Expected function safety traffic sign identification method
By improving the RT-DETR model and combining it with a spatial fusion module and a multi-scale feature compression module, the problems of missed detection and false detection in small target detection are solved, the robustness and accuracy of traffic sign detection are improved, and it is suitable for vehicle terminals.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2026-03-13
AI Technical Summary
Existing traffic sign detection methods are prone to missed detections and false detections when detecting small targets, and have poor robustness in complex environments, especially with a significant decrease in detection accuracy under actual weather and lighting conditions such as rain, fog, and night.
An improved RT-DETR model is adopted, which enhances the robustness and detection accuracy of traffic signs by introducing a spatial fusion module and a multi-scale feature compression module. This includes the HGStem module, HGBlock-space module, neck network, and RT-DETR decoder for feature extraction, fusion, and detection.
While ensuring real-time performance, it significantly improves the detection accuracy and robustness of small-sized traffic signs, especially in complex environments, enhancing the ability to identify small targets and resist interference.
Smart Images
Figure CN121661616A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer vision technology, and specifically relates to a method for recognizing traffic signs with expected functional safety. Background Technology
[0002] With the development of intelligent driving, traffic sign detection, as an important component of the perception module, is directly related to the vehicle's environmental understanding and driving safety.
[0003] Existing traffic sign detection methods mainly include those based on traditional image processing and those based on deep learning. Early methods often employed traditional image processing techniques such as edge detection, color segmentation, and shape matching to identify traffic signs. These methods performed well under conditions of sufficient lighting and uniformity, but were susceptible to background interference and changes in lighting in real-world traffic scenarios, exhibiting poor robustness. In recent years, with the development of deep learning, neural network-based traffic sign detection methods have emerged, such as Faster R-CNN, SSD, and the YOLO series of object detection algorithms. These methods significantly improved detection accuracy and generalization ability compared to traditional methods. Some methods also incorporated attention mechanisms, feature pyramids, and Transformer structures to enhance the detection performance for small and multi-scale targets. However, existing deep learning detection methods still have shortcomings. In detecting small targets in traffic signs, limitations in resolution and feature representation capabilities led to frequent false negatives and missed detections. Detection accuracy decreased significantly under real-world weather and lighting conditions such as rain, fog, and nighttime.
[0004] Therefore, while ensuring detection accuracy, it is still of great significance to improve robustness to small-sized traffic signs and real-world scenarios, and to take into account real-time performance. Summary of the Invention
[0005] To address the shortcomings of existing technologies, in a first aspect, the present invention provides a method for recognizing traffic signs with intended functional safety, comprising the following steps:
[0006] Acquire traffic sign images and perform preprocessing to obtain preprocessed traffic sign images;
[0007] A traffic sign detection model based on an improved RT-DETR is constructed, comprising a backbone network, a neck network, and an RT-DETR decoder. The backbone network is used for feature extraction, including an HGStem module, multiple HGBlock-space modules, and multiple DWConv deep convolutional modules. The HGStem module extracts features from traffic sign images to obtain feature maps. The HGBlock-space module extracts and fuses features from the input feature maps through different paths. The neck network is used for feature fusion, including a feature modeling module, multiple upsampling layers, multiple spatial fusion modules, multiple multi-scale feature compression modules, and multiple Conv convolutional sub-modules. The feature modeling module performs deep correlation modeling on the feature maps output by the backbone network, and the spatial fusion module performs spatial fusion on two input feature maps. The multi-scale feature compression module performs pooling, weighted fusion, and concatenation on the input feature maps. The RT-DETR decoder is used to detect traffic signs based on the fused feature maps output by the neck network.
[0008] The preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained.
[0009] Furthermore, the HGStem module includes multiple Conv convolutional sub-modules and a max pooling layer. The Conv convolutional sub-modules include 2D convolution, a normalization layer, and a SiLU activation function.
