Multi-target traffic perception and intelligent detection method based on aerial photography of unmanned aerial vehicle

Abstract

The invention discloses a multi-target traffic perception and intelligent detection method based on unmanned aerial vehicle aerial photography, belongs to the technical field of computer vision target detection, and solves the problem that a vehicle target in a complex traffic scene is difficult to accurately recognize and distinguish in the field of unmanned aerial vehicle speed detection, especially the detection precision of a small-target vehicle is insufficient. The problem that the detection accuracy and stability are affected due to insufficient feature expression of a vehicle target caused by limitation of an existing detection algorithm in the feature extraction and fusion process is solved. The problem that real-time and accurate calculation cannot be achieved through a traditional vehicle speed estimation method is solved, the small target detection precision and stability are improved by improving a YOLO frame and introducing a multi-dimensional attention mechanism, a dynamic adaptive feature fusion strategy and a feature alignment dynamic detection head, and efficient and accurate vehicle speed detection is achieved. In addition, the model adopts a lightweight design, so that the calculation complexity and the memory occupation are reduced, and the real-time performance and the high efficiency are ensured.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer vision target detection, and in particular relates to a multi-target traffic perception and intelligent detection method based on unmanned aerial vehicle (UAV) aerial photography. Background Art

[0002] In the field of smart transportation, vehicle speed detection is one of the key technologies for traffic management. At present, vehicle speed detection mainly relies on technologies such as laser, radar and video monitoring. Although laser and radar speed measurement are highly accurate, the equipment cost is high and is easily affected by weather conditions; video speed measurement requires a large amount of labeled data to train the model, which is costly and inefficient. UAV aerial photography technology provides a new solution for vehicle speed detection. The visual sensors it carries can monitor traffic conditions over a large area. However, UAV aerial photography technology has difficulty in accurately identifying and distinguishing vehicles in vehicle speed detection, and existing technologies still have shortcomings in detection accuracy, feature extraction and speed estimation. The vehicle speed estimation method cannot achieve real-time and accurate calculations and cannot meet the needs of smart transportation systems for real-time vehicle speed monitoring. Currently, there is a technical solution similar to the present invention, which is a vehicle speed estimation method based on UAV aerial video. The solution first constructs a data set based on the specific flight altitude of the UAV, and obtains a weight file through deep learning network training. The YOLO detection algorithm and Deep-sort tracking algorithm are then used to obtain the vehicle's trajectory. However, in complex traffic scenarios, the detection accuracy of small vehicles is insufficient, and the feature extraction and fusion strategies are imperfect, resulting in inadequate representation of vehicle target features, affecting the accuracy and stability of detection. Therefore, a new method is needed to efficiently detect vehicle targets in drone aerial imagery and accurately calculate vehicle speed to improve the effectiveness of intelligent traffic management. Based on this, this technical solution proposes a vehicle speed calculation method based on an improved YOLO model and inter-frame matching technology. Summary of the Invention

[0003] The purpose of the present invention is to provide a multi-target traffic perception and intelligent detection method based on drone aerial photography, which solves the problem that traditional methods in the field of drone vehicle speed detection are difficult to accurately identify and distinguish vehicle targets in complex traffic scenes.

[0004] The technical solution adopted is:

[0005] The present invention discloses a multi-target traffic perception and intelligent detection method based on drone aerial photography, which relates to the field of target detection. The method comprises: obtaining a data set of drone aerial photography traffic scene images and performing data enhancement; constructing a PE-YOLO (Pixel Enhanced) network model, improving feature extraction efficiency through a multi-dimensional pixel enhancement attention module (MDPE), enhancing feature expression using a dynamic adaptive feature fusion module (DAFI), improving detection accuracy using a feature alignment dynamic detection head (FADH), and designing an En-IoU loss function optimization model; training a model using the enhanced data set; and using the trained model to detect vehicle targets in an image. In addition, the present invention also relates to inter-frame matching and trajectory tracking, and calculates the real-time speed of a vehicle in combination with camera calibration parameters and time information. The present invention solves the problems of insufficient vehicle detection accuracy and poor real-time performance in the prior art through an improved network structure and speed estimation process, significantly improving the vehicle speed detection performance of drones in intelligent transportation, and is suitable for real-time monitoring applications in complex traffic scenarios.

