Traffic police command action recognition method based on improved YOLOv11

Abstract

The invention belongs to the technical field of image processing and computer vision, and particularly relates to a traffic police command action recognition method based on improved YOLOv11, and the method comprises the specific steps: S1, constructing a traffic police command action recognition data set: collecting images based on Internet public resources and actual traffic scenes, constructing a traffic police gesture recognition data set, and constructing a traffic police command action recognition data set; a standardized annotation sample is generated through image screening, cleaning and enhancement processing, and the annotation format is a YOLO format. According to the method, a diversified traffic police gesture recognition data set is constructed, and multiple image enhancement means such as image atomization, sharpening, Gaussian noise addition, random rotation and scale zooming are adopted, so that the generalization ability and robustness of the model in complex weather, multi-attitude and complex background environments are effectively improved; according to the method, a GPFB module is introduced into a YOLOv11 backbone network, global average pooling and local maximum pooling features are fused, efficient expression of multi-scale features is realized, and the capture ability of the model to traffic police gesture details is enhanced.

Description

Technical Field

[0001] The present invention relates to the technical fields of image processing and computer vision, and in particular to a method for recognizing traffic police command actions based on an improved YOLOv11. Background Art

[0002] With the continuous growth of urban traffic and the rapid development of intelligent transportation technology, accurate recognition of traffic police command actions is crucial for maintaining traffic order and improving traffic efficiency. Traditional traffic police command action recognition relies on manual monitoring or traditional image processing techniques. Limited by factors such as ambient lighting variations, occlusions, and complex and diverse actions, recognition accuracy and real-time performance are insufficient to meet practical application requirements.

[0003] In recent years, deep learning technologies, particularly object detection algorithms based on convolutional neural networks (CNNs), have made significant progress in image recognition. The YOLO family of algorithms, with their efficient single-stage detection architecture, excels in object detection tasks and has become a mainstream choice for action recognition in traffic scenarios. However, the original YOLOv11 model still suffers from insufficient detection accuracy and limited ability to recognize small objects when processing traffic police actions in complex traffic environments.

[0004] Current mainstream target detection algorithms (such as YOLOv5 and YOLOv7) are often affected by factors such as background interference, image blur, and non-standard posture when processing images of traffic police directing actions, resulting in reduced recognition accuracy. Especially in low-resolution or complex scenes, the problems of false detection and missed detection are more serious.

[0005] At the same time, traffic police command action recognition requires improving robustness to action details and environmental interference while maintaining a lightweight model and real-time inference capabilities. Therefore, to address the shortcomings of existing technologies, such as low recognition accuracy, heavy model computational burden, and poor environmental adaptability, we propose a method to improve the YOLOv11 model. This method combines a global-local pooling fusion module, a global contextual attention mechanism, and a lightweight bidirectional feature fusion structure to achieve high-precision and high-performance traffic police command action recognition, meeting the application requirements of intelligent transportation systems. Summary of the Invention

[0006] (1) Technical problems solved

[0007] In response to the shortcomings of the existing technology, the present invention provides a method for traffic police command action recognition based on improved YOLOv11. In response to the problems of slow detection speed, insufficient recognition accuracy and weak adaptability to complex environments in the existing traffic police command action recognition technology, this method introduces a global-local pooling fusion module (GPFB), a Global Context attention mechanism and a lightweight bidirectional feature fusion structure (Lite-BiFPN). Without significantly increasing the number of network parameters and the computational burden, it effectively improves the accuracy and real-time performance of action recognition, significantly reduces the missed detection rate and false detection rate, and enhances the model's robustness to diverse environments (such as different lighting, occlusion, posture changes, etc.), solving the problem of unstable recognition effect of traditional methods in practical applications.

[0008] (2) Technical solution

[0009] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions:

[0010] A traffic police command action recognition method based on improved YOLOv11, the specific steps are as follows:

[0011] S1. Constructing a traffic police command action recognition dataset: Based on images collected from public resources on the Internet and actual traffic scenes, a traffic police gesture recognition dataset is constructed. After image screening, cleaning, and enhancement, standardized annotated samples are generated. The annotation format uses the YOLO format.