[0010] For the image input to the HGStem module The HGStem module performs convolutional processing using a single Conv convolutional submodule, followed by two branches. The first branch includes two Conv convolutional submodules, and the second branch includes a max-pooling layer. The results from the first and second branches are concatenated and then passed through another Conv convolutional module to obtain the feature map output by the HGStem module. As shown in the formula below:
[0011]
[0012] in, For the Conv convolutional submodule, For max pooling layer, This is for splicing operations.
[0013] Furthermore, the HGBlock-space module includes multiple Conv convolutional sub-modules;
[0014] For the feature map of the input HGBlock-space module , feature map Copy as feature map Multiple Conv convolutional submodules are used to process the feature map. Feature extraction is performed by convolving the outputs of all preceding Conv convolutional submodules with the feature map. It is concatenated with the feature maps output from multiple Conv convolutional submodules, and then combined with the feature maps. Aggregate the features to obtain the feature map output by the HGBlock-space module. As shown in the formula below:
[0015] .
[0016] Furthermore, the spatial fusion module performs spatial fusion on the two input feature maps as follows:
[0017] The spatial fusion module will input two feature maps and spliced as Two branches are used to concatenate the feature maps. The processing involves a 1×1 convolution in the first branch, which yields the first branch feature map. ;
[0018] The second branch includes 1×1 convolutions, 3×3 convolutions, pooling layers, a KAN network, and a softmax layer. The pooling layers include average pooling and max pooling branches, which sequentially use 1×1 and 3×3 convolutions to concatenate the feature maps. The intermediate feature map is obtained through processing. Pooling layers are used to process the intermediate feature maps respectively. Perform max pooling and average pooling operations, then concatenate the results of max pooling and average pooling to obtain the feature map output by the pooling layer. The feature maps output from the pooling layer are sequentially input into the KAN network and the softmax layer to obtain the feature maps of the second branch. ;
[0019] The first branch feature map Intermediate feature map Second branch feature map By splicing the features together, we obtain the feature map output by the spatial fusion module. .
[0020] Furthermore, the multi-scale feature compression module includes an average pooling layer, a max pooling layer, a 1×1 convolution, a 3×3 convolution, a KAN network, and a softmax layer;
[0021] Feature maps of the input multi-scale feature compression module Perform max pooling and average pooling operations separately, and concatenate the results of max pooling and average pooling to obtain the first intermediate feature map. The first intermediate feature map was processed using 1×1 convolution and 3×3 convolution respectively. Convolution and concatenation are performed to obtain the second intermediate feature map. ;
[0022] The second intermediate feature map The intermediate feature maps at different scales are sequentially input into the KAN network and the Softmax layer, and then weighted and fused to obtain the third intermediate feature map. ;
[0023] The feature map input to the multi-scale feature compression module and the third intermediate feature map The feature maps are concatenated to obtain the feature maps output by the multi-scale feature compression module. .
[0024] Furthermore, the preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained, including the following steps:
[0025] The preprocessed traffic sign image is input into the backbone network for feature extraction, and the multi-scale feature map output by the backbone network is obtained.
[0026] The multi-scale feature map output by the backbone network is input into the neck network for feature fusion, resulting in the fused feature map output by the neck network.
[0027] The fused feature map output from the neck network is input into the RT-DETR decoder to obtain the traffic sign detection results.
[0028] Secondly, this application proposes an electronic device including one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the intended function of the traffic sign recognition method.
[0029] Thirdly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the intended function of the traffic sign recognition method.
[0030] Fourthly, this application proposes a computer program product, including a computer program or instructions that, when executed by a processor, implement the intended function of the safe traffic sign recognition method.
