Efficient real-time vehicle intelligent detection and identification method and system based on YOLO-BOS
By adopting the vehicle intelligent detection and recognition method based on YOLO-BOS in intelligent driving, using the BRSA attention mechanism and full-dimensional dynamic convolution, combined with the Shape IOU loss function, the error detection rate and missed detection rate problems of vehicle detection in intelligent driving are solved, and efficient and accurate vehicle identification is achieved.
Patent Information
- Application Number
- CN202411712473.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-27
- Publication Date
- 2025-03-14
AI Technical Summary
In the complex scenarios of intelligent driving, the existing technology faces high error detection rates and missed detection rates in non-ideal detection environments such as long-distance target recognition, mutual shading between vehicles, and severe weather conditions, and it is difficult to effectively balance real-time processing capabilities and high-precision detection performance.
The efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS is adopted to acquire and process road monitoring scene video data, build a YOLO-BOS model, use BRSA attention mechanism and full-dimensional dynamic convolution to enhance feature extraction and perception capabilities, and use Shape IOU loss function to optimize bounding box prediction.
It improves the accuracy and efficiency of vehicle detection and recognition, and can effectively identify vehicles in complex road scenarios, and is suitable for vehicle detection and recognition in complex backgrounds.
Smart Images
Figure CN119625688A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent driving technology, and in particular to an efficient real-time vehicle intelligent detection and recognition method and system based on YOLO-BOS. Background Art
[0002] In the complex scenarios of intelligent driving, vehicle detection is the basis for ensuring driving safety, achieving autonomous planning and decision-making, and precise control. However, in the actual application of intelligent driving, in non-ideal detection environments such as long-distance target recognition, mutual occlusion between vehicles, and severe weather conditions, existing technologies still face challenges of high false detection and missed detection rates. Therefore, how to effectively balance the real-time processing capability and high-precision detection performance of vehicle detection methods in intelligent driving has become a key technical problem to promote the further maturity and widespread application of intelligent driving technology. Summary of the invention
[0003] To this end, the present invention provides an efficient real-time vehicle intelligent detection and recognition method and system based on YOLO-BOS to solve the problem of unsatisfactory vehicle detection accuracy in existing intelligent driving.
[0004] According to the design scheme provided by the present invention, on the one hand, an efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS is provided, comprising:
[0005] Acquire road monitoring scene video data, and process the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle bounding box information;
[0006] A YOLO-BOS model is constructed, and the YOLO-BOS model is trained using a vehicle sample image to obtain a vehicle detection and recognition model, wherein the YOLO-BOS model comprises: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features, and the backbone network uses a double-layer routing in a BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion;
[0007] The image data to be detected is input into the vehicle detection and recognition model, and the vehicle detection and recognition model is used to extract the category and bounding box information of the target vehicle in the image data to be detected.
[0008] As an efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS of the present invention, further, the video data is processed, including:
[0009] Extract video images at specified frame intervals and annotate the vehicle category and bounding box information in the images;
[0010] The video image is divided into vehicle training sample image data and vehicle verification sample image data according to a preset ratio, so that the model is trained using the vehicle training sample image data, and the trained model is tested and optimized using the vehicle verification sample image data, and the YOLO-BOS model after test and optimization is used as the vehicle detection and recognition model.
[0011] As an efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS of the present invention, further, the backbone network includes: a convolutional layer, a plurality of C2F components stacked and used for feature conversion, and a fast spatial pyramid pooling layer SPPF for capturing supplementary scale image features, and the BRSA attention mechanism is arranged between the terminal C2F component and the fast spatial pyramid pooling layer SPPF.