[0012] S2. Build a network model: Based on the YOLOv11 network structure, we combine the GPFB (Global-Local Pooling Fusion Block) module, the Global Context attention mechanism, and the Lite-BiFPN multi-scale feature fusion structure to form the improved model YOLO-GGB.

[0013] S3. Training model: Use the enhanced traffic police gesture dataset to train the YOLO-GGB model, dynamically evaluate the model performance through indicators such as mAP, Precision, and Recall, and select the optimal training weights;

[0014] S4. Testing and Validation: After model training is complete, the trained weights are applied to a pre-specified test set. When a test image is input, the model automatically outputs the location of the detected object bounding box and its corresponding category label for each image. To comprehensively evaluate the model's performance in real-world scenarios, the test set covers complex conditions such as multiple lighting, multiple angles, and partial occlusion, enabling systematic analysis and verification of the model's accuracy and robustness.

[0015] Furthermore, the image enhancement processing in S1 includes multiple methods such as image fogging, sharpening, adding Gaussian noise, random rotation and scale scaling to improve the model's generalization ability in multiple weather, multiple posture and complex background scenes.

[0016] Furthermore, the method for constructing the network model in S2 is:

[0017] (1) The GPFB module is embedded in the Backbone network to fuse the global average pooling and local maximum pooling features to enhance the multi-scale feature expression capability. The calculation formula is:

[0018]

[0019] F GPFB =Conv 1×1 (Concat[F gap ,F max ])+X

[0020] Among them, X represents the input feature map, the size is C×H×W; F gap represents the global features obtained by global average pooling, which is used to model the overall context information of the image; F max represents the significant regional response features extracted by local maximum pooling, highlighting the local key areas; Concat[·] represents the concatenation of two feature maps along the channel dimension; Conv1×1(·) represents a 1×1 convolution operation, which is used to fuse the concatenated features and adjust the number of channels; +X represents the residual connection operation, which adds the fused features to the input features to retain the original information and enhance the feature expression capability; F GPFB It is the multi-scale fusion feature map finally output by the module.

[0021] (2) The Global Context Attention Mechanism is introduced in the backend of the backbone. It consists of a context modeling module and a channel weight generation module to extract global context semantic information and improve the model's ability to focus on key areas. The calculation process is as follows:

[0022] Contextual Semantic Modeling:

[0023]

[0024] Channel weight generation:

[0025] W c =σ(MLP(Z c ))

[0026] Weighted output:

[0027] Y c =W c ×Xc

[0028] Among them, X c (i, j) represents the pixel value at the cth channel and position (i, j) of the input feature map; Z c represents the global context vector of channel c; MLP(·) represents the multi-layer perceptron structure, which is used to learn the channel-level weight mapping relationship; σ(·) represents the activation function (usually Sigmoid), which is used to normalize the weight to the [0,1] interval; W c represents the global attention weight corresponding to channel c; Y c Represents the final weighted output feature map.

[0029] (3) The Lite-BiFPN module is introduced in the Neck part, and a top-down and bottom-up weighted bidirectional feature fusion path is adopted. The GhostConv compression module is combined to reduce the computational complexity. The feature fusion calculation formula is:

[0030]

[0031] Among them, P i represents the i-th input feature map (from different levels and scales); w i is the learnable fusion weight corresponding to the input feature map, which is used to control the contribution of features at different scales; ε is a very small constant (such as 1e -4 ), used to avoid the denominator being zero and ensure numerical stability; P out Represents the fused output feature map.

[0032] Furthermore, the GPFB module consists of two parallel branches: one is the global average pooling channel, and the other is the local maximum pooling channel. The output features of the two are concatenated in the channel dimension and then input into a 1×1 convolution fusion, and the original features are retained through the residual connection to enhance the gradient transfer.