[0031] The beneficial effects of adopting the above technical solution are as follows: The traffic sign recognition method for expected functional safety provided by this invention, by introducing a spatial fusion module, can fully integrate contextual information from different spatial regions, improving recognition robustness in complex backgrounds. By designing a multi-scale feature compression module, features at different scales are efficiently compressed and represented while maintaining the integrity of key information, improving the detection accuracy for small-sized traffic signs. The HGBlock-space module can effectively enhance the overall perception capability of traffic sign targets and possesses good anti-interference capabilities. In summary, this invention, while ensuring real-time performance, can significantly improve the accuracy and robustness of traffic sign detection in real-world scenarios, especially under conditions of small size and varying lighting, where its detection performance is superior, making it suitable for the practical application needs of vehicle-mounted terminals. Attached Figure Description
[0032] Figure 1 A schematic diagram of the traffic sign detection model structure based on the improved RT-DETR provided in this embodiment of the invention;
[0033] Figure 2 A schematic diagram of the HGStem module structure provided in this embodiment of the invention;
[0034] Figure 3 A schematic diagram of the Conv convolutional submodule structure provided in this embodiment of the invention;
[0035] Figure 4 A schematic diagram of the HGBlock-space module structure provided in this embodiment of the invention;
[0036] Figure 5 A schematic diagram of the spatial fusion module structure provided in this embodiment of the invention;
[0037] Figure 6 A schematic diagram of the multi-scale feature compression module structure provided in this embodiment of the invention. Detailed Implementation
[0038] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.
[0039] Example 1:
[0040] The intended functional safety of automobiles requires vehicles to accurately detect traffic signs. Existing traffic sign detection methods suffer from insufficient accuracy and poor robustness in real-world road environments. Specifically, traffic signs are typically small, making existing models prone to missed detections and false detections when detecting small targets. Furthermore, traffic signs are diverse, and their features are often indistinct at long distances or in low-resolution conditions, making accurate identification difficult. Traditional detection methods relying on handcrafted features such as color, shape, and edges have limited adaptability and cannot meet the detection needs of real-world scenarios.
[0041] This embodiment proposes a method for recognizing traffic signs with expected functional safety, which can effectively improve the detection accuracy and robustness of small-sized, multi-category traffic signs in real road environments. The method includes the following steps:
[0042] Acquire traffic sign images and perform preprocessing to obtain preprocessed traffic sign images;
[0043] Construct a traffic sign detection model based on an improved RT-DETR, such as... Figure 1 As shown, it includes the backbone network, neck network, and RT-DETR decoder;
[0044] The preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained.
[0045] The backbone network includes the HGStem module, multiple HGBlock-space modules, and multiple DWConv deep convolutional modules;
[0046] HGStem module such as Figure 2 As shown, it includes multiple Conv convolutional submodules and a max pooling layer, for an image input to the HGStem module. The HGStem module performs convolutional processing using a single Conv convolutional submodule, followed by two branches. The first branch includes two Conv convolutional submodules, and the second branch includes a max-pooling layer. The results from the first and second branches are concatenated and then passed through another Conv convolutional module to obtain the feature map output by the HGStem module. As shown in the formula below:
[0047]
[0048] in, For the Conv convolutional submodule, For max pooling layer, For splicing operations;
[0049] The Conv convolutional submodule includes 2D convolutions, normalized layers, and SiLU activation functions, such as... Figure 3 As shown, the output of the Conv convolutional submodule As shown in the formula below:
[0050]
[0051] in, For the input of the Conv convolutional submodule, For 2D convolution, For normalization layer, The SiLU activation function;
[0052] The structure of the HGBlock-space module is as follows: Figure 4 As shown, it includes multiple Conv convolutional submodules;
[0053] For the feature map of the input HGBlock-space module , feature map Copy as feature map To preserve the original feature maps, multiple Conv convolutional submodules are used to refine the feature maps. Feature extraction is performed by convolving the outputs of all preceding Conv convolutional submodules to extract multi-level, multi-scale features, and then converting the feature map... It is concatenated with the feature maps output from multiple Conv convolutional submodules, and then combined with the feature maps. Aggregate the features to obtain the feature map output by the HGBlock-space module. Feature map The feature map output by the HGBlock-space module simultaneously includes local edge details, global spatial structure, and contextual semantic information. As shown in the formula below:
[0054]
[0055] The HGBlock-space module introduces a multi-convolutional branch structure in the spatial dimension, which enables full fusion of contextual information from different spatial regions of the input feature map, enhancing the overall perception of the target. At the same time, it can effectively alleviate the problem of insufficient spatial dependence of traditional convolutional networks in complex traffic scenarios, and enhance the model's ability to perceive the spatial consistency of traffic signs and targets in different locations.
