Traffic sign target detection method and device and storage medium

Abstract

The invention relates to a traffic sign target detection method and device and a storage medium. According to the method, after a to-be-detected traffic sign image is obtained, the to-be-detected traffic sign image is input into a trained traffic sign target detection model based on CMF-YOLOv8s, and a traffic sign target detection result is obtained; the construction and training process of the traffic sign target detection model comprises the following steps: constructing a CMF-YOLOv8s model; a traffic sign data set is obtained and preprocessed; a deep learning environment is configured, and CMF-YOLOv8s model parameters are initialized; the preprocessed traffic sign data set is input into the initialized CMF-YOLOv8s model to be trained and tested; according to the CMF-YOLOv8s model, on the basis of a YOLOv8s original model, a large target feature extraction layer and a detection head in the original model are removed, and a small target detection head is added. Compared with the prior art, the method has the advantages of high detection precision, small model parameter quantity, small calculation quantity and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic sign target detection, and in particular to a traffic sign target detection method, device and storage medium. Background Art

[0002] With the rapid development of artificial intelligence, a series of technological changes have been driven in all walks of life, and a development boom has been set off in the field of autonomous driving. At present, unmanned driving technology has entered the initial application stage, among which traffic sign detection and recognition is one of the key technologies. Accurate and rapid recognition of traffic sign information on the road can provide real-time road information for intelligent vehicles, help them make safe and efficient driving decisions, reduce the probability of traffic accidents, and improve traffic safety. However, in the actual road driving process, the traffic signs captured by the on-board camera generally have large scale variations, a large proportion of small targets, and are easily disturbed by the environment, resulting in low recognition accuracy and false detection and missed detection. This may cause misjudgment of intelligent vehicles during driving, traffic accidents, and threaten the lives of passengers. Therefore, it is crucial to accurately and efficiently identify road traffic signs.

[0003] At present, there are two types of traffic sign detection methods based on machine vision. One type is the traditional machine learning method, which requires image preprocessing and complex feature definition, extraction, and screening processes. Although these methods can achieve satisfactory detection results for images with simple backgrounds, the detection accuracy of the model is poor for traffic signs in complex environments and cannot meet the requirements of actual conditions. The second type of method is based on deep learning, including one-stage detection algorithms and two-stage detection algorithms. Among them, the two-stage detection algorithm mainly includes RCNN, Fast-RCNN, and Faster-RCNN. First, a large number of candidate areas need to be generated, and then the candidate areas are identified. Therefore, the detection speed is slow and the algorithm is too complex to meet the real-time requirements of traffic sign detection. Compared with the two-stage algorithm, the one-stage detection algorithm is more efficient and has a faster detection speed. It mainly includes SSD algorithm, RetinaNet algorithm, and YOLO series algorithm.

[0004] In terms of traffic sign detection, some existing technologies introduce CBAM modules based on YOLOv8s to improve the model's ability to recognize small targets. Although this improvement improves the detection accuracy, it also increases the number of model parameters. Other existing technologies modify the spatial attention mechanism and path aggregation network of the YOLOv4 model to improve the model's detection accuracy, but there is still a phenomenon of missed detection for small targets. Therefore, the current traffic sign detection mainly has the following problems: First, the recognition ability for small targets is still weak, and there are false detections and missed detections; second, in order to increase the detection accuracy of the model, some improvements will increase the number of model parameters or the amount of calculation, affecting the detection efficiency. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a traffic sign target detection method, device and storage medium, which can improve the recognition ability of small targets such as traffic signs while reducing the model calculation amount and ensuring detection efficiency.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] According to a first aspect of the present invention, a traffic sign target detection method based on CMF-YOLOv8s is provided, comprising the following steps: S1, obtaining a traffic sign image to be detected; S2, inputting the traffic sign image to be detected into a trained traffic sign target detection model based on CMF-YOLOv8s to obtain a traffic sign target detection result; wherein the construction and training process of the traffic sign target detection model comprises: S21, constructing a CMF-YOLOv8s model; S22, obtaining a traffic sign data set and performing preprocessing; S23, configuring a deep learning loop environment and initialize CMF-YOLOv8s model parameters; S24, input the preprocessed traffic sign data set into the initialized CMF-YOLOv8s model for training and testing; wherein, the CMF-YOLOv8s model is based on the YOLOv8s original model, removes the large target feature extraction layer and the detection head in the original model, and adds a small target detection head, the large target feature extraction layer and the detection head are used to extract target information with a dimension that meets a first preset dimension threshold, and the small target detection head is used to extract target information with a dimension that meets a second preset dimension threshold.