[0033] Furthermore, the Global Context Attention Mechanism includes a context modeling module and a channel weight generation module. The former is used to obtain spatial semantic information, and the latter uses a multi-layer perceptron (MLP) to dynamically generate channel weight coefficients.

[0034] Furthermore, the Lite-BiFPN module is based on a weighted bidirectional feature fusion design, supports cross-scale feature alignment and fusion, and reduces the number of parameters and computational complexity through the GhostConv compression module.

[0035] Furthermore, during the S3 training phase, a stochastic gradient descent (SGD) optimizer was used, with an initial learning rate set to 0.01 and 300 training rounds. A cosine annealing learning rate decay strategy was introduced during training to improve late-stage model stability, and a label smoothing mechanism was incorporated to mitigate overfitting risks and enhance generalization capabilities. During training, the training effect was dynamically estimated using metrics such as accuracy, recall, and inference speed (FPS), and training parameters were adjusted appropriately to achieve optimal performance.

[0036] Furthermore, during the S4 testing phase, the trained model was used to conduct a comprehensive performance evaluation on a standard traffic police gesture test set, using metrics including mAP@0.5, precision, recall, and inference speed (FPS). Furthermore, through image visualization analysis of the test results, combined with complex scenarios such as multi-lighting, multi-angle, and occlusion, the model's robustness and adaptability were thoroughly analyzed. The performance metrics were calculated as follows:

[0037]

[0038] Among them, m is the total number of traffic police command action categories, TP is the number of positive samples correctly identified, FP is the number of negative samples incorrectly identified as positive samples, and FN is the number of positive samples incorrectly identified as negative samples.

[0039] The method of the present invention was tested on a self-made traffic police gesture dataset. The results showed that:

[0040] Compared with the original YOLOv11 model, this method improves the mAP50 index by 3.8% (from 86.8% to 90.6%).

[0041] At the same time, the number of model parameters was reduced by 37.6% and GFLOPs was reduced by 12.7%, significantly reducing the computational complexity.

[0042] These improvements effectively improve the efficiency of model inference and make it feasible and practical for efficient deployment on edge computing devices.

[0043] The traffic police command gesture recognition method described in this paper, based on an improved YOLOv11 algorithm, significantly reduces model computational complexity and inference latency while maintaining high accuracy. It is particularly suitable for deployment in vehicle-based autonomous driving systems, enabling real-time and accurate recognition of traffic police gesture commands. Furthermore, this method can also be applied to urban intelligent traffic monitoring, road management systems, human-machine collaborative robots, and 5G / V2X low-latency communication environments, demonstrating excellent scalability and engineering adaptability.

[0044] (3) Beneficial effects

[0045] Compared with the existing technology, the present invention provides a method for identifying traffic police command actions based on improved YOLOv11, which has the following beneficial effects:

[0046] 1. This paper constructs a diverse dataset for traffic police gesture recognition and uses a variety of image enhancement methods, such as image atomization, sharpening, adding Gaussian noise, random rotation, and scaling, to effectively improve the model's generalization ability and robustness in complex weather, multi-gesture, and complex background environments.

[0047] 2. This paper introduces the GPFB module into the YOLOv11 backbone network, integrating global average pooling and local maximum pooling features to achieve efficient expression of multi-scale features and enhance the model's ability to capture the details of traffic police gestures;

[0048] 3. This invention introduces the Global Context Attention Mechanism, combines context modeling with the channel weight generation module, effectively extracts global context semantic information, enhances the model's ability to focus on key areas, and thus significantly improves the accuracy and stability of traffic police action recognition;

[0049] 4. This paper adopts the Lite-BiFPN module to replace the traditional FPN+PAN structure and designs a weighted bidirectional feature fusion path. Through top-down and bottom-up cross-scale feature information fusion, it effectively improves the comprehensive utilization of shallow and deep semantics and enhances the multi-scale feature expression capability. At the same time, with the help of the GhostConv compression module, it reduces the computational complexity and improves the inference efficiency of the model.

