An image target detection method based on intelligent transportation system

Abstract

The application discloses an image target detection method based on an intelligent traffic system, and specifically comprises the following steps: acquiring an original image data set from the intelligent traffic system, pre-processing the data set to generate a feature map to be processed; constructing a Trifuse module with feature extraction, scale perception and context information capture; embedding the Trifuse module into a YOLOv5s network structure to obtain an image target detection method model, wherein the Trifuse module receives the feature map to be processed, first inputs into a CBR module to generate a processed feature map and generates enhanced local features through a local path; the enhanced local features are divided into three branches: one branch retains original shallow layer information; one branch is transmitted into a Nested-Stream nested stream; one branch is processed through a path based on a hollow convolution; and the processed feature map is input into the image target detection method model to obtain a result data set.

Description

TECHNICAL FIELD

[0001] The present application relates to target recognition technology in intelligent transportation systems, particularly a multi-scale image target detection method applied in the field of public transportation, which can effectively improve the target recognition accuracy based on big data and intelligent technology, especially for small and occluded target detection. BACKGROUND

[0002] With the rapid development of intelligent transportation systems, the demand for target detection technology in the field of public transportation is growing. Especially in complex traffic environments, traditional target detection methods face the problem of insufficient small target recognition accuracy, especially for small vehicles, pedestrians and traffic signs. Real-time data analysis based on big data can effectively optimize the traffic management system and improve the safety and efficiency of public transportation.

[0003] In autonomous driving and intelligent transportation systems, the detection of small targets such as distant pedestrians and vehicles has always been a challenge, and traditional target detection methods are difficult to achieve good detection results in complex backgrounds. Small targets usually occupy a small area in the image, and background information may interfere with the detection results. In addition, due to the dynamic and complex nature of the traffic environment, the scale change and occlusion of the target pose even higher requirements on existing methods.

[0004] In order to solve these problems, in recent years, many multi-scale image target detection methods have emerged. For example, FPN (Feature Pyramid Network) extracts feature maps of different scales through a top-down approach; while PANet (Path Aggregation Network) enhances the use of low-level features by introducing a bottom-up path, thereby improving the accuracy of small target detection. Although these methods have solved the problem of multi-scale feature fusion to some extent, due to their complex structure and large computational overhead, real-time processing capability is still limited.

[0005] The Trifuse module as a new multi-scale image target detection method can effectively fuse information of different scales through local, nested and global paths. Compared with existing multi-scale fusion structures such as FPN and BiFPN, Trifuse introduces an adaptive path selection mechanism, which can significantly improve detection efficiency while maintaining high accuracy. The Trifuse module proposed in the present application not only enhances the detection capability of small targets, but also supports real-time analysis through big data technology, optimizing target recognition in public transportation systems. SUMMARY

[0006] To solve the above problems, the application provides an image target detection method based on an intelligent transportation system, which performs multi-scale feature enhancement through local, nested and global paths of a Trifuse module, significantly improves detection accuracy, especially in small targets and complex occlusion scenes in public transportation environments, and comprises the following steps:

[0007] An original image dataset is acquired from an intelligent transportation system, the dataset is preprocessed, and a feature map to be processed is generated;

[0008] A Trifuse module with feature extraction, scale perception and context information capture is constructed; the Trifuse module is embedded into a YOLOv5s network structure to obtain an image target detection method model, the Trifuse module receives the feature map to be processed, first inputs into a CBR module to generate a processed feature map and generates enhanced local features through a local path; the CBR module is composed of a two-dimensional convolution layer without a bias term, a batch normalization layer and a nonlinear activation function; the enhanced local features are divided into three branches: one branch retains original shallow layer information; one branch is transmitted into a Nested-Stream nested stream; one branch is processed through a path based on a hollow convolution;

[0009] The outputs of all branches are transmitted into a global path for activation, the global path captures long-distance semantic dependence, and global context is re-projected into a spatial domain by using global average pooling and bilinear upsampling; finally, the outputs of all activated paths are connected and projected back to the target channel dimension through a 1×1 convolution;

[0010] The processed feature map is input into the image target detection method model to obtain a result dataset.

[0011] The input feature map is first processed through a standard Conv-BN-ReLU (CBR) module to unify the channel dimension. This mapping not only reduces redundancy, but also aligns the feature space, preparing for subsequent multi-branch processing.

