Road traffic sign target detection method based on YOLOv8

Abstract

The invention discloses a road traffic sign target detection method based on YOLOv8, and relates to the field of road target detection, in the method, a CDFF-YOLO model is constructed, and an MPA module is embedded in a feature fusion network of the model to enhance the feature fusion capability of C2f; an MSF module is constructed, GSConv and CARAFE up-sampling operators are embedded to realize a multi-dimensional sequence information fusion effect, and the detection precision of the model is improved while the parameter quantity and the calculation quantity of the model are reduced as much as possible. A DFF module is constructed, two-dimensional Fourier transform and inverse transform methods are adopted to enable feature information to be enhanced in a frequency domain and a space domain at the same time, and the detection performance of a network model for shielded small targets and small targets under uneven illumination is effectively improved. The three modules not only enhance the detection effect of the small target model, but also enable the network to stably cope with the detection scene of the small target. According to the method, the detection performance of the small road traffic sign target is excellent, the processing speed is high, and deployment of mobile terminal equipment is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of road target detection, and in particular to a road traffic sign target detection method based on YOLOv8. Background Art

[0002] Nowadays, cars have become an essential means of transportation for every household, so self-driving cars and intelligent driving systems have emerged. How to use intelligent driving systems to avoid potential safety hazards on the road has become a hot issue in recent years. Traffic signs, as important signs on the road, contain semantic information such as road conditions and traffic signals. Therefore, TSDR (traffic sign detection and recognition) is very important for understanding road conditions and preventing safety risks, and plays an indispensable role in driving safety. In practical applications, the size of traffic signs is often not large enough, and they are prone to fading, damage and occlusion. In addition, traffic signs are also affected by various bad weather conditions, including rain, snow and windy and sandy weather. At the same time, since the distance between the vehicle and the traffic sign is constantly changing during driving, the feature information is also constantly changing. Therefore, the traffic sign detection system must ensure the accuracy of detection and consider the detection efficiency in actual scenarios, which is still a major challenge. Traditional methods based on manual design and image shape and color are obviously difficult to meet these requirements.

[0003] Existing deep learning-based target detection algorithms often perform well in detecting large targets, but perform poorly in detecting small targets. The common method of introducing the P2 detection layer can improve the accuracy of small target detection to a certain extent, but the huge number of parameters and computations brought by this method will pose a huge challenge to the deployment of mobile terminals. Therefore, neither the traditional detection method of manually designed features nor the existing deep learning method has sufficient applicability in the scenario of small target detection, and cannot achieve a good coordination between detection accuracy and detection efficiency, which brings challenges to recognition and detection in the scenario of small target detection.

[0004] In the scene of traffic sign detection, small targets often have too few pixels and insufficient feature information. In this case, it is difficult to solve the problem of target detection by only using the existing dimensional data information. Since three-dimensional data information carries depth information, it contains data that shows the complete set information in the three-dimensional scene. Using three-dimensional data to detect small targets can effectively deal with the problem of lack of feature information, but it will consume more computing power and use more complex models in the algorithm processing stage. How to strike a balance between computational complexity and detection accuracy is the key point of using this technology. The C2f module of YOLOv8 was designed to handle the task of detecting small targets that need to retain spatial information at the beginning of its development. However, the C2f module still performs poorly when there are multiple small targets in the scene and the background environment is relatively complex. Therefore, solving the problem of the poor performance of the C2f module in small target detection is also a top priority in the field of target detection. Sometimes the targets to be detected will overlap or blur. Image reconstruction is generally used to solve these problems. However, traditional image reconstruction methods cannot complete real-time tasks, so how to deal with the problem of blurring and occlusion of small targets in this case requires further exploration. Summary of the invention