[0008] As a preferred technical solution, the large target feature extraction layer and detection head include a feature extraction layer and detection head with a dimension of 20×20 in the original model, and the small target detection head includes a detection head with a dimension of 160×160.

[0009] As a preferred technical solution, the CMF-YOLOv8s model includes an input end, a backbone network, a neck network and a detection head module connected in sequence, the backbone network includes a convolutional layer, an SPPF module, an improved C2f module and an improved coordinate attention mechanism, at least one of the convolutional layers is connected to the input end, the remaining convolutional layers are connected to the improved C2f module, the SPPF module is respectively connected to the improved C2f module and the improved coordinate attention mechanism, and the outputs of the improved C2f module and the improved coordinate attention mechanism are respectively connected to the neck network.

[0010] As a preferred technical solution, the improved C2f module adopts the FasterNet module to improve the original Bottleneck module; the improved coordinate attention mechanism adds global average pooling and convolution operations on the basis of the original coordinate attention mechanism, obtains global channel information, generates weights using the sigmoid function, and then multiplies these weights with the original feature map to obtain the final output feature map.

[0011] As a preferred technical solution, the neck network consists of a feature pyramid network and a path aggregation network.

[0012] As a preferred technical solution, the CMF-YOLOv8s model uses MPDIoU as the loss function.

[0013] As a preferred technical solution, in S24, the average accuracy mean and the calculation amount and parameter amount of the model are used as evaluation indicators during model testing.

[0014] As a preferred technical solution, in S22, the preprocessing process specifically includes: filtering out traffic signs whose number is greater than a preset number threshold from the traffic sign data set to obtain a traffic sign detection data set.

[0015] According to a second aspect of the present invention, there is provided a traffic sign target detection device based on CMF-YOLOv8s, comprising a memory, a processor, and a program stored in the memory, wherein the method described is implemented when the processor executes the program.

[0016] According to a third aspect of the present invention, there is provided a storage medium having a program stored thereon, wherein the program implements the method described above when executed.

[0017] Compared with the prior art, the present invention has the following beneficial effects:

[0018] 1. The present invention uses the CMF-YOLOv8s model to realize traffic sign target detection. The model is based on the original YOLOv8s model, removes the large target feature extraction layer and detection head (for example, 20×20 feature extraction layer and detection head) in the original model, and adds a small target detection head (for example, 160×160 detection head), which can effectively improve the detection accuracy of small targets such as traffic signs and reduce the number of parameters;

[0019] 2. In the CMF-YOLOv8s model provided by the present invention, the backbone network includes an improved C2f module and an improved coordinate attention mechanism. The improved coordinate attention mechanism obtains global channel information through global average pooling and convolution operations, and processes feature maps in combination with position information, which can improve the target detection accuracy of the model in complex environments. The improved C2f module uses the FasterNet module to improve the original Bottleneck module, which effectively reduces the number of parameters while improving the detection accuracy;

[0020] 3. The present invention introduces the loss function MPDIoU, which comprehensively considers factors such as non-overlapping area, center point distance, and width and height deviation through vertex coordinates, which can optimize the calculation process and further improve the model detection accuracy;

[0021] 4. Compared with other classic one-stage target detection algorithms, the present invention has the lowest model parameter amount and computational complexity, which is conducive to deployment in resource-constrained terminal devices. At the same time, it has higher detection accuracy for small targets and occluded targets in complex environments, and the detection effect of traffic sign images outside the data set is also better than traditional algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of a process for providing a method in Example 1 of the present invention;

[0023] Figure 2 A schematic diagram of a network structure of the improved coordinate attention mechanism in Embodiment 1 of the present invention;

[0024] Figure 3 Schematic diagram of the structure of the Bottleneck module and the FasterNet module in Example 1 of the present invention;

[0025] Figure 4 A schematic diagram of a network structure of the improved C2f module in Embodiment 1 of the present invention;

[0026] Figure 5 This is a diagram of the CMF-YOLOv8s network structure in Example 1 of the present invention;

[0027] Figure 6 A schematic diagram of a network structure of the SPPF module in Embodiment 1 of the present invention;