[0006] The "small target" mentioned in the present invention specifically refers to a detection object that meets the following geometric and semantic characteristics:

[0007] Absolute size constraints:

[0008] The target bounding rectangle area is ≤ 32 2 Pixels (corresponding to the small category standard of the COCO dataset).

[0009] The minimum bounding rectangle aspect ratio ∈ [0.5, 2] (excluding interference from extremely slender targets).

[0010] Relative proportion constraints:

[0011] The target pixel ratio is ≤5% of the input image area (derived based on the typical observation distance of the UAV).

[0012] The central region saliency score is ≤ 0.3 (quantified by ViT visual features).

[0013] The present invention has made improvements and optimizations in terms of detection accuracy, feature expression, real-time performance, adaptability and model lightweighting, and can more accurately identify and distinguish vehicle targets, especially improving the detection accuracy of small target vehicles in complex traffic scenarios; by introducing a multi-dimensional attention mechanism and an adaptive feature fusion strategy, the characteristics of vehicle targets are more fully expressed, the mutual interference between classification and regression tasks is reduced, and the accuracy and stability of detection are improved; combined with the drone platform and the optimized speed estimation model, the demand for real-time vehicle speed monitoring in the intelligent transportation system is better met; the model is lightweight in design, which reduces the computational complexity and memory usage while ensuring high accuracy, making it more suitable for actual vehicle-mounted systems and drone platforms.

[0014] Its advantages are:

[0015] This technical solution addresses the difficulty faced by traditional methods in the field of drone speed detection in accurately identifying and distinguishing vehicle targets in complex traffic scenarios, particularly the inadequate detection accuracy for small vehicle targets. It also addresses the limitations of existing detection algorithms in the feature extraction and fusion processes, which lead to insufficient feature representation of vehicle targets, affecting the accuracy and stability of detection. Furthermore, it addresses the inability of traditional vehicle speed estimation methods to achieve real-time and accurate calculations, a deficiency that hinders their ability to meet the real-time speed monitoring requirements of intelligent transportation systems. By improving the YOLO framework, introducing a multi-dimensional attention mechanism, a dynamic adaptive feature fusion strategy, and a feature-aligned dynamic detection head, the accuracy and stability of small target detection are improved. Combined with the drone platform and speed estimation model, efficient and accurate vehicle speed detection is achieved. Furthermore, the model adopts a lightweight design to reduce computational complexity and memory usage, ensuring real-time and high efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flowchart of the present invention.

[0017] Figure 2 This is the overall structure diagram of PE-YOLO.

[0018] Figure 3 This is the model evaluation index diagram.

[0019] Figure 4 The following is a comparison of the detection effects of YOLOv11 and PE-YOLO in the first scenario.

[0020] Figure 5 The following is a comparison of the detection effects of YOLOv11 and PE-YOLO in the second scenario. DETAILED DESCRIPTION

[0021] The multi-target traffic perception and intelligent detection method based on drone aerial photography is characterized by comprising the following steps:

[0022] Step 1: Collect a dataset of drone aerial images and perform data augmentation on the dataset.

[0023] Step 2: Build the backbone network of the PE-YOLO network model to extract the feature map of the input image.

[0024] Step 3: Add a multi-dimensional pixel enhanced attention module MDPE after the backbone network.

[0025] Step 4: Construct a neck network to further process the feature map and fuse features of different scales.

[0026] Step 5: Add the dynamic adaptive feature fusion module DAFI after the neck network.

[0027] Step 6: Construct the feature alignment dynamic detection head FADH.

[0028] Step 7: Design En-IoU loss function.