[0005] In view of the above problems and technical requirements, the inventors have proposed a road traffic sign target detection method based on YOLOv8. In view of the problems of low recall rate and inaccurate positioning of the original target detection algorithm when detecting small targets, the present invention effectively improves the performance of C2f in the feature fusion network, and designs a new convolutional layer structure C2f-MPA using the MPA module; at the same time, in view of the problem of lack of network feature fusion capability, the feature fusion module is reconstructed and the MSF module is designed. In addition, in order to deal with the problem of small traffic sign targets being blocked and uneven lighting in the detection scene, the DFF module is designed, thereby constructing a new CDFF-YOLO model. The technical solution of the present invention is as follows:

[0006] A road traffic sign target detection method based on YOLOv8, comprising the following steps:

[0007] Step 1: Obtain road images in different scenarios through a camera fixed on the front of the smart car, scale all images to a predetermined size, and normalize the pixel values ​​to form a data set;

[0008] Step 2: Divide the data set into training set and test set according to the predetermined ratio;

[0009] Step 3: Construct a cross-dimensional dual-domain feature fusion CDFF-YOLO model, including a backbone network, a feature fusion network, and three decoupled detection heads; among them:

[0010] The backbone network uses the darknet53 network and uses the residual connection method;

[0011] The feature fusion network uses a CDFF network that integrates multiple modules to fully integrate the feature information output by the backbone network. The modules include spatial depth conversion convolution SPDConv, C2f-multi-branch perception attention MPA module, multi-dimensional spatial information fusion MSF module and dual-domain feature fusion DFF module; SPDConv is used to transmit the feature information of the P2 detection layer; the C2f-MPA module is used to realize cross-dimensional information interaction through a multi-branch structure and multiple attention mechanisms; the MSF module uses generalized sparse convolution GSConv and content-aware feature reorganization CARAFE upsampling operator to stack and fuse sequence information of multiple dimensions to improve the detection accuracy of small targets; the DFF module uses the two-dimensional Fourier transform method to fuse the feature information of the frequency domain and the spatial domain to enhance the detection accuracy of small targets affected by interference;

[0012] The three decoupled detection heads detect the scaled feature maps, decouple the localization branch and the classification branch respectively; the decoupled head extracts the target location and category information respectively, learns them separately through different network branches, and finally fuses them;

[0013] Step 4: Input the training set images into the CDFF-YOLO model for training, including:

[0014] 1) Randomly initialize the network parameters of the model, and then input the training set images into the CDFF-YOLO model in batches;

[0015] 2) The input feature information is extracted by the backbone network and input into the feature fusion network through the SPPF layer. In addition to forward propagation, the feature information of the P3, P4, and P5 detection layers is also directly passed into the MSF module to be fused with the subsequent feature information. In addition, the feature information of the P2 detection layer is fused with the feature information of the P3 detection layer input from the backbone network after SPDConv convolution processing, and then processed by the DFF module.

[0016] 3) After being processed by the CIOU loss function and the DFL loss, the task alignment sample matching strategy is used to allocate positive and negative samples, and the IoU value between each sample and the true box is calculated. Then, the classification score of each sample is multiplied by the IoU value to obtain the task alignment score t. Finally, the top N samples in the t ranking are taken for each true box as positive samples;

[0017] 4) Calculate the loss function value with Loss as the loss function and update the weights of the unfrozen network part. In order to improve the network training efficiency, the cosine annealing learning rate is used for training. According to the number of iterations in the training process, the learning rate gradually decreases in the form of half a cosine cycle.

[0018] 5) Repeat steps 1) to 4) until the set number of iterations is reached;

[0019] Step 5: Use the test set to test the trained CDFF-YOLO model, and achieve effective cross-dimensional and dual-domain detection of road traffic signs based on the network model with qualified test evaluation indicators.