[0012] The outputs of all branches are transmitted into a global path for activation, the global path captures long-distance semantic dependence, and global context is re-projected into a spatial domain by using global average pooling and bilinear upsampling; finally, the outputs of all activated paths are connected and projected back to the target channel dimension through a 1×1 convolution.

[0013] Preferably, the local path adopts a lightweight 1×1 convolution followed by a mixed convolution composed of a 3×3 convolution.

[0014] The processed feature maps are then processed through local paths, which apply lightweight 1×1 convolutions followed by 3×3 convolutions to enhance the spatial context of shallow layers. This design is particularly effective in preserving fine-grained boundary and texture details, which is crucial for small object detection.

[0015] Preferably, the path based on dilated convolution adopts a hybrid parallel-cascaded dilated structure, including three dilated convolutions adjusted by an adaptive serpentine dilation mechanism, assuming dilation rates of 3, 6, and 9, respectively, as expressed in the following formula:

[0016] ;

[0017] This represents the initial input data for the dilated convolution path. , and They represent dimensions as follows: The convolutional kernels have C input channels and C output channels, and the kernel size is 3. , and Let represent the bias parameters added to the convolution operation, and d represent different hole ratios. This represents the output feature map of the first dilated convolutional layer. This represents the output feature map of the second dilated convolutional layer. This represents the output feature map of the third dilated convolutional layer, and BN represents batch normalization. This represents the activation function.

[0018] The specific formula for the adaptive serpentine expansion mechanism is as follows:

[0019] ;

[0020] in, It is the input feature map. It is a convolution operation used to transform the input feature map. Extract spatial information from it. This is an activation function used to add nonlinearity;

[0021] The selection of the void ratio set D is predicted using the following formula:

[0022] ;

[0023] Here, FC is a fully connected layer used to process the convolutional feature maps. Mapped to a set of void rates D, where D contains different sets of void rates;

[0024] ;

[0025] The learning module dynamically selects the hole rate set D based on the image content, and Select represents the selection operation;

[0026] ;

[0027] in, and These are the minimum and maximum values ​​in the void ratio set D, respectively. These two values ​​will be used in the subsequent void ratio selection process to ensure the increasing rule of void ratio.

[0028] ;

[0029] In the first convolutional layer, the hole rate Selected as It is usually the minimum void ratio in set D (e.g., =3);

[0030] ;

[0031] In the second convolutional layer, the void ratio The selection is based on the first layer void ratio and ensure it is greater than At the same time, it is less than the void ratio of the third layer. ;

[0032] ;

[0033] In the third convolutional layer, the void ratio Selected as , which is usually the maximum void ratio in set D.

[0034] Preferably, the specific calculation process for the global path includes:

[0035] A preliminary compressed feature is obtained by linearly transforming the channels using 1×1 convolution, and then adaptive global average pooling is used to extract the global context representation. The specific formula is as follows:

[0036] ;

[0037] in, This represents the weight matrix of a 1×1 convolution. ∈ The input channel is C, and the output channel is C. This indicates that average pooling is performed across the entire spatial dimension. This represents a two-dimensional convolution operation. Let X represent the bias term, and let X represent the input feature map, X∈ B represents the batch size, C represents the channel, H represents the height, and W represents the width. Representing global features, it is the output after a 1×1 convolution and a gap, with dimensions of [missing information]. .

[0038] The adaptive global average pooling extraction of global context representation enables the channel reconstruction branch, which enhances the channel expressiveness through a fully connected layer. The specific formula is as follows:

[0039] ;

[0040] in, Represents the global features of the input. This represents the weight matrix of the fully connected layer. This represents the bias term of the fully connected layer. Indicates reconstruction. This indicates the enhanced channel features. .

[0041] The enhanced channel features are further remapped to the original spatial size H×W through bilinear upsampling, as shown in the following formula:

[0042] ;

[0043] in, This indicates bilinear interpolation upsampling. This indicates the enhanced channel features. This represents the feature map after upsampling.