[0012] As an efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS of the present invention, further, the backbone network uses the double-layer routing in the BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, including:
[0013] For the extracted image features, the corresponding query vector, key vector and value vector are obtained through linear mapping;
[0014] The topK routing attention mechanism is used to obtain the routing index matrix. The routing index matrix is used to perform tensor aggregation on the key vector and the value vector to obtain the corresponding intermediate vector. The image feature vector is obtained by using fine-grained token-to-token attention on the intermediate vector and the query vector. The image feature vector is then subjected to maximum pooling and average pooling operations to obtain the output image features.
[0015] As an efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS of the present invention, further, the head network includes four C2F_2 components for feature conversion fusion that are superimposed, wherein the first C2F_2 component includes a convolution layer, an upsampling layer, a splicing layer and a feature conversion C2F-2 layer, the second C2F_2 component includes a full-dimensional dynamic convolution layer, an upsampling layer, a splicing layer and a feature conversion C2F_2 layer, the third C2F_2 component includes a convolution layer, a splicing layer and a feature conversion C2F_2 layer, and the fourth C2F_2 component includes a full-dimensional dynamic convolution layer, a splicing layer and a feature conversion C2F_2 layer.
[0016] As an efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS of the present invention, further, the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion, including:
[0017] After the input image features are processed by the GAP global average pooling layer, FC and ReLU, they are sent to four parallel branch dimensional routes to obtain attention in four dimensions. The four dimensions include: convolution kernel dimension, spatial dimension, input channel dimension and output channel dimension. Each parallel branch dimensional route consists of a fully connected layer and a sigmoid function.
[0018] The four dimensions of attention are multiplied by the convolution kernel one by one, and the convolution kernel shape and weight are adaptively adjusted according to the input feature content to obtain the image features output by the head network.
[0019] As an efficient real-time vehicle intelligent detection and recognition method based on YOLO-BOS of the present invention, further, Shape IOU is used as the target loss function in the training process of the YOLO-BOS model, so as to guide the model to perform training optimization by utilizing the ratio of the overlapping area between the real box and the predicted box and the total area after the real box and the predicted box are merged during the training process.
[0020] On the other hand, the present invention also provides an efficient real-time vehicle intelligent detection and recognition system based on YOLO-BOS, comprising: a sample construction module, a model training module and a vehicle detection module, wherein:
[0021] A sample construction module is used to obtain road monitoring scene video data and process the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle bounding box information;
[0022] A model training module is used to build a YOLO-BOS model and train the YOLO-BOS model using a vehicle sample image to obtain a vehicle detection and recognition model. The YOLO-BOS model includes: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features. The backbone network uses a double-layer routing in a BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion.
[0023] The vehicle detection module is used to input the image data to be detected into the vehicle detection and recognition model, and use the vehicle detection and recognition model to extract the category and boundary box information of the target vehicle in the image data to be detected.
[0024] Beneficial effects of the present invention:
[0025] The present invention adds the BRSA attention mechanism to the backbone network, and can adaptively remove irrelevant areas to improve the efficiency of feature extraction by enhancing the ability to capture and integrate spatial feature information, and increase spatial feature information for more comprehensive feature extraction; embed the full-dimensional dynamic convolution ODConv to reconstruct the head network, and dynamically adjust the size and weight of the convolution kernel in four dimensions of the kernel space to better adapt to the complex and changeable situations of road targets; use the ShapeIOU loss function to more carefully characterize the geometric similarity between the prediction box and the target box, which can more accurately capture the size change and shape distortion of the bounding box, and improve the accuracy of the target detection task. The scheme of this case is further experimented with two data sets, and the experimental results show that the scheme of this case can improve the accuracy of vehicle detection and recognition, and can be applied to vehicle detection and recognition under illumination changes and complex background road scenes. BRIEF DESCRIPTION OF THE DRAWINGS
[0026] Figure 1 This is a schematic diagram of the efficient real-time vehicle intelligent detection and recognition process based on YOLO-BOS in the embodiment;
[0027] Figure 2 The structure of YOLO-BOS in the embodiment is shown;
[0028] Figure 3 The working principle of BRSA in the embodiment is shown;
[0029] Figure 4 FIG. 4 is a schematic diagram of the ODConv structure in the embodiment. DETAILED DESCRIPTION
[0030] In order to make the purpose, technical solutions and advantages of the present invention clearer and more understandable, the present invention is further described in detail below in conjunction with the accompanying drawings and technical solutions.