[0056] The neck network includes a feature modeling module AIFI, multiple upsampling layers, multiple spatial fusion modules MFCompression, multiple multi-scale feature compression modules SFusion, and multiple Conv convolutional sub-modules;
[0057] The AIFI feature modeling module in the RT-DETR model is used to perform deep correlation modeling on the output feature map of the backbone network, thereby enabling information interaction and supplementation across space and channels.
[0058] The spatial fusion module MFCompression is used to fully fuse information from feature maps at different levels, thereby preserving low-level detailed features while maintaining high-level semantic information, improving the recognition accuracy of small targets and long-distance traffic signs. Its structure is as follows: Figure 5 As shown, it includes multiple 1×1 convolutions, 3×3 convolutions, pooling layers, KAN networks, and softmax layers;
[0059] Two feature maps from the input space fusion module MFCompression and spliced as The concatenated feature map is processed using two branches. The first branch includes a 1×1 convolution to achieve channel compression and feature alignment. The feature map of the first branch is obtained after processing by the first branch. ;
[0060] The second branch includes 1×1 convolutions, 3×3 convolutions, pooling layers, a KAN network, and a softmax layer. The pooling layers include average pooling and max pooling branches, which process the intermediate feature maps output from the 1×1 and 3×3 convolutions. Max pooling and average pooling operations are performed separately to capture salient target regions and global background information. The results of max pooling and average pooling are then concatenated to obtain the feature map output by the pooling layer. The feature map output from the pooling layer is input into the KAN network. The KAN network can model the complex spatial dependencies between features by fitting relationships through an adaptive nonlinear function. The softmax layer is used to process the output of the KAN network to obtain the feature map of the second branch. ;
[0061] The first branch feature map Intermediate feature map Second branch feature map By concatenating the features, we obtain the feature map output by the spatial fusion module MFCompression. ;
[0062] The spatial fusion module MFCompression dynamically balances the influence of features at different scales by introducing multi-scale convolution and adaptive weighting mechanisms. Specifically, the KAN network more accurately models the nonlinear relationships between multi-layer features, thereby improving the semantic consistency of the fusion results. Max pooling and average pooling provide multi-view information, preserving salient regions and background context features. MFCompression enables the traffic sign detection model to maintain higher detection stability when dealing with small traffic signs and to maintain high feature discrimination power under real-world conditions such as lighting changes and occlusion, thus improving the overall robustness of the model.
[0063] The multi-scale feature compression module SFusion is used to adaptively compress and weight feature maps of different scales during feature fusion, thereby significantly improving the model's detection performance for small-sized traffic signs while maintaining the integrity of key information. Its structure is as follows: Figure 6 As shown, it includes average pooling layers, max pooling layers, 1×1 convolutions, 3×3 convolutions, a KAN network, and a softmax layer;
[0064] The feature map of the input multi-scale feature compression module SFusion Max pooling and average pooling operations are performed separately. These two types of pooling operations extract the saliency information of the features and the global balance information, respectively. The results of max pooling and average pooling are then concatenated to obtain the first intermediate feature map. The first intermediate feature map was processed using 1×1 convolution and 3×3 convolution respectively. Convolution and concatenation are performed to capture spatial information at different scales, resulting in a second intermediate feature map. ;
[0065] The second intermediate feature map The intermediate feature maps at different scales are sequentially input into the KAN network and the Softmax layer, and weighted fusion is performed to achieve adaptive feature compression and filtering, resulting in a third intermediate feature map. ;
[0066] The feature map input to the multi-scale feature compression module SFusion and the third intermediate feature map The feature maps are concatenated to obtain the feature maps output by the multi-scale feature compression module SFusion. ;
[0067] The multi-scale feature compression module SFusion effectively suppresses redundant components in features while maintaining overall information representation capabilities, enabling the network to focus more on responses in key regions. The multi-scale feature compression mechanism allows the network to simultaneously perceive large-scale background and fine-grained local structures, making it particularly suitable for detecting small-sized traffic signs. By combining various convolution and pooling operations, it effectively integrates saliency information and global balance information. In real-world environments such as varying lighting conditions, the multi-scale feature compression module SFusion improves feature stability through multi-scale fusion, significantly enhancing the model's generalization performance and environmental adaptability.