[0028] Figure 7 It is a schematic diagram of the structure of ConvModule in the CMF-YOLOv8s model in Example 1 of the present invention;

[0029] Figure 8 The traffic sign category diagram selected according to the TT 100K data set in Example 1 of the present invention;

[0030] Fig. 9 A comparison diagram of traffic sign detection results in Example 1 of the present invention;

[0031] Among them: part (a) is the original image, part (b) is the detection result of the YOLOv5s model, part (c) is the detection result of the YOLOv7 model, part (d) is the detection result of the YOLOv8s model, and part (e) is the detection result of the CMF-YOLOv8s model of this embodiment. DETAILED DESCRIPTION

[0032] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0033] Embodiment 1:

[0034] like Figure 1 As shown, this embodiment provides a traffic sign target detection method based on CMF-YOLOv8s, which specifically includes the following implementation steps:

[0035] Step S1, obtaining a traffic sign image to be detected. Optionally, the traffic sign image to be detected is real-time data or historical data.

[0036] Step S2: input the traffic sign image to be detected into the trained traffic sign target detection model based on CMF-YOLOv8s to obtain the traffic sign target detection result.

[0037] Among them, the construction and training process of the traffic sign target detection model includes:

[0038] Step S21, construct a CMF-YOLOv8s model.

[0039] The CMF-YOLOv8s model (i.e., Channel-CA, MPDIoU Loss, and Fast-C2fbased YOLOv8s model) provided in this embodiment is based on the original YOLOv8s model, removes the large target feature extraction layer and detection head in the original YOLOv8s model, adds a small target detection head, and introduces an improved coordinate attention mechanism (Channel-CA) and an improved C2f module (Fast-C2f module), and introduces the MPDIoU loss function, which reduces the number of parameters and calculations while improving the recognition accuracy of the model.

[0040] The large target feature extraction layer and the detection head are used to extract target information whose dimensions meet the first preset dimension threshold, and the small target detection head is used to extract target information whose dimensions meet the second preset dimension threshold. Optionally, the large target feature extraction layer and the detection head refer to: a feature extraction layer and a detection head with a dimension of 20×20 in the original model, and the small target detection head refers to: a detection head with a dimension of 160×160.

[0041] Figure 2 One of the network structures of the improved coordinate attention mechanism (Channel-CA) is shown. Specifically, the improved coordinate attention mechanism adds global average pooling (Avg Pool) and convolution operation (Conv2d) on the basis of the original coordinate attention mechanism, obtains global channel information, generates weights using the sigmoid function, and then multiplies these weights with the original feature map to obtain the final output feature map.

[0042] The improvement principle of Channel-CA is analyzed as follows:

[0043] In the information embedding stage, the coordinate attention mechanism first performs average pooling (X Avg Pool and Y Avg Pool) on the input feature map in the horizontal and vertical directions to encode the position information. The output is shown in the following formula:

[0044]

[0045] In the formula, x c (h,i) represents the c-th channel feature map x c In the horizontal direction of height h, x c (j,w) represents the vertical feature with width w, and Represent the output after pooling in the horizontal and vertical directions respectively.

[0046] In the coordinate attention generation stage, the concatenated feature map is passed through the convolution layer and the Sigmoid activation function to generate an intermediate feature map:

[0047] f=δ(F1([z h ,z w ])) (3)

[0048] Where δ represents the Sigmoid activation function and [·,·] represents the concatenation operation of the feature map.

[0049] Re-decompose the intermediate feature map f into tensors f in two directions h and f w , and then pass the convolution layer to make f h and f w The number of channels of the two tensors matches the input feature map, and the attention weights in the horizontal and vertical directions are:

[0050] g h =δ(F h (f h )) (4)

[0051] g w =δ(F w (f w )) (5)

[0052] In the formula, δ represents the Sigmoid activation function, g h and g w Represent the attention weights in the horizontal and vertical directions respectively, and F represents a 1×1 convolution. Finally, the attention weights are applied to the input feature map for weighting, and the output feature map is as follows:

[0053] y c (i,j)=x c (i,j)×g c (i)×g c (j) (6)

[0054] Coordinate attention adjusts the correlation between different channels by considering the position information of features in the horizontal and vertical directions, but ignores the global channel relationship. Therefore, based on the original coordinate attention mechanism, global average pooling and convolution operations are used to obtain the global correlation between channels, and multiply it with the original feature map as a weight. This improvement can consider both spatial position information and global channel dimension information, thereby better helping the network learn the characteristics of the target.