[0031] In the technical field of intelligent driving systems, accurate recognition of vehicle targets in road scenes is one of its core technologies. However, in complex environments, vehicle detection faces huge challenges due to changes in lighting, weather conditions, road conditions, and vehicle types of various colors, shapes, and sizes. To this end, an embodiment of the present invention provides an efficient and real-time vehicle intelligent detection and recognition method based on YOLO-BOS, comprising:
[0032] S101, obtaining road monitoring scene video data, and processing the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle boundary box information.
[0033] Specifically, the video data processing may be designed to include:
[0034] Extract video images at specified frame intervals and annotate the vehicle category and bounding box information in the images;
[0035] The video image is divided into vehicle training sample image data and vehicle verification sample image data according to a preset ratio, so that the model is trained using the vehicle training sample image data, and the trained model is tested and optimized using the vehicle verification sample image data, and the YOLO-BOS model after test and optimization is used as the vehicle detection and recognition model.
[0036] like Figure 1 In the process scheme shown, the dataset used is the UA-DETRAC dataset, which is derived from the monitoring scenes of 24 road overpasses in Beijing and Tianjin, China, covering more than 10 hours of video recording and containing the outer frame annotation information of more than 1.21 million target objects. The video data is recorded at a rate of 25 frames per second, with a resolution of 960×540 pixels, and each image presents multiple vehicles of different types, with varying degrees of occlusion and truncation. The dataset is specifically subdivided into four label categories, namely cars, buses, vans, and other categories. Regarding the annotation information, it records in detail the category of each target in the image and the precise coordinate values of the bounding box surrounding the target. The specific format is: [category, x_center point, y_center point, width, height]. Among them, "category" is used to identify the type of target; "(x_center point, y_center point)" indicates the coordinate position of the center point of the annotation box; "width" and "height" represent the width and height of the annotation box respectively.
[0037] The dataset contains multiple continuous time series videos, and the differences in vehicles and background changes in adjacent video frames are not significant. If it is used directly for training the target detection model, it may lead to data redundancy and overfitting. In order to effectively improve the generalization ability of the model and shorten the training cycle of the model, this case adopts a specific data processing strategy during the implementation process: extract an image every 10 frames, and divide the extracted images into training sets and verification sets according to the ratio of 8:2, so as to obtain an optimized dataset. Specifically, the processed dataset contains 8639 training images and 2231 verification images.
[0038] S102, constructing a YOLO-BOS model, and training the YOLO-BOS model using a vehicle sample image to obtain a vehicle detection and recognition model, wherein the YOLO-BOS model comprises: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features, and the backbone network uses the double-layer routing in the BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion.
[0039] Among them, Figure 2 As shown, the backbone network includes: a convolutional layer, several C2F components stacked and used for feature conversion, and a fast spatial pyramid pooling layer SPPF for capturing complementary scale image features, and the BRSA attention mechanism is set between the terminal C2F component and the fast spatial pyramid pooling layer SPPF. The head network includes four C2F_2 components stacked and used for feature conversion fusion, wherein the first C2F_2 component includes a convolutional layer, an upsampling layer, a splicing layer, and a feature conversion C2F-2 layer, the second C2F_2 component includes a full-dimensional dynamic convolutional layer, an upsampling layer, a splicing layer, and a feature conversion C2F_2 layer, the third C2F_2 component includes a convolutional layer, a splicing layer, and a feature conversion C2F_2 layer, and the fourth C2F_2 component includes a full-dimensional dynamic convolutional layer, a splicing layer, and a feature conversion C2F_2 layer.