[0020] The beneficial technical effects of the present invention are:

[0021] The present invention redesigns the neck feature fusion network structure CDFF. This feature fusion network integrates SPDConv and replaces the C2f module with an improved C2f-MPA module, which effectively improves the detection performance of the network. At the same time, the MSF module and the DFF module are embedded. The MSF module can enhance the target detection capability without increasing the network model parameters and calculation amount, while the DFF module effectively improves the model's detection performance for complex detection scenes, such as when the target is blocked and the lighting is uneven. The neck adopts a multi-branch method to fuse feature information at multiple scales, and uses SPDConv to effectively introduce the P2 small target detection layer information, while avoiding the excessive parameters and calculation amount brought by the introduction of the P2 detection layer, and effectively improves the network's detection capability for small target objects. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 Flowchart of the road traffic sign target detection method provided in this application.

[0023] Figure 2 This is the structure diagram of the cross-dimensional dual-domain feature fusion CDFF-YOLO model provided in this application.

[0024] Figure 3 (a) is the neck network diagram of the traditional YOLO v8 model.

[0025] Figure 3 (b) CDFF network diagram provided for this application.

[0026] Figure 3 (c) The structure diagram of the spatial depth conversion convolution SPDConv provided in this application.

[0027] Figure 4 The C2f-multi-branch perceptual attention MPA structure diagram provided for this application.

[0028] Figure 5 This is the MPA module structure diagram provided for this application.

[0029] Figure 6Schematic diagram of the efficient channel attention mechanism ECA and spatial attention mechanism SAM provided for this application.

[0030] Figure 7 This is the structural diagram of the multidimensional spatial information fusion MSF module provided in this application.

[0031] Figure 8 This is a structural diagram of the dual-domain feature fusion DFF module provided in this application.

[0032] Fig. 9 This is the DFF-Block structure diagram provided for this application. DETAILED DESCRIPTION

[0033] The specific implementation of the present invention will be further described below in conjunction with the accompanying drawings.

[0034] The present application provides a road traffic sign target detection method based on YOLOv8, including: collecting pictures obtained by smart car cameras, pre-processing the pictures and dividing the data sets; constructing a CDFF (Cross-dimensional and Dual-domain Feature Fusion)-YOLO model, in which the feature fusion network structure CDFF is redesigned. This feature fusion network integrates SPDConv (SPace-to-Depth Convolution, spatial depth conversion convolution), SPDConv can eliminate the inherent step size of traditional convolution to reduce the loss of feature information, and can more effectively transmit P2 layer information compared to the traditional upsampling layer. At the same time, the C2f module is replaced with an improved C2f-MPA (Multi-branch Perceptual Attention, multi-branch perception attention) module, and the MSF (Multi-dimension Spatial information Fusion) module and the DFF (Dual-domain Feature Fusion, dual-domain feature fusion) module are embedded. In the C2f-MPA module, the MPA module is designed to replace the original Bottleneck module. Through the multi-branch structure and the spatial attention mechanism combined with the channel attention mechanism, cross-dimensional information interaction is realized, and the feature fusion ability of the network is enhanced. In the MSF module, GSConv (Generalized-Sparse Convolution) and CARAFE (Content-Aware ReAssembly of Features) upsampling operators are used to stack and fuse sequence information of multiple dimensions. While effectively utilizing the feature information of the backbone part, the network's detection accuracy for small targets is improved. In the DFF module, the two-dimensional Fourier transform method is used to process the feature information to enhance the detection accuracy of small targets that are disturbed (such as occluded and unevenly illuminated). The training set images are input into the improved CDFF-YOLO model for training, and finally the trained network model is output for road target detection. Combined with Figure 1 As shown, the specific implementation steps of this method are described in detail below, including the following contents:

[0035] Step 1: Obtain road pictures in different scenes through the camera fixed on the front of the smart car, scale all pictures to a uniform size of 640*640, and normalize the pixel values ​​to form a data set.

[0036] Step 2: Divide the dataset into training set and test set in a ratio of 8:2 for subsequent training and testing of the model.