[0068] The RT-DETR decoder is used for traffic sign detection based on the fused feature map output by the neck network.
[0069] The preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained, including the following steps:
[0070] The preprocessed traffic sign image Xdata is input into the backbone network for feature extraction, resulting in a multi-scale feature map output by the backbone network. The specific method is as follows:
[0071] The HGStem module is used to process the Xdata data to obtain the feature map. ;
[0072] Input the feature map X1 into the first HGBlock-space module to obtain the feature map. ;
[0073] Input the feature map X2 into the first DWConv depthwise convolution module to obtain the feature map. ;
[0074] The feature map X3 is input into the second and third HGBlock-space modules in series to obtain the feature map. ;
[0075] The feature map X4 is input into the second DWConv depthwise convolution module to obtain the feature map. ;
[0076] The feature map X5 is input into the cascaded fourth HGBlock-space module to the eighth HGBlock-space module to obtain the feature map. ;
[0077] The feature map X6 is input into the third DWConv depthwise convolution module to obtain the feature map. ;
[0078] The feature map X7 is fed into the cascaded ninth and tenth HGBlock-space modules to obtain the feature map. ;
[0079] The feature map X4 output by the third HGBlock-space module, the feature map X6 output by the eighth HGBlock-space module, and the feature map X8 output by the tenth HGBlock-space module are selected as the multi-scale feature maps output by the backbone network.
[0080] The multi-scale feature map output from the backbone network is input into the neck network for feature fusion to obtain the fused feature map output from the neck network. The specific method is as follows:
[0081] The neck network is based on a feature pyramid structure, including two paths: bottom-up and top-down. In the bottom-up path of the neck network, the first Conv convolutional submodule processes the feature map X8 to obtain the feature map. ;
[0082] The feature map X9 is input into the feature modeling module AIFI based on the Transformer encoder layer. The feature modeling module AIFI is used to enhance the global modeling capability of image features extracted by the convolutional neural network, and thus obtain the feature map. ;
[0083] The feature map X10 is input into the second Conv convolutional submodule to obtain the feature map. ;
[0084] The feature map X11 is input into the first upsampling layer to obtain the feature map. ;
[0085] The feature map X6 is processed using the third Conv convolutional submodule to obtain the feature map. ;
[0086] Feature map X12 and feature map X13 are input into the first spatial fusion module MFCompression to obtain feature map. ;
[0087] The feature map X14 is input into the first multi-scale feature compression module SFusion to obtain the feature map. ;
[0088] The feature map X15 is input into the fourth Conv convolutional submodule to obtain the feature map. ;
[0089] The feature map X16 is input into the second upsampling layer to obtain the feature map. ;
[0090] The feature map X4 is processed using the fifth Conv convolutional submodule to obtain the feature map. ;
[0091] Feature maps X17 and X18 are input into the second spatial fusion module MFCompression to obtain feature maps. ;
[0092] In the top-down path, the feature map X19 is input into the second multi-scale feature compression module SFusion to obtain the feature map. ;
[0093] The feature map X20 is processed using the fifth Conv convolutional submodule to obtain the feature map. ;
[0094] Feature maps X16 and X21 are input into the third spatial fusion module MFCompression to obtain feature maps. ;
[0095] The feature map X22 is input into the third multi-scale feature compression module SFusion to obtain the feature map. ;
[0096] The feature map X23 is processed using the sixth Conv convolutional submodule to obtain the feature map. ;
[0097] Feature map X11 and feature map X24 are input into the fourth spatial fusion module MFCompression to obtain feature maps. ;
[0098] The feature map X25 is input into the fourth multi-scale feature compression module SFusion to obtain the feature map. ;
[0099] The feature map X20 output by the second multi-scale feature compression module SFusion, the feature map X23 output by the third multi-scale feature compression module SFusion, and the feature map X26 output by the fourth multi-scale feature compression module SFusion are used as the fused feature map output by the neck network.