[0040] The double-layer routing strategy in the BRSA attention mechanism is used before the SPPF of the backbone network to fine-tune the weights of each position, aiming to strengthen useful feature information, thereby significantly improving the model's attention and detection effect. Secondly, in the head network part, a full-dimensional dynamic convolution ODConv is embedded after C2F, which can effectively expand the model's receptive field and perception range, enabling it to capture more useful features and provide strong support for accurate vehicle detection. Finally, the Shape-IOU loss function is used to calculate the loss by paying special attention to the shape and size of the bounding box itself, so that the model can be more accurately guided to perform bounding box prediction regression, significantly improving the performance of the model.
[0041] Among them, the BRSA mechanism can selectively screen features according to task requirements and use SAM attention to adjust the importance of spatial position in order to more accurately enhance useful features such as the shape, color, and texture of the target vehicle. Figure 3The working process of BRSA is demonstrated: first, Q, K, and V are obtained through linear mapping, and the top k routing attention mechanism is used on the adjacent matrix Ar to obtain the routing index matrix Ir. K and V are clustered using Ir to obtain Kg and Vg. Token-to-token attention is applied to Q, Kg, and Vg to obtain O. Finally, maximum pooling and average pooling operations are performed to obtain the final extracted features.
[0042] After the C2F of the head network, a full-dimensional dynamic convolution is designed. This convolution can adaptively adjust the convolution parameters according to the input image, and calculate four types of attention along all four dimensions of the kernel space (convolution kernel dimension, spatial dimension, input channel dimension, and output channel dimension) in parallel, assigning different attention weights to each convolution kernel, thereby enhancing the model's ability to obtain input information and capture contextual clues. The schematic diagram of the full-dimensional dynamic convolution is shown in Figure 1. Figure 4 First, the input feature map X is processed by the GAP global average pooling layer to reduce the feature dimension, and then processed by the FC and ReLU modules and sent to four parallel branch routes (composed of a fully connected layer and a sigmoid function) to obtain four attention α wi , α si , α ci , α fi , each attention is multiplied by the convolution kernel one by one, and the shape and weight of the convolution kernel are adaptively adjusted according to the content of the input feature, and finally the final output feature map Y is obtained.
[0043] Using Shape IOU as the target loss function for model training, the loss is calculated by focusing on the shape and size of the bounding box itself, making the bounding box regression more accurate. The calculation method of Shape IOU can be expressed as follows:
[0044]
[0045]
[0046] Among them, scale is the scaling factor, which is related to the size of the target in the dataset, w,h,w gt ,h gt Represents the width and height of the predicted box and the real box respectively, (w gt ) scale and (h gt ) scale They represent w after adjustment using the scaling factor gt and h gt , x c ,y c ,x c gt ,y c gtThey represent the center coordinates of the predicted box and the real box respectively, c represents the diagonal length of the minimum bounding rectangle of the predicted box and the real box, ww and hh represent the weight coefficients in the horizontal and vertical directions respectively, and their values are related to the shape of the real box, and distance shape is the distance loss function, Ω shape is the shape loss function, and the corresponding bounding box regression loss L Shape-IOU as follows:
[0047] L Shape-IOU =1-IOU+distance shape +0.5*Ω shape
[0048] After training is completed, the test set images are input into the YOLO-BOS detection model, and the final detection results are output based on the preset confidence threshold conditions. The results include the predicted target category and the coordinate information of the predicted bounding box.
[0049] S103: input the image data to be detected into a vehicle detection and recognition model, and use the vehicle detection and recognition model to extract the category and bounding box information of the target vehicle in the image data to be detected.
[0050] Furthermore, based on the above method, an embodiment of the present invention also provides an efficient real-time vehicle intelligent detection and recognition system based on YOLO-BOS, comprising: a sample construction module, a model training module and a vehicle detection module, wherein:
[0051] A sample construction module is used to obtain road monitoring scene video data and process the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle bounding box information;
[0052] A model training module is used to build a YOLO-BOS model and train the YOLO-BOS model using a vehicle sample image to obtain a vehicle detection and recognition model. The YOLO-BOS model includes: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features. The backbone network uses a double-layer routing in a BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion.