[0037] Step 3: Build the CDFF-YOLO model, such as Figure 2 As shown in the figure, the model includes a backbone network, a feature fusion network and three decoupled detection heads. The backbone network selects the darknet53 network, which contains 53 convolutional layers, and a batch normalization layer and a SilU activation function are added after each convolutional layer. The combination of the three is the CBS module. The network uses the residual connection method to overcome the problem of gradient disappearance. The feature fusion network uses a CDFF network that integrates multiple modules to fully integrate the feature information output by the backbone network. The multiple modules include SPDConv, C2f-MPA module, MSF module and DFF module. The three decoupled detection heads detect the scaled feature map, decouple the positioning branch and the classification branch respectively; the decoupled head extracts the target position and category information respectively, learns them separately through different network branches, and finally fuses them, which can effectively reduce the number of parameters and computational complexity, while enhancing the generalization ability and robustness of the model.

[0038] contrast Figure 3 (a) and Figure 3 (b) It can be seen that this embodiment adds SPDConv for transmitting the characteristic information of the P2 detection layer. The specific structure of SPDConv is as follows: Figure 3 As shown in (c), SPDConv is a stepless space-to-depth convolution layer that can eliminate the inherent step size of traditional convolution to reduce the loss of feature information. The original C2f module is composed of a CBS module and a Bottleneck module in a multi-branch structure. In this embodiment, the Bottleneck is replaced with an MPA structure to obtain a C2f-MPA module, which is used to achieve cross-dimensional information interaction through a multi-branch structure and multiple attention mechanisms, effectively improving the detection performance of the network while reducing the number of model parameters and computational complexity. The structural diagram of the C2f-MPA module is shown in Figure 4 As shown in FIG, the input feature information passes through the first CBS module, the channel separation Split module and n sequentially connected MPA modules in sequence. After the output of the n MPA modules is tensor-joined with the Split module, it passes through the second CBS module to obtain the output feature map. Figure 5As shown in the figure, in the MPA module, the input feature information is forward processed by four branches. The information after complex processing has more critical feature information, while the information after simple processing retains richer information. Finally, it is fused to effectively maintain the validity of the information. In two of the branches that require complex processing, the feature patch parameter P is used by the LGA (Local and Global Attention) module to perform feature segmentation marking and feature selection to extract key features. The other branch that requires simple processing is composed of multiple convolution connections, and the information processed by the four branches is fused to reduce the loss of feature information. The fused information is processed by the efficient channel attention mechanism ECA and the spatial attention mechanism SAM, and then output after batch normalization and activation function. The structures of ECA and SAM are shown as follows Figure 6 shown.

[0039] The MSF module uses GSConv and CARAFE upsampling operators to stack and fuse sequence information of multiple dimensions, thus improving the detection accuracy of small targets without increasing the number of parameters as much as possible. Figure 7 As shown in the figure, the feature information of the P4 and P5 detection layers is processed by GSConv and then upsampled by CARAFE, and then expanded in dimension to make the extracted information richer. In addition, the feature information of the P3 detection layer is directly stacked and fused with the processed P4 and P5 information after convolution and dimension expansion. The stacked feature information is finally subjected to a 3D convolution for multi-scale feature extraction and fusion, which effectively enriches the feature information.

[0040] The DFF module uses the two-dimensional Fourier transform method to fuse the frequency domain and spatial domain feature information, effectively improving the network model's detection performance for common small targets that are obscured or small targets under uneven illumination. Figure 8 , Fig. 9 As shown in the figure, the DFF module processes the input feature information simply and then enters the X and Y branches in a ratio of 3:1 for subsequent processing and fusion. In the DFF-block of the X branch, the spatial channel attention mechanism SCAM and the frequency channel attention mechanism FCAM under the global branch are constructed. Then, the parameter learning method is used to use two learnable parameters α and β to enhance the learning ability of the module, and the dual-domain information is efficiently extracted and fused. Finally, it is merged with the forward information processed by multiple depth-by-depth convolutions DwConv under the local branch as the output of the DFF-block. Compared with ordinary convolution, DwConv can effectively reduce the number of parameters while utilizing feature information. The DFF module uses the two-dimensional Fourier transform method to transform and enhance the feature information. The formula is as follows:

[0041]

[0042] Among them, F(u,υ) represents the value of the signal in the frequency domain after the two-dimensional fast Fourier transform, and f(x,y) represents the value of the original image in the spatial domain before the transform, u and v represent the coordinates in the frequency domain, x and y represent the coordinates in the spatial domain; the range of u and x is from 0 to M-1 and the range of variables v and y is from 0 to N-1; j is an imaginary number symbol.