[0044] The enhanced local features are split into two streams: one is directly fed into the fusion stage as a shortcut connection, while the other is combined with the subsequent multi-scale fusion process, providing complementary spatial cues. Meanwhile, Nested-Stream processes the features via a dilated convolution-based path, consisting of three dilated convolutions adjusted by an adaptive serpentine dilation mechanism, assuming dilation rates of 3, 6, and 9. The dilated convolutions are arranged in parallel and a cascading mechanism is introduced, where the output of the branch with a dilation rate of 3 is fed into the branch with a dilation rate of 6, and the output of the latter is then fed into the branch with a dilation rate of 9. This design improves the interaction of multi-scale features while maintaining spatial consistency.

[0045] The outputs of all branches, including nested dilated features, are then passed to the global path, which captures long-range semantic dependencies to complement the information from the local and dilated paths. This path utilizes global average pooling and bilinear upsampling to reproject the global context back into the spatial domain.

[0046] Finally, the outputs of all activation paths are concatenated and projected back to the target channel dimension via a 1×1 convolution. This compression step allows for the efficient fusion of diverse contextual representations while maintaining computational efficiency. The final representation is then used in subsequent prediction stages.

[0047] Based on the above three-branch structure, a Trifuse module with feature extraction, scale awareness, and capture of contextual information is constructed.

[0048] The Trifuse module uses a three-branch structure to extract features from targets at different scales, improving the accuracy and comprehensiveness of feature extraction. It solves the problem of how to accurately extract features from targets at different scales, thus meeting the needs of multi-scale target detection.

[0049] Specifically, the three branches in the Trifuse module are used to input the received preprocessed dataset into the CBR module and a module in the Local-Stream consisting of 1×1 convolutions followed by 3×3 convolutions, respectively, to enhance shallow spatial context information. Subsequently, the processed feature maps branch into two streams: one flows as a shortcut connection into the final fusion stage, and the other is processed by the Nested-Stream, which consists of three dilated convolutions adjusted by an adaptive serpentine dilation mechanism. This invention employs a hybrid parallel and nested processing mechanism. This design improves the interaction of multi-scale features while maintaining spatial consistency. Finally, it is processed by the Global-Stream to capture long-range semantic dependencies and supplement information from the Local-Stream and Nested-Stream. This path utilizes global average pooling and bilinear upsampling to reproject the global context into the spatial domain.

[0050] The specific processing flow of the preprocessed dataset in the three branches of the Trifuse module ensures the preservation of key features and the extraction of multi-scale features, and solves the problem of specific steps and processes for data processing in the Trifuse module.

[0051] Finally, the Trifuse module is embedded into the YOLOv5s network structure to obtain the optimized target detection method model based on three-branch fusion.

[0052] The above technical solution illustrates the output process of the preprocessed feature map through the three-branch flow, which helps to gradually extract feature details and fuse multi-scale features, solving the problem of gradually extracting and fusing the original multi-scale features in the YOLOv5 network structure.

[0053] According to a second aspect of the present invention, a computer program product is provided, on which one or more computer programs are stored, which, when executed by a computer processor, implement the method described above.

[0054] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects:

[0055] By cleverly designing these three branch flows, the Trifuse module achieves comprehensive utilization of features at different scales. This multi-scale feature extraction mechanism provides the detection model with more comprehensive and accurate information, significantly improving the model's performance in multi-scale object detection tasks. Attached Figure Description

[0056] The accompanying drawings are included to provide a further understanding of the embodiments and are incorporated in and constitute a part of this specification. The drawings illustrate embodiments and, together with the description, serve to explain the principles of the invention. Other embodiments and many anticipated advantages of the embodiments will be readily recognized as they become better understood through reference to the following detailed description. Elements in the drawings are not necessarily to scale. The same reference numerals refer to corresponding similar parts.

[0057] Figure 1 This is a flowchart illustrating an image target detection method based on an intelligent transportation system according to an embodiment of the present invention.

[0058] Figure 2 This is a schematic diagram of the Trifuse module structure of an image target detection method based on an intelligent transportation system according to an embodiment of the present invention.

[0059] Figure 3 This is the Local-Stream of the Trifuse module in an image target detection method based on an intelligent transportation system according to an embodiment of the present invention.

[0060] Figure 4 This is the Nestedl-Stream of the Trifuse module in an image target detection method based on an intelligent transportation system according to an embodiment of the present invention.

[0061] Figure 5 This is the Global-Stream of the Trifuse module in an image target detection method based on an intelligent transportation system according to an embodiment of the present invention.