[0055] In this embodiment, the improved C2f module (Fast-C2f module) uses the FasterNet module to improve the original Bottleneck module to reduce the number of parameters and the amount of calculation of the model. Figure 3 Part (a) shows the network structure of the Bottleneck module. Figure 3Part (b) shows the network structure of the FasterNet module. Figure 4 One of the network structure diagrams of the Fast-C2f module is shown.

[0056] The CMF-YOLOv8s model uses MPDIoU as the loss function. Specifically, the introduction of MPDIoU is to replace the CIoU in the original YOLOv8s network with MPDIoU. The calculation formula of the CIoU loss function is:

[0057]

[0058] In the formula, IoU represents the intersection-over-union ratio between the predicted box and the real box. represents the Euclidean distance between the center point of the predicted box and the center point of the real box, c represents the length of the diagonal of the smallest rectangle covering the predicted box and the real box, v is used to measure the consistency of the aspect ratio between the predicted box and the real box, where w and h represent the width and height of the predicted box respectively. and is the width and height of the real frame; α is a weight used to adjust the impact of aspect ratio consistency on the loss.

[0059] In order to simplify the calculation of the similarity between the predicted box and the true box and improve the accuracy of bounding box regression, this method introduces the MPDIoU loss function, which is calculated as follows:

[0060]

[0061] Where A∩B and A∪B represent the intersection and union areas of the predicted box and the true box, respectively; d1 represents the Euclidean distance between the upper left corner of the predicted box and the true box; d2 represents the Euclidean distance between the lower right corner of the predicted box and the true box; w and h represent the width and height of the input image, respectively.

[0062] The MPDIoU loss function comprehensively considers multiple factors of bounding box regression, including non-overlapping area, center point distance, and width and height difference, through the vertex coordinates of the predicted box and the true box. This method not only covers the factors considered by CIoU, but also optimizes the calculation process. In addition, when the aspect ratio of the predicted box and the true box is consistent, MPDIoU ensures that the predicted box inside the true box has a lower loss value than the predicted box outside, thereby reducing redundant prediction boxes and improving the matching accuracy of the bounding box.

[0063] Based on the above improvements, Figure 5 One of the overall network structures of the CMF-YOLOv8s model is shown.

[0064] Specifically:

[0065] The CMF-YOLOv8s model includes the input end, backbone network, neck network and detection head module connected in sequence. The input end first resizes the input image to 640×640 pixels and uses the Mosaic algorithm for data enhancement.

[0066] The backbone network includes convolutional layers, SPPF (Spatial Pyramid Pooling Fast) modules, improved C2f modules (Fast-C2f modules) and improved coordinate attention mechanisms (Channel-CA). The backbone network extracts image features through convolutional layers, Fast-C2f modules and SPPF modules. The first convolutional layer (conv) is connected to the input end, and the remaining three convolutional layers are connected to the corresponding Fast-C2f modules. The input and output of the SPPF module are connected to the last Fast-C2f module and Channel-CA respectively. The outputs of Channel-CA and the first two Fast-C2f modules are connected to the neck network respectively. The Fast-C2f module enables the network to have rich image information while ensuring lightweight. At the same time, since the data sets used in the experiment are mostly small targets, the 20×20 feature extraction layer and detection head that extract large target information in the original network are deleted, and a branch is derived from the 160×160 feature extraction layer of the backbone network and a detection head is added in this branch as a small target detection head, thereby improving the recognition accuracy of small targets. In this embodiment, a Channel-CA module is added after the original backbone network SPPF module for feature extraction to improve the recognition accuracy of the network model. Figure 6 One of the network structures of the SPPF module is shown. Figure 7 Shows the structure of ConvModule in the CMF-YOLOv8s model.

[0067] The neck network consists of a feature pyramid network (FPN) and a path aggregation network (PAN). FPN first constructs a feature pyramid to extract multi-scale feature maps, and then fuses these feature maps through upsampling. PAN adopts a bottom-up structure to fuse multi-scale feature maps through convolutional layers. The detection head is a decoupled head structure responsible for loss calculation and target detection box screening.