[0043] The working process of SCAM and FCAM is as follows:

[0044]

[0045] Among them, X FA and X SA They represent the feature information after FCAM and SCAM processing respectively; abs() represents the absolute value, IF() and F() represent two-dimensional fast Fourier transform and two-dimensional inverse fast Fourier transform respectively; X globle Represents the feature information input to the global branch; Represents the 1*1 convolution layer under FCAM and SCAM; AAP() represents adaptive average pooling.

[0046] Step 4: Input the training set images into the CDFF-YOLO model for training, including:

[0047] 1) The network parameters of the model are randomly initialized, and then the training set images are input into the CDFF-YOLO model in batches. 2) The input feature information is extracted through the backbone network Backbone and input into the feature fusion network through the SPPF layer. In addition to forward propagation, the feature information of the P3, P4, and P5 detection layers is also directly passed into the MSF module to be fused with the subsequent feature information. In addition, the feature information of the P2 detection layer is fused with the feature information of the P3 detection layer input from the backbone network after SPDConv convolution processing, and then processed by the DFF module. In the process of information processing, C2f-MPA processes the information through a multi-branch structure, and uses ECA and SCA to process the information. 3) After being processed by the CIOU loss function and the DFL loss, the task alignment sample matching strategy is used to allocate positive and negative samples, and the IoU value of each sample and the true box is calculated. Then the classification score of each sample is multiplied by the IoU value to obtain the task alignment score t. Finally, the top 10 samples of each true box are taken as positive samples. 4) Calculate the loss function value with Loss as the loss function and update the weights of the unfrozen network part. In order to improve the efficiency of network training, the cosine annealing learning rate is used for training. According to the number of iterations in the training process, the learning rate gradually decreases in the form of half a cosine cycle. The cosine annealing strategy of the learning rate is conducive to network convergence and avoids falling into the local optimum and being difficult to jump out. In addition, when the cosine annealing strategy is used to train the network, the learning rate changes smoothly and is easy for the network to learn. 5) Repeat steps 1) to 4) until the set number of iterations is reached.

[0048] Step 5: Use the test set to test the trained CDFF-YOLO model to obtain indicators such as average accuracy, recall rate, and number of detections per second. Based on the network model with qualified test evaluation indicators, effective cross-dimensional and dual-domain detection of road traffic signs can be achieved.

[0049] In order to demonstrate the detection effect of small traffic sign targets based on YOLOv8 in this invention. All experiments were conducted on the TT100k dataset, and the same experimental environment and experimental parameters were maintained. The TT100K dataset was compiled and released by the Tsinghua-Tencent Joint Laboratory, and contains a total of 100,000 Tencent Street View panoramic images. These images were taken by six high-pixel wide-angle digital SLR cameras (DSLR) under different lighting and weather conditions in multiple cities in China. In addition, the TT100K dataset contains 30,000 traffic signs, and the final size of the panoramic image after cropping is 2048×2048. The TT100K dataset contains various traffic sign categories, covering various traffic scenes. There are 221 different categories in the entire dataset, but due to the limited number of samples in some categories, only 45 categories with a large number of samples are retained. Among them, conventional objects account for 5.6%, while small targets account for 94.4%, which shows that small targets account for a large proportion in the dataset. The dataset is divided into a training set and a test set in a ratio of 8:2.