[0062] Figure 6 This is a performance comparison chart between an image target detection method based on an intelligent transportation system according to an embodiment of the present invention and YOLOv5s, YOLOv8n, YOLOv11s, and YOLOv12s.

[0063] Figure 7 This is a comparison chart showing the performance difference between an image target detection method based on an intelligent transportation system according to an embodiment of the present invention and the original model YOLOv5s.

[0064] Figure 8 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention. Detailed Implementation

[0065] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0066] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0067] Figure 1 This is a flowchart illustrating an image target detection method based on an intelligent transportation system according to an embodiment of the present invention, as shown below. Figure 1 As shown:

[0068] S1. Dataset Acquisition;

[0069] S2. Preprocess the dataset;

[0070] S3. Construct a Trifuse module that enables feature extraction, scale awareness, and capture of contextual information;

[0071] The Trifuse module uses different branch flows to extract features from different targets.

[0072] The Trifuse module includes three branching streams: Local-Stream, Nested-Stream, and Global-Stream. Figure 2 As shown.

[0073] The Trifuse module has three branches that process the received preprocessed dataset separately, including a hybrid convolution consisting of 1×1 and 3×3 layers, three dilated convolutions with dynamically selected dilation rates, and global average pooling and bilinear upsampling. The specific process is as follows:

[0074] The input feature map is first processed through a standard Conv–BN–ReLU (CBR) module, consisting of a 2D convolutional layer (Conv2d) without bias, a batch normalization layer (BatchNorm2d), and a non-linear activation function (ReLU). The convolutional operation uses an auto-padding strategy to maintain spatial invariance, while supporting configurable kernel size, stride, dilation rate, and grouped convolutions to adapt to different task requirements and unify channel dimensions. This mapping not only reduces redundancy but also aligns the feature space, preparing for subsequent multi-branch processing. The processed feature map is then processed through local paths, which apply a lightweight 1×1 convolution followed by a 3×3 convolution, forming a hybrid convolution, such as... Figure 3 As shown, this approach enhances the shallow spatial context. This design is particularly effective at preserving fine-grained boundary and texture details, which is crucial for small object detection. The enhanced local features, combined with the subsequent multi-scale fusion process, provide complementary spatial cues. The processed feature map is then divided into two paths: one serves as a shortcut, directly preserving the original shallow information for subsequent fusion with the backbone features. The fusion stage employs channel-dimensional concatenation, connecting all branch outputs to this shortcut feature along the channel direction, followed by compression and fusion via a 1×1 convolution, unifying the number of channels to the output dimension. This branch retains more fine-grained features and constructs a shallow, high-resolution representation, serving as the information foundation for subsequent fusion modules. This strategy differs from element-wise addition in traditional residual structures, preserving more spatial details unperturbed by convolution, making it particularly suitable for accurate boundary and texture perception in small object detection. The other path is fed into the Nested-Stream for subsequent operations.

[0075] like Figure 1 As shown, simultaneously, another stream is processed via a path based on dilated convolution.

[0076] This dilated convolution path includes three dilated convolutions, with dilation rates of 3, 6, and 9, adjusted by an adaptive serpentine dilation mechanism. Figure 4 As shown, a novel hybrid parallel-cascade void structure is employed:

[0077] ;

[0078] This represents the initial input data for the dilated convolution path. , and They represent dimensions as follows: The convolutional kernels have C input channels and C output channels, and the kernel size is 3. , and Let represent the bias parameters added to the convolution operation, and d represent different hole ratios. This represents the output feature map of the first dilated convolutional layer. This represents the output feature map of the second dilated convolutional layer. This represents the output feature map of the third dilated convolutional layer, and BN represents batch normalization. This represents the activation function.

[0079] The specific formula for the adaptive serpentine expansion mechanism is as follows:

[0080] ;

[0081] in, It is the input feature map. It is a convolution operation used to transform the input feature map. Extract spatial information from it. This is an activation function used to add nonlinearity;

[0082] The selection of the void ratio set D is predicted using the following formula:

[0083] ;

[0084] Here, FC is a fully connected layer used to process the convolutional feature maps. Mapped to a set of void rates D, where D contains different sets of void rates;

[0085] ;

[0086] The learning module dynamically selects the hole rate set D based on the image content, and Select represents the selection operation;

[0087] ;

[0088] in, and These are the minimum and maximum values ​​in the void ratio set D, respectively. These two values ​​will be used in the subsequent void ratio selection process to ensure the increasing rule of void ratio.