[0068] Step S22, obtaining a traffic sign dataset and performing preprocessing.

[0069] In this step, the preprocessing process specifically includes: filtering out traffic signs whose number is greater than a preset number threshold from the traffic sign dataset to obtain a traffic sign detection dataset. Optionally, the traffic sign dataset uses the TT100K traffic sign image dataset, selects the top 15 traffic sign images to construct a dataset, and divides the dataset into a training set, a validation set, and a test set in a ratio of 8:1:1. The selected traffic signs are as follows: Figure 8 Specifically, 15 types of traffic signs with a large number of images are selected from traffic sign images as the training data set, with a total of 9050 images.

[0070] Step S23, configure the deep learning environment and initialize the CMF-YOLOv8s model parameters.

[0071] In this embodiment, the operating system used in the experiment is Windows 11 64-bit, the processor is Intel(R) Core(TM) i9-13980H, the graphics card is NVIDIA GeForce RTX 4060, the RAM size is 32G, the Python version used in the experiment is 3.9.7, the CUDA version is 11.6, the training cycle (epochs) is set to 100, the batch size (batchsize) is 32, the stochastic gradient descent (SGD) optimizer is used, and the initial learning rate is 1e-2.

[0072] Step S24, input the preprocessed traffic sign dataset into the initialized CMF-YOLOv8s model for training and testing.

[0073] Specifically, the training set divided from the traffic sign detection data set in step S23 is input into the initialized CMF-YOLOv8s model for training. During the training process, the validation set is used for verification, and the trained model is tested with the test set.

[0074] This embodiment uses the average precision mean and the amount of calculation and parameter of the model as evaluation indicators.

[0075] The calculation method of the mean average precision is:

[0076]

[0077] Where R is the recall rate, P is the precision, TP is the number of correctly predicted positive samples, FN is the number of incorrectly predicted negative samples, FP is the number of incorrectly predicted positive samples, AP is the area of ​​the precision and recall curve, mAP is the mean AP of all detection categories, and N is the number of detection categories.

[0078] By using steps S21 to S24, a trained traffic sign target detection model based on CMF-YOLOv8s is obtained. On the basis of the original YOLOv8s, the model adds an improved coordinate attention mechanism, a lightweight FasterNet module, an MPDIoU loss function, and a small target detection head to improve the detection accuracy of traffic signs in complex scenes, and the parameter amount and calculation amount of the CMF-YOLOv8s model are also lower than those of YOLOv8s. Based on this, the real-time traffic sign image to be detected or the historical data of the traffic sign image to be detected is input into the trained improved model to obtain the corresponding traffic sign target detection result.

[0079] In this embodiment, the trained CMF-YOLOv8s model is evaluated, and it can be found that the mAP of this method reaches 91.2% on the traffic sign detection dataset, the computational complexity is 23.8GFLOPs, and the model parameter volume is 2.7M, which meets the high precision and low computational complexity requirements of traffic sign detection. Fig. 9 The figure shows the comparison of the detection results obtained by using other existing methods and the method provided by this embodiment. Fig. 9 The results show that other existing methods have missed detection, while the CMF-YOLOv8s model with Channel-CA and small target detection layer can accurately identify traffic signs in images. In the images of the third and fourth rows, some traffic signs are blocked by green belts and railings, and the background of the traffic signs is relatively complex. YOLOv5s, YOLOv7 and YOLOv8s have missed detection to varying degrees, while the method of this embodiment correctly identifies all of them. The main reason is that the addition of the Channel-CA module enables the backbone network to better extract target features and better identify blocked targets.

[0080] Embodiment 2:

[0081] This embodiment provides a traffic sign target detection device based on CMF-YOLOv8s, including a memory, a processor, and a program stored in the memory. When the processor executes the program, the method in the embodiment is implemented.