[0050] 5. This paper uses indicators such as precision (P), recall (R), and mAP to dynamically evaluate and optimize model performance. Combined with complex test scenarios such as multi-lighting, multi-angle, and occlusion, it verifies the high recognition rate and good robustness of the model in practical applications.

[0051] 6. The overall network structure of the present invention is optimized and reasonable, taking into account both detection accuracy and inference speed. It is suitable for real-time traffic police command action recognition in resource-constrained environments and has good engineering application prospects and promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is an overall flow chart of the traffic police command action recognition method based on improved YOLOv11 according to the present invention;

[0053] Figure 2 Schematic diagram of the network structure of the YOLO-GGB improved model of the present invention;

[0054] Figure 3This is a structural diagram of the GPFB (global-local pooling fusion) module of the present invention;

[0055] Figure 4 This is a structural diagram of the Global Context Attention Mechanism described in the present invention;

[0056] Figure 5 This is a structural diagram of the Lite-BiFPN feature fusion module described in the present invention;

[0057] Figure 6 This is a bar chart comparing the accuracy of the method described in the present invention and the original YOLOv11 model on multiple performance indicators. DETAILED DESCRIPTION

[0058] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0059] Example

[0060] like Figure 1-6 As shown, an embodiment of the present invention proposes a method for identifying traffic police command actions based on improved YOLOv11, which mainly includes the following steps:

[0061] Step 1: Construct a traffic police command action dataset;

[0062] A self-built dataset of traffic police command gesture images was used as the training basis. The original images were primarily sourced from publicly available internet data resources. The Beijing University of Technology's Intelligent Interaction Team collected, filtered, cleaned, and annotated the data, creating a standard action image library covering eight common traffic police gesture commands.

[0063] To improve the model's generalization and robustness in complex scenarios, such as those with varying weather conditions, multiple poses, and complex backgrounds, we incorporated various image enhancement strategies during the image dataset construction process. These enhancements include, but are not limited to, image fogging, image sharpening, Gaussian noise addition, random rotation, and scaling. These enhancements effectively expand the diversity of training samples and enhance the model's adaptability to environmental changes.

[0064] After image enhancement, all images are labeled with bounding boxes and categories using the Label Img tool, ultimately forming a training dataset consisting of 1,000 original images and 4,000 enhanced images, totaling 5,000 training samples, providing sufficient data support for subsequent model training.

[0065] Step 2: Build an improved YOLOv11 network model;

[0066] Based on the original YOLOv11 network, three structural improvements are introduced, including:

[0067] GPFB global-local pooling fusion module;

[0068] Global Context (GC) attention mechanism;

[0069] Lite-BiFPN lightweight bidirectional feature fusion structure.

[0070] The integration of the above improved modules constitutes the YOLO-GGB network structure. The design and function of each module are described as follows:

[0071] (1) Introducing the GPFB module:

[0072] The original Backbone of YOLOv11 has problems such as limited ability to extract high-level semantic information and insufficient expression of detailed features. This can easily lead to a decrease in recognition accuracy, especially in scenes with large motion amplitudes, complex backgrounds, or strong interference. Therefore, the GPFB module is introduced into the Backbone.

[0073] The GPFB module consists of two parallel branches: a global average pooling branch and a local maximum pooling branch. The global average pooling branch extracts the overall contextual semantic information of the image, while the local maximum pooling branch emphasizes the response characteristics of local salient regions, thereby capturing both global semantics and local edge details. The two output features are concatenated along the channel dimension and fused through a 1×1 convolution with the number of channels adjusted to produce an intermediate fused feature map.

[0074] To further preserve the original structural information of the input features and enhance the network's feature learning capabilities, the GPFB module design introduces a residual connection structure, which adds the fused feature map to the original input feature map element-by-element to form the final output. This not only enhances the continuity of feature expression but also helps alleviate the vanishing gradient problem in deep networks, improving the model's ability to model complex traffic police gestures and its robustness.