[0089] ;

[0090] In the first convolutional layer, the hole rate Selected as It is usually the minimum void ratio in set D (e.g., =3);

[0091] ;

[0092] In the second convolutional layer, the void ratio The selection is based on the first layer void ratio and ensure it is greater than At the same time, it is less than the void ratio of the third layer. ;

[0093] ;

[0094] In the third convolutional layer, the void ratio Selected as , which is usually the maximum void ratio in set D.

[0095] This invention introduces an organic combination of parallel and nested mechanisms in its structure. Specifically, it first designs three parallel dilated convolutional branches on the main path. By introducing an adaptive serpentine dilation mechanism, this mechanism uses a learning module to adaptively select the dilation rate of each convolutional layer based on the spatial information of the input image. Assuming convolutional kernels with dilation rates of 3, 6, and 9 are used, they directly act on the uniformly encoded input feature map, thereby extracting semantic information under different receptive fields. This part retains the efficiency of ASPP-like structures in spatial parallel modeling. At the same time, this invention further introduces nested paths, using the output of the branch with a dilation rate of 3 as the input of the next stage of convolution, thus entering the modules with dilation rates of 6 and 9, realizing the gradual expansion and depth enhancement of semantic information at the hierarchical level. The hybrid structure of this invention, while ensuring the breadth of spatial coverage, constructs a contextual progression link through cross-layer input connections, realizing the explicit semantic transmission between receptive field levels. This approach effectively alleviates the problem of information isolation between paths of different scales in parallel structures, enhances the semantic continuity between scales and the consistency of the global structure, and is particularly suitable for small target detection and complex boundary scenes that require fine-grained semantic parsing. This design improves the interaction of multi-scale features while maintaining spatial consistency.

[0096] It is worth noting that traditional dilated convolutions, due to their sparse sampling characteristics, are prone to local missing spatial sampling points while expanding the receptive field, leading to "Semantic Disconnection" or "Gridding Artifacts." This problem is particularly problematic in small object detection tasks, potentially causing broken object edges and interrupted semantic transmission, severely impacting detection accuracy. Contrary to initial expectations, dilated convolutions do not always translate into direct computational efficiency. While theoretically they reduce the need for additional parameters, implementation can introduce subtle performance bottlenecks. Sparse sampling patterns can lead to increased memory access complexity and potential computational overhead, especially in hardware-constrained environments.

[0097] Cascaded dilated convolutional layers in deep convolutional neural networks (DCNNs) exponentially expand the receptive field with increasing layer count, contrasting with linear networks. These studies show that starting from the first layer, the dilation rate of each feedforward layer doubles until it reaches an upper limit before the repeating pattern. Notably, mesh artifacts have a more significant impact on the model when using cascaded dilated convolutions. This is because in all DCNNs using cascaded dilated convolutions, the dilation rate of consecutively stacked layers shares a common factor of 2.

[0098] To address the aforementioned issues, the Nested-Stream proposed in this invention incorporates an anti-artifact filter mechanism in its design: an anti-artifact filter is added to the dilated convolution of each layer, and the anti-artifact filter specifically includes:

[0099] Using structured attention-weighted convolution as result-guided compensation:

[0100] ;

[0101] in, This represents the activation function. and Let each represent a learnable feature mapping function. This is the transpose of the matrix during computation. The key-value vector used to generate the attention weight matrix, where c represents the feature factor scaling factor;

[0102] Define a nonlocal convolution with a global receptive field to integrate low-frequency background semantics in the entire image and smooth the breaks between different layers. Specifically, smoothing the breaks between different layers involves processing the weights of each convolution kernel to maximize the influence of neighboring pixels, thereby reducing drastic changes in the image. A possible smoothing formula is: ,in, This represents the smoothed feature map output. This represents the pixel values ​​in the input feature map. The weights of the smoothing filter are usually smoothing functions such as mean or Gaussian distribution, K represents the length of the filter, and t represents the current spatial or temporal position to be calculated in the convolution operation.