[0053] The vehicle detection module is used to input the image data to be detected into the vehicle detection and recognition model, and use the vehicle detection and recognition model to extract the category and boundary box information of the target vehicle in the image data to be detected.
[0054] To verify the effectiveness of the scheme, the following is a further explanation based on experimental data:
[0055] The dataset uses the above-mentioned UA-DETRAC dataset. The videos in the dataset are processed, one image is extracted every 10 frames, and divided into training set and validation set in a ratio of 8:2. The Windows 10 operating system is used as the platform, the Python 3.8 environment is used, and the detection algorithm is constructed in combination with the Pytorch deep learning framework. To ensure the comparability of the experimental data, the hyperparameters of the experiment are uniformly set: the image input size is set to 640×640 pixels, the training cycle is set to 100 epochs, and the batch size is set to 32. In addition, SGD (stochastic gradient descent) is selected as the optimizer, the learning rate is set to 0.01, and the number of working threads is set to 8.
[0056] The performance of this solution is compared with other attention mechanisms, loss functions and neural networks, and ablation experiments are carried out on the three innovative points of BRSA, ODC onv and loss function in this solution. The experimental results of the comparison experiment and ablation experiment are shown in Tables 1 to 4.
[0057] Table 1 Performance comparison of BRSA and other attention mechanisms
[0058]
[0059] Table 2 Performance comparison of Shape-IOU and other loss functions
[0060]
[0061] Table 3 Comparison of detection accuracy and model scale of different target detection models
[0062]
[0063] Table 4 Ablation experiments of three innovative points
[0064]
[0065] The above experimental data show that the proposed solution has better detection performance than the existing model, can improve the model detection accuracy and has good generalization ability, can meet the task of detecting targets with variable forms in complex road scenes, and provides a new solution for vehicle identification in intelligent driving systems.
[0066] Unless otherwise specifically stated, the relative steps, numerical expressions and values of the components and steps set forth in these embodiments do not limit the scope of the present invention.
[0067] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.
[0068] The units and method steps of each example described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. A person of ordinary skill in the art may use different methods to implement the described functions for each specific application, but such implementation is not considered to be beyond the scope of the present invention.
[0069] Those skilled in the art will appreciate that all or part of the steps in the above method can be completed by instructing related hardware through a program, and the program can be stored in a computer-readable storage medium, such as a read-only memory, a disk or an optical disk. Optionally, all or part of the steps in the above embodiment can also be implemented using one or more integrated circuits, and accordingly, each module / unit in the above embodiment can be implemented in the form of hardware or in the form of software function modules. The present invention is not limited to any specific form of combination of hardware and software.
[0070] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims
1. An efficient and real-time vehicle intelligent detection and recognition method based on YOLO-BOS, characterized in that: Include: Acquire road monitoring scene video data, and process the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle bounding box information; A YOLO-BOS model is constructed, and the YOLO-BOS model is trained using a vehicle sample image to obtain a vehicle detection and recognition model, wherein the YOLO-BOS model comprises: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features, and the backbone network uses a double-layer routing in a BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion; The image data to be detected is input into the vehicle detection and recognition model, and the vehicle detection and recognition model is used to extract the category and bounding box information of the target vehicle in the image data to be detected.
2. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 is characterized in that: Process the video data, including: Extract video images at specified frame intervals and annotate the vehicle category and bounding box information in the images; The video image is divided into vehicle training sample image data and vehicle verification sample image data according to a preset ratio, so that the model is trained using the vehicle training sample image data, and the trained model is tested and optimized using the vehicle verification sample image data, and the YOLO-BOS model after test and optimization is used as the vehicle detection and recognition model.
3. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 is characterized in that: The backbone network includes: a convolutional layer, a plurality of C2F components stacked and used for feature conversion, and a fast spatial pyramid pooling layer SPPF for capturing complementary scale image features, and a BRSA attention mechanism is arranged between the terminal C2F component and the fast spatial pyramid pooling layer SPPF.
4. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 or 3, characterized in that: The backbone network uses the two-layer routing in the BRSA attention mechanism to adjust the weights of each spatial position in the image to filter and enhance key features, including: For the extracted image features, the corresponding query vector, key vector and value vector are obtained through linear mapping; The topK routing attention mechanism is used to obtain the routing index matrix. The routing index matrix is used to perform tensor aggregation on the key vector and the value vector to obtain the corresponding intermediate vector. The image feature vector is obtained by using fine-grained token-to-token attention on the intermediate vector and the query vector. The image feature vector is then subjected to maximum pooling and average pooling operations to obtain the output image features.
5. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 is characterized in that: The head network includes four superimposed C2F_2 components for feature conversion fusion, wherein the first C2F_2 component includes a convolution layer, an upsampling layer, a splicing layer and a feature conversion C2F-2 layer, the second C2F_2 component includes a full-dimensional dynamic convolution layer, an upsampling layer, a splicing layer and a feature conversion C2F_2 layer, the third C2F_2 component includes a convolution layer, a splicing layer and a feature conversion C2F_2 layer, and the fourth C2F_2 component includes a full-dimensional dynamic convolution layer, a splicing layer and a feature conversion C2F_2 layer.
6. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 or 5, characterized in that: The head network uses full-dimensional dynamic convolution to expand the model's receptive field and perception range to capture relevant features in feature fusion, including: After the input image features are processed by the GAP global average pooling layer, FC and ReLU, they are sent to four parallel branch dimensional routes to obtain attention in four dimensions. The four dimensions include: convolution kernel dimension, spatial dimension, input channel dimension and output channel dimension. Each parallel branch dimensional route consists of a fully connected layer and a sigmoid function. The four dimensions of attention are multiplied by the convolution kernel one by one, and the convolution kernel shape and weight are adaptively adjusted according to the input feature content to obtain the image features output by the head network.
7. The efficient real-time vehicle intelligent detection and identification method based on YOLO-BOS according to claim 1 is characterized in that: During the training process of the YOLO-BOS model, Shape IOU is used as the target loss function. During the training process, the ratio of the overlapping area between the true box and the predicted box and the total area after the true box and the predicted box are used to guide the model to perform training optimization.
8. An efficient real-time vehicle intelligent detection and recognition system based on YOLO-BOS, characterized in that: Contains: sample construction module, model training module and vehicle detection module, among which, A sample construction module is used to obtain road monitoring scene video data and process the video data to obtain vehicle sample images annotated with vehicle category labels and vehicle bounding box information; A model training module is used to build a YOLO-BOS model and train the YOLO-BOS model using a vehicle sample image to obtain a vehicle detection and recognition model. The YOLO-BOS model includes: an input end for preprocessing an input image, a backbone network for extracting features from the preprocessed image, a head network for fusing the extracted image features, and a prediction output end for classifying and predicting the fused output image features. The backbone network uses a double-layer routing in a BRSA attention mechanism to adjust the weights of each spatial position in the image to screen and strengthen key features, and the head network uses full-dimensional dynamic convolution to expand the model receptive field and perception range to capture relevant features in feature fusion. The vehicle detection module is used to input the image data to be detected into the vehicle detection and recognition model, and use the vehicle detection and recognition model to extract the category and boundary box information of the target vehicle in the image data to be detected.
9. An electronic device, characterized in that: include: at least one processor, and a memory coupled to the at least one processor; The memory stores a computer program, and the computer program can be executed by the at least one processor to implement the method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the method according to any one of claims 1 to 7 can be implemented.