[0075] The calculation expression of this module is as follows:

[0076]

[0077] F GPFB =Conv 1×1 (Concat[F gap ,F max ])+X;

[0078] Among them, X represents the input feature map, the size is C×H×W; F gap represents the global features obtained by global average pooling, which is used to model the overall context information of the image; F max represents the significant regional response features extracted by local maximum pooling, highlighting the local key areas; Concat[·] represents the concatenation of two feature maps along the channel dimension; Conv1×1(·) represents a 1×1 convolution operation, which is used to fuse the concatenated features and adjust the number of channels; +X represents the residual connection operation, which adds the fused features to the input features to retain the original information and enhance the feature expression capability; F GPFB It is the multi-scale fusion feature map finally output by the module.

[0079] Through residual connections, the GPFB module not only enhances the feature fusion capability, but also significantly improves the network's modeling capability and robustness for complex traffic police gesture motion features.

[0080] (2) Adding GC attention mechanism:

[0081] To improve the model's ability to focus on key areas during traffic police command actions, the Global Context (GC) attention mechanism was introduced in the Neck phase of YOLOv11. This module incorporates spatial context information to enhance the semantic expression of feature channels, effectively suppressing background interference and highlighting the characteristics of the action area, thereby improving recognition accuracy and robustness.

[0082] Its core mechanism involves first acquiring the global spatial semantic information of the input feature map through the context modeling module to generate a global context feature vector. Then, using the channel weight generation module, a multi-layer perceptron (MLP) is used to dynamically calculate the weight coefficients of each channel and apply semantic attention to the original feature map using a channel-by-channel weighted mapping. This mechanism effectively suppresses background interference and highlights the discriminative features of traffic police gestures, thereby improving the recognition accuracy and robustness of the model. The GC attention mechanism consists of the two core modules mentioned above, which respectively implement spatial semantic information modeling and channel attention weight generation.

[0083] The calculation process is as follows:

[0084] Contextual Semantic Modeling:

[0085]

[0086] Channel weight generation:

[0087] W c =σ(MLP(Z c ))

[0088] Weighted output:

[0089] Yc =W c ×X c

[0090] Among them, X c (i, j) represents the pixel value at the cth channel and position (i, j) of the input feature map; Z c represents the global context vector of channel c; MLP(·) represents the multi-layer perceptron structure, which is used to learn the channel-level weight mapping relationship; σ(·) represents the activation function, usually Sigmoid, which is used to normalize the weight to the [0,1] interval; W c represents the global attention weight corresponding to channel c; Y c Represents the final weighted output feature map.

[0091] (3) Replace FPN+PAN with Lite-BiFPN structure:

[0092] In the Neck architecture, the FPN+PAN feature fusion method used in the original YOLOv11 is replaced with the Lite-BiFPN architecture. Lite-BiFPN is based on a weighted bidirectional feature fusion design and uses a top-down and bottom-up dual-path parallel mechanism to achieve cross-scale feature alignment and fusion. Compared to the traditional PAN architecture, Lite-BiFPN not only maintains high-level semantic information but also effectively preserves low-level position information. By introducing a learnable weighted fusion mechanism, it assigns different weights to features at different levels, thereby improving the ability to express multi-scale information. Its feature fusion process can be expressed as:

[0093]

[0094] Among them, P i represents the i-th input feature map (from different levels and scales); w i is the learnable fusion weight corresponding to the input feature map, which is used to control the contribution of features at different scales; ε is a very small constant (such as 1e -4 ), used to avoid the denominator being zero and ensure numerical stability; P out Represents the fused output feature map.

[0095] To further reduce model computational complexity and inference latency, Lite-BiFPN introduces the GhostConv convolution module, which compresses convolution operations, effectively reducing the number of parameters and computational resource consumption, significantly improving the model's deployment efficiency on resource-constrained embedded devices. This module improves multi-scale feature fusion performance while balancing the model's lightweight and real-time requirements, enhancing the model's practical value.