[0103] Finally, upsampling is used to restore the low frequencies and compensate for edge blurring and void skipping areas. The upsampling formula is: Up(Conv3) 3,( )), where Conv3 3 is the standard 3 3. Convolution, extracting local information, where s represents the stride. The downsampling convolution with a stride of 2 is used to extract low-frequency information, while the upsampling operation restores the low-frequency information to its original size to fill in the jump points caused by holes. The anti-artifact filter does not require additional training parameters, and its weight calculations are regular. Memory access and computational overhead are optimized during large-scale computations, especially when combined with hardware acceleration and convolution optimization techniques. Simultaneously, to further alleviate the weakness of large-hole convolution paths in representing high-frequency information such as texture and boundaries, this module introduces Local-Stream, which compensates for shallow spatial context through a hybrid convolution of conventional 1×1 and 3×3 layers. This branch, combined with nested paths, integrates features through a multi-branch fusion structure, ultimately achieving a unified output with a 1×1 compressed convolution. This effectively achieves complementary enhancement under different spatial receptive fields, improving the overall modeling ability for small targets and detailed structures, and significantly suppressing the sampling blind spot problem caused by dilated convolutions.

[0104] Furthermore, such as Figure 1 As shown, the output of all branches, including nested hollow features, is then passed to the global path, as follows: Figure 5 As shown, the global path captures long-range semantic dependencies to supplement the information of local and nested paths. This path utilizes global average pooling and bilinear upsampling to reproject the global context back into the spatial domain. The specific computational process is shown below:

[0105] Let the input feature map be: X∈ This module first performs a linear transformation on the channels using a 1×1 convolution to obtain preliminary compressed features, and then applies adaptive global average pooling to extract global context representations.

[0106] ;

[0107] in, This represents the weight matrix of a 1×1 convolution. ∈ The input channel is C, and the output channel is C. This indicates that average pooling is performed across the entire spatial dimension. This represents a two-dimensional convolution operation. Let X represent the bias term, and let X represent the input feature map, X∈ B represents the batch size, C represents the channel, H represents the height, and W represents the width. Representing global features, it is the output after a 1×1 convolution and a gap, with dimensions of [missing information]. .

[0108] Furthermore, enabling the channel reconstruction branch (use_fc=True) further enhances the channel expressiveness through a fully connected layer:

[0109] ;

[0110] in, Represents the global features of the input. This represents the weight matrix of the fully connected layer. This represents the bias term of the fully connected layer. Indicates reconstruction. This indicates the enhanced channel features. .

[0111] Express the obtained channels Remodeling And remapped back to the original spatial size H×W through bilinear upsampling:

[0112] ;

[0113] in, This indicates bilinear interpolation upsampling. This indicates the enhanced channel features. This represents the feature map after upsampling.

[0114] like Figure 1 As shown, finally, the outputs of all activation paths are concatenated and projected back to the target channel dimension via a 1×1 convolution. This compression step enables the efficient fusion of diverse contextual representations while maintaining computational efficiency. The final representation is then used in subsequent prediction stages.

[0115] S4. Embed the Trifuse module into the YOLOv5s network structure to obtain the optimized target detection method model based on three-branch fusion.

[0116] S5. Input the preprocessed dataset into an optimized object detection method model based on three-branch fusion to obtain the final dataset.

[0117] Based on one embodiment of the present invention, such as Figure 6 As shown, in order to more comprehensively evaluate the advantages and disadvantages of the proposed YOLOv5s+Trifuse method in complex scenarios compared with YOLOv5s, YOLOv8n, YOLOv11s and YOLOv12s, the Cityscapes dataset will be used for verification to compare the performance differences of these five methods on the dataset.

[0118] The YOLOv5s model achieved an mAP50 of 61.70%, a recall of 54.50%, and a precision of 77.20% on this dataset.

[0119] The YOLOv8n model achieved an mAP50 of 62.00%, a recall of 55.20%, and a precision of 69.70% on this dataset.

[0120] The YOLOv11s model achieved an mAP50 of 63.00%, a recall of 56.60%, and a precision of 74.70% on this dataset.

[0121] The YOLOv12s model achieved an mAP50 of 60.60%, a recall of 53.10%, and a precision of 76.40% on this dataset.

[0122] The YOLOv5s+Trifuse model achieved an mAP50 of 63.20% on this dataset, a 1.5% improvement over YOLOv5s; a recall of 55.90%, a 1.4% improvement over YOLOv5s; and a precision of 79.20%, a 2% improvement over YOLOv5s.