[0029] Step 8: Use the data augmentation dataset to train the PE-YOLO network model until the model training is completed.

[0030] Step 9: Input the image to be detected into the trained model, output the location and category information of the target, crop and preprocess the vehicle target image to extract the deep feature vector.

[0031] Step 10: Target matching and trajectory tracking: perform inter-frame matching based on the extracted feature vectors, calculate cosine similarity to associate vehicle targets, combine position information to assist in judgment, and initialize, update, or end the vehicle trajectory.

[0032] Step 11: Speed ​​calculation and result output: Determine the camera's internal and external parameters to complete coordinate transformation, calculate vehicle displacement and speed, and integrate vehicle information into a structured output format.

[0033] The step 1 comprises:

[0034] The hue, saturation, and brightness of the images in the drone aerial image dataset are randomly transformed within a certain range of proportions to simulate the image effects under different lighting and weather conditions; the images in the dataset are randomly translated, randomly scaled, and randomly flipped horizontally to enhance the diversity of the dataset and the model's adaptability to changes in target position; the mosaic data augmentation method is used to combine multiple images into a new image to fuse the features of multiple images and enrich the training samples; the brightness and contrast of the image are randomly adjusted to improve the model's adaptability to different lighting conditions; and the image is randomly cropped to generate sub-images of different sizes and positions, thereby simulating different ways in which the target appears in the image.

[0035] The step 2 includes:

[0036] Step 2-1: The backbone network receives the preprocessed image data and normalizes the pixel values.

[0037] Step 2-2: Extract shallow features of the image through the convolution layer. The convolution kernel size is usually 3×3 or 1×1, and the stride is 1 or 2. It is used to extract basic features such as edges and textures, and use batch normalization and activation functions to enhance feature expression capabilities.

[0038] Steps 2-3: Use multiple layers of convolution and pooling to gradually downsample. Each time the feature map is downsampled, the size is halved while the number of channels is gradually increased to extract deeper semantic features. The C3k2 module replaces the traditional C2f module, using two smaller convolution kernels instead of one large one to accelerate computation while maintaining performance.

[0039] Steps 2-4: Output multiple layers of feature maps at different downsampling stages to form a feature pyramid structure to support subsequent multi-scale target detection.

[0040] The step 3 includes:

[0041] Step 3-1: Integrate the MDPE (Multi-Dimensional Pixel Enhanced) module after the backbone network. First, perform global average pooling on the feature map to obtain the global feature vector of each channel, and then process it through the fully connected layer to generate channel attention weights.

[0042] Step 3-2: Sort by channel importance and assign the channels equally to high- and low-importance feature heads. The high-importance feature head captures long-range dependencies through self-attention, while the low-importance feature head highlights spatial features through spatial attention.

[0043] Step 3-3: The processed high-importance and low-importance feature maps are concatenated and fused with the help of the channel attention mechanism. The fused feature maps are passed through the convolution layer to adjust the number of channels and processed by batch normalization and activation function.

[0044] Step 3-4: Receive the original feature map and the fused attention feature map, expand the attention feature map in the channel dimension, and concatenate it to form a new feature tensor. The concatenated feature tensor is rearranged and input into a convolution layer with a 7x7 convolution kernel using a reflected padding pattern. The output is converted to attention weights using a sigmoid activation function, and the original feature map is weighted and adjusted to generate the final enhanced feature map. It can be expressed as:

[0045] f final =σ(Conv(f concat ))·f original

[0046] Among them, f concat is the concatenated feature map, f original is the original feature map, f final It is the final enhanced feature map.

[0047] Step 3-5: Add the original input and the enhanced feature map through the residual connection to generate the final output feature map.

[0048] The step 4 includes:

[0049] Step 4-1: The neck network receives the multi-layer feature map output by the backbone network and performs feature fusion and further processing.

[0050] Step 4-2: Process the feature map through the C3k2 module to enhance the feature expression capability. The C3k2 module uses two smaller convolution kernels instead of one larger convolution kernel to speed up processing while maintaining performance.