[0050] According to the third step above, the CDFF-YOLO model is constructed, the pre-trained weights are loaded and frozen into the backbone network darknet53, and the parameters of the feature fusion network and the detection head are initialized. According to the fourth step above, the training set images are input into the CDFF-YOLO network for training. During the model training process, 16 samples are used in each batch, the initial learning rate is set to 0.01, and the final learning rate is set to 0.005. During the period, it undergoes a half cosine cycle change, and the number of iterations is set to 300 times. After the training is completed, the detection effect of the model is verified with the test set. In addition, in order to verify the superiority of the present invention in the road target detection task, CDFF-YOLOv8 is compared with a variety of other models. Each method is trained with the same training set and tested with the same test set. The evaluation indicators of the test results are the single picture detection time FPS, the parameters for large, medium and small objects, the amount of calculation FLOPS, the average accuracy mAP@50 and mAP@50:95 (including prediction and post-processing). The test results are shown in Table 1:

[0051] Table 1 Evaluation indicators of detection results of each algorithm

[0052]

[0053] From the results in Table 1, it can be seen that the parameters and calculation amount of the CDFF-YOLO algorithm proposed in the present invention are not much different from those of these improved algorithms. Compared with some algorithms, the calculation amount and parameter amount are even reduced, and mAP@0.50:0.95 and mAP@0.50 are both in the leading position. At the same time, the detection speed of the algorithm proposed in the present invention also maintains 157FPS, which is enough to meet the real-time requirements and has the ability of real-time detection after being deployed on mobile devices.

[0054] From the above, it can be seen that this application proposes a road traffic sign target detection method based on YOLOv8 for the problem of missed detection and false detection in the process of small target detection of traffic signs, which increases the detection accuracy of the model while minimizing the loss of model detection efficiency. The C2f module in the YOLOv8 feature fusion network is replaced by the C2f-MPA module to improve the detection performance without reducing the detection efficiency. The MSF module is designed to realize the superposition of multi-dimensional feature information, and the feature fusion ability of the network is enhanced without increasing the number of parameters and the amount of calculation. The model uses the DFF module for feature fusion processing, and the dual-domain information processing realized by Fourier transform and inverse transform enables the model to cope with the situation of target objects being blocked and uneven lighting in small target scenes, thereby enhancing the detection performance of the model. Through comparative verification, the improved YOLOv8 algorithm model can effectively improve the detection accuracy of small targets, and the higher FPS and reasonable number of parameters and calculation amount are conducive to the subsequent mobile terminal deployment of the network model, solving the problem of small target detection of traffic signs.

[0055] The above is only a preferred embodiment of the present application, and the present invention is not limited to the above embodiments. It is understood that other improvements and changes directly derived or associated by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included in the protection scope of the present invention.