[0123] The performance changes between YOLOv5s and YOLOv8, and between YOLOv11s and YOLOv12s are relatively small. However, the YOLOv5s+Trifuse model achieves higher accuracy.

[0124] Compared to other models, YOLOv5s+Trifuse performs better.

[0125] like Figure 6 As shown, experimental results demonstrate that combining the YOLOv5s model with the Trifuse-TFL module significantly improves performance across multiple metrics, particularly in mean AP50 and precision at an intersection-over-union ratio of 0.5. Specifically, when the Trifuse module is introduced, the model achieves 63.20% mAP50, 55.90% recall, and 79.20% precision, representing improvements of 1.5%, 1.4%, and 2%, respectively, compared to using YOLOv5s alone.

[0126] This result highlights the significant role of the Trifuse module in optimizing model performance, particularly its efficiency in handling multi-scale object detection problems.

[0127] YOLOv5s-Trifuse also demonstrated high performance in recall rate, fully demonstrating the overall advantages of the Trifuse module.

[0128] More importantly, in terms of accuracy, YOLOv5s-Trifuse achieved 79.20%, significantly outperforming other configurations. This result not only verifies the effectiveness of the Trifuse module in improving detection accuracy but also reflects its potential in reducing false positive rates and optimizing detection results.

[0129] like Figure 7As shown, in order to intuitively compare the detection performance differences between the multi-scale object detection model YOLOv5s-Trifuse and the original model YOLOv5s, this invention performs a visualization analysis on the Cityscapes dataset.

[0130] The image above shows the original YOLOv5s model, while the image below shows the multi-scale detection model YOLOv5s-Trifuse proposed in this invention. By comparing the multi-scale detection results of the two images, this invention observes that the YOLOv5s-Trifuse model significantly outperforms the original YOLOv5s model on the Cityscapes dataset.

[0131] Figure 7 In the image below, the YOLOv5s-Trifuse model successfully detected a small, distant target that was not identified in the image above.

[0132] Furthermore, the detection accuracy of the image below is relatively improved. This invention, by inserting the Trifuse module into the network structure of the YOLOv5s model, endows the model with stronger adaptability and richer contextual information, especially significantly improving the detection performance of small-sized targets and occluded targets; thus enabling the model to perform excellently in target detection tasks.

[0133] In summary, by comparing the results of the two images, the significant advantages of the YOLOv5s-Trifuse model in multi-scale object detection are clearly demonstrated.

[0134] The following is for reference. Figure 8 It shows a schematic diagram of the structure of a computer system 800 suitable for implementing an electronic device according to embodiments of the present invention. Figure 8 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of the present invention.

[0135] like Figure 8 As shown, the computer system 800 includes a central processing unit (CPU) 801, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 802 or programs loaded from storage section 808 into random access memory (RAM) 803. The RAM 803 also stores various programs and data required for the operation of the system 800. The CPU 801, ROM 802, and RAM 803 are interconnected via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0136] The following components are connected to I / O interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a liquid crystal display (LCD) and speakers, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card such as a LAN card and a modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to I / O interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 810 as needed so that computer programs read from it can be installed into storage section 808 as needed.

[0137] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable storage medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by central processing unit (CPU) 801, it performs the functions defined in the methods of this invention. It should be noted that the computer-readable storage medium of this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can also be any computer-readable storage medium other than a computer-readable storage medium that can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. Program code contained on a computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0138] Computer program code for performing the operations of this invention can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Python, and C++—as well as conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0139] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0140] The modules described in the embodiments of the present invention can be implemented in software or in hardware.

[0141] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable storage medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: acquire an original image dataset from an intelligent transportation system; preprocess the dataset to generate a feature map to be processed; input the feature map to be processed into a CBR module to generate a processed feature map; construct a Trifuse module with feature extraction, scale awareness, and capture of contextual information; embed the Trifuse module into a YOLOv5s network structure to obtain an image object detection method model; and input the processed feature map into the image object detection method model to obtain a result dataset.

[0142] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention is not limited to the specific combination of the above-described technical features, but also includes other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in this invention.