[0051] Step 4-3: Use the feature pyramid structure to fuse feature maps of different scales to generate a multi-scale aggregated feature map.

[0052] The step 5 comprises:

[0053] Step 5-1: The DAFI (Dynamic Adaptive Feature Integration) module receives the multi-scale feature maps from the backbone network and the MDPE module, denoted as F = {F1, F2, …, Fn}.

[0054] Step 5-2: Each feature map is processed by the feature enhancement component. In this component, the input feature map first passes through a depthwise separable convolutional layer to halve the number of channels. A ReLU activation function is then applied to introduce nonlinearity. Next, a depthwise separable convolutional layer restores the number of channels to the original number. Finally, a sigmoid activation function is used to generate feature weights. The output is the weighted sum of the input and enhanced features.

[0055] Step 5-3: The enhanced feature map is subjected to 1×1 convolution to adjust the number of channels and activated by the ReLU function

[0056] Step 5-4: For the enhanced feature map F i On the one hand, a 3×3 pooling kernel and a maximum pooling operation with a stride of 2 are used to generate a downsampled feature map. It can be expressed as:

[0057]

[0058] On the other hand, the feature map size is doubled by bicubic interpolation upsampling operation to obtain the upsampled feature map It can be expressed as:

[0059]

[0060] Finally, the downsampled and upsampled feature maps are spliced ​​in the channel dimension to obtain:

[0061]

[0062] Step 5-5: After the downsampled and upsampled feature maps are concatenated in the channel dimension, a depth-wise separable convolution is performed on the concatenated feature maps to adjust the number of channels.

[0063] Step 5-6: Select a feature map and generate dynamic weights through the convolution layer and Sigmoid activation function:

[0064] W=σ(Conv weight (F1)

[0065] Merging another feature map by weighted fusion can be expressed as:

[0066] F fused =W·F1+(1-W)·F2

[0067] Steps 5-7: The adaptively fused feature map is subjected to 1×1 convolution to adjust the number of channels, and then batch normalization and SiLU activation function operations are performed in sequence.

[0068] Step 6:

[0069] Step 6-1: The FADH (Feature Alignment Dynamic Head) module receives the fused feature map F from the DAFI module final .

[0070] Step 6-2: Decompose the feature map into classification feature map F through the task decomposition module (TAP) cls and regression feature map F reg :

[0071] F cls ,F reg =TAP(F final )

[0072] Step 6-3: Regression feature map F reg Apply dynamic alignment convolution (DAC) to generate the offset map O and mask map M:

[0073] O,M=DAC(F reg )

[0074] Step 6-4: Use the offset map O and the mask map M to dynamically align the regression feature map Freg:

[0075]

[0076] Step 6-5: Align the regression feature map and classification feature map F cls Splicing to generate the final detection feature map F detect :

[0077]

[0078] Step 6-6: Detect feature map F through convolutional layer detect Processing is performed to generate bounding box predictions and category predictions:

[0079] B, C=Conv detect (F detect )

[0080] Among them, B represents the bounding box prediction and C represents the category prediction.

[0081] The step 7 includes:

[0082] Step 7-1: The En-IoU (Enhanced-IoU) loss function module receives the bounding box prediction B and category prediction C from the FADH module, as well as the corresponding ground truth annotations B* and C*.

[0083] Step 7-2: Calculate the inner intersection-over-union (InnerIoU) between the predicted bounding box and the true bounding box, which can be expressed as:

[0084]

[0085] Among them, A inter is the intersection area of ​​the predicted bounding box and the true bounding box, A gt is the area of ​​the ground-truth bounding box.

[0086] Step 7-3: Calculate the aspect ratio difference between the predicted bounding box and the true bounding box, which can be expressed as:

[0087]

[0088] Among them, w pred and h pred is the width and height of the predicted bounding box, w gt and h gt are the width and height of the ground-truth bounding box.