[0100] The fused feature maps X20, X23, and X26 output from the neck network are input into the RT-DETR decoder to obtain the traffic sign detection results. As shown in the formula below:
[0101]
[0102] This embodiment constructs a traffic sign detection model based on an improved RT-DETR. On the basis of the RT-DETR model, a spatial fusion module MFCompression is introduced to enhance the feature interaction at different spatial levels, thereby improving the representation ability of traffic sign regions. At the same time, combined with the feature compression module SFusion, multi-scale features are effectively aggregated and redundancy suppressed, reducing computational complexity while ensuring detection accuracy. An HGBlock-space module is designed to replace the original HGBlock module to enhance spatial extraction capabilities and improve the accuracy and robustness of traffic sign detection in real-world scenarios.
[0103] Example 2:
[0104] This embodiment proposes an electronic device, including: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to perform the intended functional traffic sign recognition method.
[0105] The electronic device may be a mobile phone, computer, or tablet computer, etc., and includes a memory and a processor. The memory stores a computer program that, when executed by the processor, implements the intended functional safety traffic sign recognition method as described in the embodiments. It is understood that the electronic device may also include an input / output (I / O) interface and communication components.
[0106] The processor is used to execute all or part of the steps in the intended functional safety traffic sign recognition method as described in the above embodiments. The memory is used to store various types of data, which may include, for example, instructions for any application or method in the electronic device, as well as application-related data.
[0107] The processor may be implemented as an Application Specific Integrated Circuit (ASIC), Digital Signal Processor (DSP), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor, or other electronic components, and is used to execute the intended functional safety traffic sign recognition method described in the above embodiments.
[0108] Example 3:
[0109] This embodiment proposes a computer-readable storage medium that stores executable instructions. When these instructions are executed, if they are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
[0110] The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the intended functional safety traffic sign recognition method described in the various embodiments of this application.
[0111] The aforementioned storage media include: flash memory, hard disk, multimedia card, card-type memory (e.g., SD (Secure Digital Memory Card) or DX (Memory Data Register, MDR) memory), random access memory (RAM), static random-access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic storage, disk, optical disk, server, APP (Application) application store, and other media capable of storing program verification codes, on which computer programs are stored. When the computer programs are executed by a processor, they can implement the various steps of the above-described expected function of the safe traffic sign recognition method.
[0112] Example 4:
[0113] This embodiment proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the intended function of a safe traffic sign recognition method.
[0114] Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or part of the technical solution, can be embodied in the form of a computer program product.
[0115] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.
[0116] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of this disclosure and its equivalents, then the intent of this disclosure also includes these modifications and variations.
Claims
1. A method for recognizing traffic signs with intended functional safety, characterized in that, Includes the following steps: Acquire traffic sign images and perform preprocessing to obtain preprocessed traffic sign images; A traffic sign detection model based on an improved RT-DETR is constructed, comprising a backbone network, a neck network, and an RT-DETR decoder. The backbone network is used for feature extraction and includes an HGStem module, multiple HGBlock-space modules, and multiple DWConv deep convolutional modules. The HGStem module extracts features from the traffic sign image to obtain a feature map. The HGBlock-space module extracts and fuses features from the input feature map using different paths. The neck network is used for feature fusion. It includes a feature modeling module, multiple upsampling layers, multiple spatial fusion modules, multiple multi-scale feature compression modules, and multiple Conv convolutional sub-modules; the spatial fusion module is used to spatially fuse two input feature maps; the multi-scale feature compression module performs pooling, weighted fusion, and concatenation on the input feature maps; the RT-DETR decoder is used to detect traffic signs based on the fused feature map output by the neck network; The preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained.