Claims

1. A road traffic sign target detection method based on YOLOv8, characterized in that: The method comprises: Step 1: Obtain road images in different scenarios through a camera fixed on the front of the smart car, scale all images to a predetermined size, and normalize the pixel values ​​to form a data set; Step 2: Divide the data set into training set and test set according to the predetermined ratio; Step 3: Construct a cross-dimensional dual-domain feature fusion CDFF-YOLO model, including a backbone network, a feature fusion network, and three decoupled detection heads; among them: The backbone network selects the darknet53 network and uses the residual connection method; The feature fusion network uses a CDFF network that integrates multiple modules to fully integrate the feature information output by the backbone network. The multiple modules include spatial depth conversion convolution SPDConv, C2f-multi-branch perception attention MPA module, multi-dimensional spatial information fusion MSF module and dual-domain feature fusion DFF module; SPDConv is used to transmit the feature information of the P2 detection layer; the C2f-MPA module is used to realize cross-dimensional information interaction through a multi-branch structure and multiple attention mechanisms; the MSF module uses generalized sparse convolution GSConv and content-aware feature reorganization CARAFE upsampling operator to stack and fuse sequence information of multiple dimensions to improve the detection accuracy of small targets; the DFF module uses a two-dimensional Fourier transform method to fuse frequency domain and spatial domain feature information to enhance the detection accuracy of small targets affected by interference; The three decoupled detection heads detect the scaled feature maps, decouple the localization branch and the classification branch respectively; the decoupled head extracts the target location and category information respectively, learns them separately through different network branches, and finally fuses them; Step 4: Input the training set images into the CDFF-YOLO model for training, including: 1) Randomly initialize the network parameters of the model, and then input the training set images into the CDFF-YOLO model in batches; 2) The input feature information is extracted by the backbone network and input into the feature fusion network through the SPPF layer. In addition to forward propagation, the feature information of the P3, P4, and P5 detection layers is also directly transmitted to the MSF module to be fused with the subsequent feature information. In addition, the feature information of the P2 detection layer is fused with the feature information of the P3 detection layer input from the backbone network after SPDConv convolution processing, and then processed by the DFF module. 3) After being processed by the CIOU loss function and the DFL loss, the task alignment sample matching strategy is used to allocate positive and negative samples, and the IoU value between each sample and the true box is calculated. Then, the classification score of each sample is multiplied by the IoU value to obtain the task alignment score t. Finally, the top N samples in the t ranking are taken for each true box as positive samples; 4) Calculate the loss function value with Loss as the loss function and update the weights of the unfrozen network part. In order to improve the network training efficiency, the cosine annealing learning rate is used for training. According to the number of iterations in the training process, the learning rate gradually decreases in the form of half a cosine cycle. 5) Repeat steps 1) to 4) until the set number of iterations is reached; Step 5: Use the test set to test the trained CDFF-YOLO model, and achieve cross-dimensional and dual-domain effective detection of road traffic signs based on the network model with qualified test evaluation indicators.

2. The road traffic sign target detection method based on YOLOv8 according to claim 1, characterized in that: In the C2f-MPA module: The input feature information passes through the first CBS module, the channel separation Split module and n sequentially connected MPA modules in sequence. After the output of the n MPA modules is tensor-joined with the Split module, it passes through the second CBS module to obtain the output feature map. In the MPA module, the input feature information is forward processed by four branches, in which the patch parameter P is used for feature segmentation marking and feature selection to extract key features through the local and global attention LGA module in two branches; the other branch is composed of multiple convolution connections, and the information processed by the four branches is fused. The fused information is processed by the efficient channel attention mechanism ECA and the spatial attention mechanism SAM, and then batch normalized and activated before output.

3. The road traffic sign target detection method based on YOLOv8 according to claim 1, characterized in that: In the MSF module: After the feature information of the P4 and P5 detection layers is processed by the GSConv, CARAFE upsampling and dimension expansion are performed; in addition, the feature information of the P3 detection layer is directly stacked and fused with the processed P4 and P5 information after convolution processing and dimension expansion. The stacked feature information is finally subjected to a 3D convolution for multi-scale feature extraction and fusion to enrich the feature information.

4. The road traffic sign target detection method based on YOLOv8 according to claim 1, characterized in that: In the DFF module: After the input feature information is processed, it enters the X and Y branches according to the predetermined ratio for subsequent processing and fusion; in the DFF-block of the X branch, the spatial channel attention mechanism SCAM and the frequency channel attention mechanism FCAM under the global branch are constructed, and then the parameter learning method is used to enhance the learning ability of the module using two learnable parameters α and β, and the dual-domain information is extracted and fused, and finally merged with the forward information after multiple depth-by-depth convolutions under the local branch as the output of the DFF-block; the action flow of SCAM and FCAM is as follows: Among them, X FA and X SA represent the feature information processed by the FCAM and the SCAM respectively; abs() represents the absolute value, IF() and F() represent the two-dimensional fast Fourier transform and the two-dimensional inverse fast Fourier transform respectively; X globle Represents the feature information input to the global branch; Represents the 1*1 convolution layer under FCAM and SCAM; AAP() represents adaptive average pooling.

Images