[0089] Step 7-4: Combine the internal intersection-over-union ratio and the aspect ratio difference term to calculate the En-IoU loss, which can be expressed as:

[0090] En-IoU=1-Inner IoU+λ·Δ AR

[0091] Among them, λ is a weight parameter used to balance the contributions of the two terms.

[0092] Step 7-5: Combine the En-IoU loss with the classification loss to form the final total loss function, which can be expressed as:

[0093] Total Loss=Cls Loss+En-IoU Loss

[0094] The step 8 comprises:

[0095] Step 8-1: Divide the data-augmented drone aerial image dataset into a training set and a validation set.

[0096] Step 8-2: Use the training set to train the PE-YOLO network model, update the model parameters through back propagation, and minimize the total loss function.

[0097] Step 8-3: During the training process, regularly use the validation set to evaluate the model performance and adjust hyperparameters such as the learning rate to optimize the training process.

[0098] Step 8-4: After training is completed, save the trained model weights for subsequent detection tasks.

[0099] The step 9 includes:

[0100] Step 9-1: Input the drone aerial image to be detected into the trained PE-YOLO network model and output the initial position and category information of all targets in the image.

[0101] Step 9-2: Perform post-processing operations on the detection results, filter out overlapping detection boxes through non-maximum suppression (NMS), retain the best detection results, and then extract the detected vehicle target image and perform standardization preprocessing.

[0102] Step 9-3: Use a pre-trained deep neural network as a feature extractor, pass the standardized pre-processed image blocks into the network, and extract high-dimensional vectors through multi-layer convolution modules.

[0103] Step 10:

[0104] Step 10-1: In the object association of adjacent frame sequences, a strategy combining feature pyramid matching and motion trajectory prediction is adopted, including the following steps:

[0105] (1) Construct a three-layer feature representation of the target feature vector, which includes the bottom-level spatial detail features, the middle-level semantic and position information features, and the high-level global semantic features;

[0106] (2) Predict the target position through Kalman filtering, perform feature similarity search in the predicted area, and obtain candidate matching targets;

[0107] (3) For candidate matching targets, calculate their cosine similarity at each level of the feature pyramid and obtain the final matching score through weighted fusion. The weight coefficient is dynamically adjusted according to the target size. The matching score is calculated as follows:

[0108]

[0109] Among them, w i is the weight coefficient of the i-th layer feature, l i is the cosine similarity of the i-th layer feature.

[0110] At the same time, combined with position information to assist matching, the distance between the current position and the position in the previous frame is calculated. If the distance exceeds a certain threshold, it is considered not the same target. Based on the matching score and motion consistency check results, the final target association result is output.

[0111] Step 10-2: When a new vehicle target is detected and successfully matched in a subsequent frame, a new track is initialized, a unique ID is assigned, and the initial position and timestamp information is recorded. The track status is updated based on the vehicle target position and timestamp information in the current frame. If the vehicle target goes out of view or is not matched for a long time, the track is terminated.

[0112] Step 11:

[0113] Step 11-1: Determine the camera's intrinsic and extrinsic parameters. Using these parameters, convert the vehicle's target position in the image coordinate system to its actual position in the world coordinate system. Calculate the vehicle's displacement in three-dimensional space based on the change in the vehicle's target world coordinate position in consecutive frames:

[0114]

[0115] Step 11-2: Calculate the time interval Δt between frames using the timestamp information of the current frame and the previous frame. Calculate the vehicle's speed using the displacement and time interval:

[0116] speed=Δd / Δt

[0117] And update the calculated speed information to the trajectory state.

[0118] Step 11-3: Integrate the location, category information, and real-time speed of the vehicle target into a structured output format. The output content includes vehicle ID, location, category, speed, and other information.