[0096] Step 3: Train the improved model:

[0097] The constructed YOLO-GGB network model was trained using the enhanced traffic police gesture dataset. The training phase was configured as follows: the initial learning rate was set to 0.01, the optimizer used was stochastic gradient descent (SGD), the batch size was 16, and the number of training epochs was 300. A composite loss function was used, consisting of position loss over union (CIoU), category loss (BCE), and confidence loss (BCE). A cosine annealing learning rate decay strategy was introduced during training to improve model stability and convergence speed. Label smoothing was also employed to mitigate overfitting and enhance model generalization. Model performance was dynamically evaluated during training using metrics such as accuracy, recall, mean average precision (mAP), and inference speed (FPS). Early stopping on the validation set was used to control the training process and adjust training parameters appropriately to ensure optimal training weights.

[0098] Step 4: Testing and Performance Evaluation:

[0099] After model training is complete, the trained weights are applied to a pre-specified test set to comprehensively verify the YOLO-GGB model's performance. The test set covers complex scenarios such as multiple lighting conditions, multiple angles, and partial occlusion, comprehensively evaluating the model's recognition accuracy and robustness in real-world applications. Evaluation metrics include accuracy (Precision), recall (Recall), mean average precision (mAP@0.5, mAP@0.5:0.95), model parameter count (Params), and inference speed (FPS).

[0100] The mAP calculation formula is as follows:

[0101]

[0102] Among them, TP is the number of true positive examples, FP is the number of false positive examples, FN is the number of false negative examples, m is the number of categories, and AP is the average accuracy of a single category.

[0103] The test results show that YOLO-GGB has significantly improved performance compared to the original YOLOv11 n model while maintaining low inference latency. Figure 6 As shown in FIG, this figure is a histogram comparison of the key performance indicators of the method described in the present invention and the original YOLOv11 model.

[0104] from Figure 6It can be seen that the YOLO-GGB model has improved by about 3.8% and 4.8% in both mAP@0.5 and mAP@0.5:0.95 indicators; at the same time, the number of model parameters and computational complexity (GFLOPs) have been reduced by about 37.6% and 12.7% respectively, and the inference speed has been increased by about 13.7%, demonstrating excellent lightweight and efficient recognition capabilities.

[0105] In summary, the traffic police command action recognition method based on improved YOLOv11 proposed in the present invention can significantly reduce the model complexity and delay while ensuring recognition accuracy, has good real-time and lightweight characteristics, is suitable for various actual deployment scenarios such as smart cameras and vehicle-mounted terminals, and has strong practicality and promotion value.

[0106] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for traffic police command action recognition based on improved YOLOv11, characterized by: The steps include: S1. Constructing a traffic police command action recognition dataset: A traffic police gesture recognition dataset is constructed based on images collected from public resources on the Internet and actual traffic scenes. After the images are screened, cleaned, and enhanced, standardized annotated samples are generated for training. The annotation format is YOLO format. S2. Build a network model: Based on the YOLOv11 network structure, introduce the GPFB module, Global Context attention mechanism, and Lite-BiFPN multi-scale feature fusion structure to form the improved model YOLO-GGB; S3. Training model: Use the enhanced image dataset to train the YOLO-GGB model, evaluate the model performance through mAP, Precision, and Recall indicators, and select the optimal training weights; S4. Test and Verification: Use the divided test set to perform recognition verification on the trained model, output the target's bounding box position information and corresponding category label, and evaluate the recognition accuracy and robustness of the model in the traffic police command action recognition task.

2. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 1, characterized in that: The image enhancement processing in S1 includes image fogging, sharpening, adding Gaussian noise, random rotation, and scale scaling to improve the model's generalization ability in multiple weather, multiple posture, and complex background scenes.

3. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 1, characterized in that: The network structure improvements of the improved model YOLO-GGB model in S2 include: (1) The GPFB module is embedded in the Backbone network to fuse the global average pooling and local maximum pooling features to enhance the multi-scale feature expression capability. The calculation formula is: F GPFB =Conv 1×1 (Concat[F gap ,F max ])+X; Among them, X represents the input feature map, the size is C×H×W; F gap represents the global features obtained by global average pooling, which is used to model the overall context information of the image; F max represents the significant regional response features extracted by local maximum pooling, highlighting the local key areas; Concat[·] represents the concatenation of two feature maps along the channel dimension; Conv1×1(·) represents a 1×1 convolution operation, which is used to fuse the concatenated features and adjust the number of channels; +X represents the residual connection operation, which adds the fused features to the input features to retain the original information and enhance the feature expression capability; F GPFB It is the multi-scale fusion feature map finally output by the module. (2) The Global Context Attention Mechanism is introduced in the backend of the backbone. It consists of a context modeling module and a channel weight generation module to extract global context semantic information and improve the model's ability to focus on key areas. The calculation process is as follows: Contextual Semantic Modeling: Channel weight generation: IN c =σ(MLP(Z c )) Weighted output: Y c =W c ×X c Among them, X c (i, j) represents the pixel value at the cth channel and position (i, j) of the input feature map; Z c represents the global context vector of channel c; MLP(·) represents the multi-layer perceptron structure, which is used to learn the channel-level weight mapping relationship; σ(·) represents the activation function, usually Sigmoid, which is used to normalize the weight to the [0,1] interval; W c represents the global attention weight corresponding to channel c; Y c Represents the final weighted output feature map. (3) The Lite-BiFPN module is introduced in the Neck part, and a top-down and bottom-up weighted bidirectional feature fusion path is adopted. The GhostConv compression module is combined to reduce the computational complexity. The feature fusion calculation formula is: Among them, P i represents the i-th input feature map, which comes from different levels and scales; w i is the learnable fusion weight corresponding to the input feature map, which is used to control the contribution of features at different scales; ε is a very small constant (such as 1e -4 ), used to avoid the denominator being zero and ensure numerical stability; P out Represents the fused output feature map.

4. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 3, characterized in that: The GPFB module consists of two parallel branches: one is the global average pooling channel and the other is the local maximum pooling channel. The output features of the two are concatenated along the channel dimension and then input into a 1×1 convolution for fusion. The original features are retained through a residual connection structure to enhance the gradient transfer capability.

5. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 3, characterized in that: The Global Context Attention Mechanism consists of a context modeling module and a channel weight generation module. The former obtains spatial semantic information through global context modeling, and the latter uses a multi-layer perceptron (MLP) to achieve dynamic channel weighting.

6. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 3, characterized in that: The Lite-BiFPN module is designed based on a weighted bidirectional feature fusion strategy, including top-down and bottom-up paths, supports cross-scale feature alignment and fusion, and reduces the number of parameters and computational complexity through the GhostConv compression module.

7. A method for identifying traffic police command actions based on improved YOLOv11 according to claim 1, characterized in that: The SGD optimizer was used in the S3 training process, with an initial learning rate of 0.01 and the number of training rounds set to 300. Cosine annealing learning rate decay and label smoothing techniques were introduced to improve model stability and generalization performance.

8. A method for recognizing traffic police command actions based on improved YOLOv11 according to claim 1, characterized in that: The S4 test phase evaluates the model's detection performance on a standard traffic police action test set using metrics including mAP@0.5, Precision, Recall, and inference speed (FPS). Image visualization results are combined to analyze the model's adaptability and robustness under different lighting, angle, and occlusion conditions. The evaluation performance indicators include precision P, recall R and mAP. The specific calculation formula is as follows: Among them, m is the total number of traffic police command action categories, TP is the number of positive samples correctly identified, FP is the number of negative samples incorrectly identified as positive samples, and FN is the number of positive samples incorrectly identified as negative samples.

Images