU by 1.08% compared with the baseline model.

Figure 5 is the average intersection-over-union (mIoU) curve of the original model and the improved model, which increased from 78.96% to 80.04%. The present invention trains and verifies the improved DeepLabV3+ network on the Cityscapes data set to obtain the mPA curve. The final model improves mPA by 2.1% compared with the baseline model.

Figure 6 shows the loss function curve after a total of 30,000 iterations of ablation experiments on the loss function, feature extraction network and attention mechanism. It can be seen that the convergence effect of the Focal Loss loss function is better. After the FocalLoss function replaced the cross-entropy loss function, the mIoU increased to 80.04% while maintaining the accuracy of image segmentation. The average value of loss decreased from 0.0925 to 0.0185, and the average value of val_loss decreased from 0.1078 to 0.1074. It can be seen that using the Focal Loss loss function to deal with the imbalance of positive and negative samples has a certain effect.

Figure 7 shows some of the visualization results of the Cityscapes data set. From left to right are the original image, the pre-improved segmentation structure, and the improved segmentation results. It can be seen that the accuracy of the details of the model of the present invention is significantly improved compared with the baseline model.

The present invention is a traffic scene semantic segmentation method based on deep learning. This method improves the DeepLabV3+ network from three aspects: ① In the encoder, the MobileNetV2 lightweight network architecture is used to extract features and reduce the amount of calculation. In the ASPP module, depth-separable convolution is used instead of ordinary convolution to reduce the number of parameters. ② Add CBAM and channel attention mechanism SE-attention to solve the problem of less correlation between different features when fusing multi-scale features, and improve the prediction accuracy of the model. ③Replace the cross-entropy loss function with the Focal Loss loss function to solve the problem of imbalance of positive and negative samples and imbalance of difficult-to-classify samples. The experiment used the PASCAL VOC data set to train and test the improved network model. The results show that the improved DeepLabV3+ network model has an average intersection-over-union ratio (mIoU) of 80.04%, an accuracy of 92.8%, and the number of parameters reduced to 6.13M. Compared with the original DeepLabV3+, the mIoU and accuracy rates are increased by 1.08% and 2.1% respectively, the parameter amount is reduced by 48.58M, and the calculation amount is reduced by 10.82GFlop.

The specific embodiments of the present invention are given above. It should be noted that the present invention is not limited to the following specific embodiments. All equivalent transformations based on the solution of this application fall within the protection scope of the present invention.

Claims (9)

1. A real-time semantic segmentation method of traffic scenes based on deep learning, which is characterized by including the following steps: Step 1: Obtain training images and preprocess the training images; Step 2: Construct a traffic scene semantic segmentation network. The encoder during construction uses the MobileNetV2 backbone feature extraction network to form the DeeplabV3+ network model of the backbone network MobielNetV2; use training images to train the traffic scene semantic segmentation network to obtain a semantic segmentation network model; Step 3: Perform real-time semantic segmentation of traffic scenes based on the semantic segmentation network model. 2. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 1, characterized in that in step 1, the preprocessing process specifically includes converting the training image into a tensor format and adjusting the image size. Normalization is performed, and data augmentation is performed by adding noise through random flipping. 3. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 1, characterized in that step 2 specifically includes the following steps: Step 2.1, the encoder when constructing the traffic scene semantic segmentation network uses MobileNetV2 lightweight network architecture to extract features, and improves Xception's DeeplabV3+ network into MobielNetV2's DeeplabV3+ network; MobielNetV2's DeeplabV3+ network obtains advanced semantic features and Low-level semantic features; Input high-level semantic features into the ASPP module, which introduces deep convolutional neural network, CBAM attention mechanism module and SE attention mechanism module; Step 2.2, input the training image into the DeeplabV3+ network model and use the Focal Loss loss function for training; after entering the deep convolutional neural network, the corresponding low-level semantic features are obtained; Step 2.3, perform a 1×1 convolution operation through the ASPP module in step 2.1 to obtain multi-scale context information; Step 2.4, through a 3×3 convolution operation and 4 times bilinear upsampling, the output segmented image is obtained; Step 2.5, perform decoding operations through the Decoder module to obtain the results after semantic segmentation. 4. A method of real-time semantic segmentation of traffic scenes based on deep learning according to claim 3, characterized in that in step 2.1, the input and output of the residual part of the MobileNetV2 lightweight network are directly connected. 5. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 3, characterized in that, in step 2.1, the DeeplabV3+ network of MobielNetV2 first uses 1×1 convolution for dimensionality increase, and then uses 3×3 Depthwise separable convolution is used for feature extraction, and finally 1×1 convolution is used to reduce dimensionality; After completing the feature extraction of MobielNetV2, two effective feature layers are obtained. One of the effective feature layers is the result of compressing the height and width of the input image twice, and the other effective feature layer is the result of compressing the height and width of the input image four times. 6. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 3, characterized in that, in step 2.1, the depth-separable convolution parameter formula in the ASPP module is, h·w·C _in ·C _out +C _in ·C _out In the formula, the depth of the input tensor is C _in , the depth of the output tensor is C _out , h and w are the width and height of the image respectively. 7. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 3, characterized in that in step 2.1, the CBAM attention mechanism module formula in the ASPP module is, M _c (F)=σ(MLP(AvgPool(F) MLP(MaxPool(F))＝ Among them, M _c (F) is the channel attention module, M _s (F) is the spatial attention module, the input function is represented by F; the Sigmoid activation function is represented by σ, W ₀ and W ₁ represent the parameters of the fully connected layer, and/> They are the features after average pooling and maximum pooling respectively; f ^7×7 represents a convolution operation with a convolution kernel size of 7×7. The spatial attention module is more concerned about the feature map at the spatial level. 8. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 1, characterized in that, in step 2, the SE attention mechanism module formula in the ASPP module is, Among them, u represents the convolved feature map, the number of channels of u is C, and W×H is the spatial dimension of u. 9. A real-time semantic segmentation method of traffic scenes based on deep learning according to claim 1, characterized in that in step 2.2, the Focal Loss loss function formula is, CE(p,y)＝CE(p _t )＝-log( Among them, p refers to the probability that the predicted sample is a positive sample; y is the real label, and p _t represents the possibility that the sample belongs to the real category.