Claims

1. An image target detection method based on intelligent transportation systems, characterized in that, include: The original image dataset is obtained from the intelligent transportation system, and the dataset is preprocessed to generate feature maps to be processed. A Trifuse module with feature extraction, scale awareness, and context capture capabilities is constructed. The Trifuse module is embedded into a YOLOv5s network structure to obtain an image object detection method model. The Trifuse module receives the feature map to be processed, first inputting it into a CBR module to generate a processed feature map, and then generating enhanced local features through local paths. The CBR module consists of a two-dimensional convolutional layer without bias, a batch normalization layer, and a non-linear activation function. The enhanced local features are divided into three branches: one branch retains the original shallow information. One stream is fed into a nested stream; the other is processed through a path based on dilated convolutions. This path uses a hybrid parallel-cascaded dilated structure, including three dilated convolutions. The dilation rate is selected by an adaptive serpentine dilation mechanism, with dilation rates of 3, 6, and 9 expressed by the following formulas: ； This represents the initial input data for the dilated convolution path. , and They represent dimensions as follows: The convolutional kernels have C input channels and C output channels, and the kernel size is 3. , and Let represent the bias parameters added to the convolution operation, and d represent different hole ratios. This represents the output feature map of the first dilated convolutional layer. This represents the output feature map of the second dilated convolutional layer. This represents the output feature map of the third dilated convolutional layer, and BN represents batch normalization. Indicates the activation function; The outputs of all branches are fed into the global path for activation. The global path captures long-range semantic dependencies and uses global average pooling and bilinear upsampling to reproject the global context back into the spatial domain. Finally, the outputs of all activated paths are concatenated and projected back to the target channel dimension through a 1×1 convolution. The processed feature map is input into the image object detection method model to obtain the result dataset.

2. The image target detection method according to claim 1, characterized in that, The local path uses a lightweight 1×1 convolution followed by a hybrid convolution consisting of 3×3 convolutions.

3. The image target detection method according to claim 1, characterized in that, The specific formula for the adaptive serpentine expansion mechanism is as follows: ； in, It is the input feature map. It is a convolution operation used to transform the input feature map. Extract spatial information from it. This is an activation function used to add non-linearity; The selection of the void ratio set D is predicted using the following formula: ； Here, FC is a fully connected layer used to process the convolutional feature maps. Mapped to a set of void rates D, where D contains different sets of void rates; ； The learning module dynamically selects the hole rate set D based on the image content, and Select represents the selection operation; ； in, and These are the minimum and maximum values ​​in the void ratio set D, respectively; ； In the first convolutional layer, the hole rate Selected as ; ； In the second convolutional layer, the void ratio The selection is based on the first layer void ratio and guarantee Greater than At the same time, it is less than the void ratio of the third layer. ; ； In the third convolutional layer, the void ratio Selected as , where is the maximum void ratio in set D.

4. The image target detection method according to claim 1, characterized in that, The specific calculation process for the global path includes: Preliminary compressed features are obtained by linearly transforming the channels using 1×1 convolution, and global contextual representation is extracted using adaptive global average pooling. The specific formula is as follows: ； in, This represents the weight matrix of a 1×1 convolution. ∈ The input channel is C, and the output channel is C. This indicates that average pooling is performed across the entire spatial dimension. This represents a two-dimensional convolution operation. Let X represent the bias term, and let X represent the input feature map, X∈ B represents the batch size, C represents the channel, H represents the height, and W represents the width. Representing global features, it is the output after a 1×1 convolution and a gap, with dimensions of [missing information]. .

5. The image target detection method according to claim 4, characterized in that, The adaptive global average pooling extraction of global context representation enables the channel reconstruction branch, which enhances the channel expressiveness through a fully connected layer. The specific formula is as follows: ； in, Represents the global features of the input. This represents the weight matrix of the fully connected layer. This represents the bias term of the fully connected layer. Indicates reconstruction. This indicates the enhanced channel features. .

6. The image target detection method according to claim 5, characterized in that, The enhanced channel features are remapped back to the original spatial size H×W through bilinear upsampling, as expressed by the following formula: ； in, This indicates bilinear interpolation upsampling. This indicates the enhanced channel features. This represents the feature map after upsampling.

7. A computer program product, characterized in that, It stores a computer program that, when executed by a processor, implements the method as described in any one of claims 1-6.

Images