2. The method for recognizing traffic signs with intended functional safety according to claim 1, characterized in that, The HGStem module includes multiple Conv convolutional sub-modules and a max pooling layer. The Conv convolutional sub-modules include 2D convolution, a normalization layer, and a SiLU activation function. For the image input to the HGStem module The HGStem module performs convolutional processing using a single Conv convolutional submodule, followed by two branches. The first branch includes two Conv convolutional submodules, and the second branch includes a max-pooling layer. The results from the first and second branches are concatenated and then passed through another Conv convolutional module to obtain the feature map output by the HGStem module. As shown in the formula below: in, For the Conv convolutional submodule, For max pooling layer, This is for splicing operations.
3. The method for recognizing traffic signs with intended functional safety according to claim 1, characterized in that, The HGBlock-space module includes multiple Conv convolutional sub-modules; For the feature map of the input HGBlock-space module , feature map Copy as feature map Multiple Conv convolutional submodules are used to process the feature map. Feature extraction is performed by convolving the outputs of all preceding Conv convolutional submodules with the feature map. It is concatenated with the feature maps output from multiple Conv convolutional submodules, and then combined with the feature maps. Aggregate the features to obtain the feature map output by the HGBlock-space module. As shown in the formula below: 。 4. The method for recognizing traffic signs with intended functional safety according to claim 1, characterized in that, The spatial fusion module performs spatial fusion on the two input feature maps as follows: The spatial fusion module will take two feature maps as input to the spatial fusion module. and spliced as Two branches are used to concatenate the feature maps. The processing involves a 1×1 convolution in the first branch, which yields the first branch feature map. ; The second branch includes 1×1 convolutions, 3×3 convolutions, pooling layers, a KAN network, and a softmax layer. The pooling layers include average pooling and max pooling branches, which sequentially use 1×1 and 3×3 convolutions to concatenate the feature maps. The intermediate feature map is obtained through processing. Pooling layers are used to process the intermediate feature maps respectively. Perform max pooling and average pooling operations, then concatenate the results of max pooling and average pooling to obtain the feature map output by the pooling layer. The feature maps output from the pooling layer are sequentially input into the KAN network and the softmax layer to obtain the feature maps of the second branch. ; The first branch feature map Intermediate feature map Second branch feature map By splicing the features together, we obtain the feature map output by the spatial fusion module. .
5. The method for recognizing traffic signs with intended functional safety according to claim 1, characterized in that, The multi-scale feature compression module includes an average pooling layer, a max pooling layer, a 1×1 convolution, a 3×3 convolution, a KAN network, and a softmax layer; The feature map of the input multi-scale feature compression module Perform max pooling and average pooling operations separately, and concatenate the results of max pooling and average pooling to obtain the first intermediate feature map. The first intermediate feature map was processed using 1×1 convolution and 3×3 convolution respectively. Convolution and concatenation are performed to obtain the second intermediate feature map. ; The second intermediate feature map The intermediate feature maps at different scales are sequentially input into the KAN network and the Softmax layer, and then weighted and fused to obtain the third intermediate feature map. ; The feature map input to the multi-scale feature compression module and the third intermediate feature map The feature maps are concatenated to obtain the feature maps output by the multi-scale feature compression module. .
6. The method for recognizing traffic signs with intended functional safety according to claim 1, characterized in that, The preprocessed traffic sign image is input into a traffic sign detection model based on improved RT-DETR for detection, and the traffic sign detection results are obtained, including the following steps: The preprocessed traffic sign image is input into the backbone network for feature extraction, and the multi-scale feature map output by the backbone network is obtained. The multi-scale feature map output by the backbone network is input into the neck network for feature fusion, resulting in the fused feature map output by the neck network. The fused feature map output from the neck network is input into the RT-DETR decoder to obtain the traffic sign detection results.
7. An electronic device, characterized in that, It includes one or more processors and a memory for storing instructions that, when executed by the one or more processors, cause the one or more processors to perform the intended functional safety traffic sign recognition method according to any one of claims 1 to 6.
8. A computer-readable storage medium, characterized in that, The system stores executable instructions that, when executed, cause the processor to perform the intended functional safety traffic sign recognition method as described in any one of claims 1 to 6.
9. A computer program product, characterized in that, Includes a computer program or instructions that, when executed by a processor, implement the intended function of the traffic sign recognition method as described in any one of claims 1 to 6.