[0082] Optionally, the device includes a data acquisition module, a target detection module, and a model construction and training module. The data acquisition module is used to acquire the traffic sign image to be detected; the target detection module is used to input the traffic sign image to be detected into the trained traffic sign target detection model based on CMF-YOLOv8s to obtain the traffic sign target detection result; the model training module is used to implement the construction and training process of the traffic sign target detection model, and the module specifically includes:

[0083] A model building unit is used to build a CMF-YOLOv8s model. The CMF-YOLOv8s model is based on the original YOLOv8s model, and the large target feature extraction layer and the detection head in the original model are removed, and a small target detection head is added. The large target feature extraction layer and the detection head are used to extract target information whose dimensions meet the first preset dimension threshold, and the small target detection head is used to extract target information whose dimensions meet the second preset dimension threshold; a data set acquisition and processing unit is used to acquire a traffic sign data set and perform preprocessing; an environment and parameter configuration unit is used to configure a deep learning environment and initialize CMF-YOLOv8s model parameters; and a training and testing unit is used to input the preprocessed traffic sign data set into the initialized CMF-YOLOv8s model for training and testing.

[0084] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of each module and unit can refer to the corresponding process in the aforementioned embodiment 1, and will not be repeated here.

[0085] Embodiment 3:

[0086] This embodiment provides a storage medium on which a program is stored, and when the program is executed, one or more steps of the method in embodiment 1 are implemented. Storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, read-only compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. According to the definition in the present invention, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carriers.

[0087] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A traffic sign target detection method based on CMF-YOLOv8s, characterized in that: The following steps are involved: S1, obtaining the image of the traffic sign to be detected; S2, inputting the traffic sign image to be detected into a trained traffic sign target detection model based on CMF-YOLOv8s to obtain a traffic sign target detection result; The construction and training process of the traffic sign target detection model includes: S21, build CMF-YOLOv8s model; S22, obtaining a traffic sign dataset and performing preprocessing; S23, configure the deep learning environment and initialize the CMF-YOLOv8s model parameters; S24, inputting the preprocessed traffic sign dataset into the initialized CMF-YOLOv8s model for training and testing; Among them, the CMF-YOLOv8s model is based on the YOLOv8s original model, removes the large target feature extraction layer and detection head in the original model, and adds a small target detection head. The large target feature extraction layer and detection head are used to extract target information whose dimension meets the first preset dimension threshold, and the small target detection head is used to extract target information whose dimension meets the second preset dimension threshold.

2. The traffic sign target detection method based on CMF-YOLOv8s according to claim 1, characterized in that: The large target feature extraction layer and detection head include a feature extraction layer and detection head with a dimension of 20×20 in the original model, and the small target detection head includes a detection head with a dimension of 160×160.

3. The traffic sign target detection method based on CMF-YOLOv8s according to claim 1, characterized in that: The CMF-YOLOv8s model includes an input end, a backbone network, a neck network and a detection head module connected in sequence, the backbone network includes a convolutional layer, an SPPF module, an improved C2f module and an improved coordinate attention mechanism, at least one of the convolutional layers is connected to the input end, the remaining convolutional layers are connected to the improved C2f module, the SPPF module is respectively connected to the improved C2f module and the improved coordinate attention mechanism, and the outputs of the improved C2f module and the improved coordinate attention mechanism are respectively connected to the neck network.

4. The traffic sign target detection method based on CMF-YOLOv8s according to claim 3 is characterized in that: The improved C2f module adopts the FasterNet module to improve the original Bottleneck module; the improved coordinate attention mechanism adds global average pooling and convolution operations on the basis of the original coordinate attention mechanism, obtains global channel information, generates weights using the sigmoid function, and then multiplies these weights with the original feature map to obtain the final output feature map.

5. The traffic sign target detection method based on CMF-YOLOv8s according to claim 3, characterized in that: The neck network consists of a feature pyramid network and a path aggregation network.

6. The traffic sign target detection method based on CMF-YOLOv8s according to claim 1, characterized in that: The CMF-YOLOv8s model uses MPDIoU as the loss function.

7. The traffic sign target detection method based on CMF-YOLOv8s according to claim 1, characterized in that: In S24, the average precision mean and the amount of calculation and parameter of the model are used as evaluation indicators during model testing.

8. The traffic sign target detection method based on CMF-YOLOv8s according to claim 1, characterized in that: In S22, the preprocessing process specifically includes: filtering out traffic signs whose number is greater than a preset number threshold from the traffic sign data set to obtain a traffic sign detection data set.

9. A traffic sign target detection device based on CMF-YOLOv8s, comprising a memory, a processor, and a program stored in the memory, characterized in that: When the processor executes the program, the method according to any one of claims 1 to 8 is implemented.

10. A storage medium having a program stored thereon, characterized in that: When the program is executed, the method according to any one of claims 1 to 8 is implemented.

Images