[0119] To verify the PE-YOLO algorithm's performance improvement in small object detection, we compared it with the YOLOv11 algorithm and representative algorithms such as Faster R-CNN, SSD, and YOLOv8. The experimental dataset was the VisDrone 2019 public dataset, and the experimental results are shown in Table 1. This comparative experiment shows that SSD performs poorly for small object detection, with a detection accuracy of 23.3%. The Faster R-CNN model is large, at 89.2 MB, making it difficult to achieve lightweight performance. The YOLOv8 algorithm achieved a detection accuracy of 33.0%, which was lower than that of the YOLOv11 algorithm in all public datasets. Therefore, YOLOv11 was selected as the original algorithm to be improved in this invention. The PE-YOLO algorithm achieved a detection accuracy of 37.1%, which is 13.8 percentage points higher than the mAP value of the SSD algorithm and 4.1 percentage points higher than that of the YOLOv8 algorithm. Finally, compared to the baseline model YOLOv11, our method achieves a 2.9 percentage point improvement in mAP50, while maintaining slightly slower inference speed. In summary, our method significantly improves detection accuracy while maintaining a high inference speed, while maintaining a moderate model size and demonstrating excellent practicality.

[0120] Table 1 Comparative experimental results of different algorithms

[0121] Model mAP50 / % Size / M <![CDATA[Params / 10 5 ]]> FPS Faster R-CNN 33.8 89.2 43.75 141.3 SSD 23.3 28.2 24.15 171.2 YOLOv8 33.0 6.25 3.05 177.4 YOLOv11 34.2 5.32 2.57 185.6 PE-YOLO 37.1 5.74 2.92 173.4

[0122] Improved algorithm performance analysis

[0123] Experimental results

[0124] The following is a comparison of the performance of the algorithm of the present invention and the performance of the benchmark algorithm in the VisDrone2019 public dataset. Figure 3 As shown, Figure 3 The figure compares key performance indicators of the two object detection models, YOLOv11 (orange) and PE-YOLO (blue), during training, including precision and recall. The convergence of the detection models is evident, with both models reaching a steady state of convergence with increasing iterations. Under identical training parameter settings, PE-YOLO significantly outperforms YOLOv11 in relevant metrics. This result validates PE-YOLO's ability to detect small objects in public datasets, effectively improving small object detection from drones, and demonstrating its high practical value.

[0125] Detection effect diagram

[0126] To verify the performance of the present invention in real-world contexts, representative backgrounds from the VisDrone2019 test set were selected, including occlusion, dense stacking, low brightness, and complex backgrounds. The results are shown in the table below, comparing the proposed method with the baseline model YOLOv11n algorithm. Figure 4-5 As shown in the comparison chart, PE-YOLO can significantly improve detection accuracy compared to YOLOv11n, especially in complex backgrounds where small targets account for a large proportion and are easily confused. It can accurately identify targets that YOLOv11n misses or misdetects. This verifies the advantages of the present invention in dealing with small target detection backgrounds.

Claims

1. Multi-target traffic perception and intelligent detection method based on UAV aerial photography, characterized by The following steps are involved: Step 1: Collect a dataset of drone aerial images and perform data augmentation on the dataset; Step 2: Build the backbone network of the PE-YOLO network model to extract the feature map of the input image; Step 3: Add a multi-dimensional pixel enhancement attention module MDPE after the backbone network; Step 4: Construct a neck network to further process the feature map and fuse features of different scales; Step 5: Add the dynamic adaptive feature fusion module DAFI after the neck network; Step 6: Construct the feature alignment dynamic detection head FADH; Step 7: Design En-IoU loss function; Step 8: Use the data augmentation dataset to train the PE-YOLO network model until the model training is completed; Step 9: Input the image to be detected into the trained model, output the location and category information of the target, crop and preprocess the vehicle target image to extract the depth feature vector; Step 10: Target matching and trajectory tracking: perform inter-frame matching based on the extracted feature vectors, calculate cosine similarity to associate vehicle targets, combine position information to assist in judgment, and initialize, update, or end the vehicle trajectory; Step 11: Speed ​​calculation and result output: Determine the camera's internal and external parameters to complete coordinate transformation, calculate vehicle displacement and speed, and integrate vehicle information into a structured output format.

2. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 1 comprises: The hue, saturation, and brightness of the images in the drone aerial image dataset are randomly transformed within a certain range of proportions to simulate the image effects under different lighting and weather conditions; the images in the dataset are randomly translated, randomly scaled, and randomly flipped horizontally to enhance the diversity of the dataset and the model's adaptability to changes in target position; the mosaic data augmentation method is used to combine multiple images into a new image to fuse the features of multiple images and enrich the training samples; the brightness and contrast of the image are randomly adjusted to improve the model's adaptability to different lighting conditions; and the image is randomly cropped to generate sub-images of different sizes and positions, thereby simulating different ways in which the target appears in the image.

3. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 2 includes: Step 2-1: The backbone network receives the preprocessed image data and normalizes the pixel values; Step 2-2: Extract shallow features of the image through the convolution layer. The convolution kernel size is usually 3×3 or 1×1, and the stride is 1 or 2. It is used to extract basic features such as edges and textures, and use batch normalization and activation functions to enhance feature expression capabilities; Steps 2-3: Use multi-layer convolution and pooling operations to gradually downsample. After each downsampling, the size of the feature map is halved, while the number of channels is gradually increased to extract deeper semantic features. The C3k2 module is used to replace the traditional C2f module, using two smaller convolution kernels instead of one large convolution kernel to accelerate calculations while maintaining performance. Steps 2-4: Output multiple layers of feature maps at different downsampling stages to form a feature pyramid structure to support subsequent multi-scale target detection.

4. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 3 includes: Step 3-1: Integrate the MDPE (Multi-Dimensional Pixel Enhanced) module after the backbone network; first perform global average pooling on the feature map to obtain the global feature vector of each channel, and then process it through the fully connected layer to generate channel attention weights; Step 3-2: Sort by channel importance and assign the channels equally to high- and low-importance feature heads; the high-importance feature head captures long-range dependencies through self-attention, and the low-importance feature head highlights spatial features through spatial attention; Step 3-3: The processed high-importance and low-importance feature maps are spliced ​​together and assisted in fusion through the channel attention mechanism; the fused feature map is passed through the convolution layer to adjust the number of channels, and is processed through batch normalization and activation function; Step 3-4: Receive the original feature map and the fused attention feature map, expand the attention feature map in the channel dimension and concatenate it to form a new feature tensor; the concatenated feature tensor is rearranged and input into the convolution layer with a 7x7 convolution kernel using a reflection padding mode; the output is converted into attention weights by the Sigmoid activation function, and the original feature map is weighted and adjusted to generate the final enhanced feature map; it can be expressed as: f final =σ(Conv(f concat ))·f original Among them, f concat is the concatenated feature map, f original is the original feature map, f final is the final enhanced feature map; Step 3-5: Add the original input and the enhanced feature map through the residual connection to generate the final output feature map.

5. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 4 includes: Step 4-1: The neck network receives the multi-layer feature map output by the backbone network and performs feature fusion and further processing; Step 4-2: Process the feature map through the C3k2 module to enhance the feature expression capability; the C3k2 module uses two smaller convolution kernels instead of one larger convolution kernel to speed up processing and maintain performance; Step 4-3: Use the feature pyramid structure to fuse feature maps of different scales to generate a multi-scale aggregated feature map.

6. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 5 comprises: Step 5-1: The DAFI (Dynamic Adaptive Feature Integration) module receives the multi-scale feature maps from the backbone network and the MDPE module; denoted as F = {F1, F2, …, Fn}; Step 5-2: Each feature map is processed by the feature enhancement component. In the feature enhancement component, the input feature map first passes through a depthwise separable convolution layer to halve the number of channels, and then applies the ReLU activation function to introduce nonlinearity; then, it passes through another depthwise separable convolution layer to restore the number of channels to the original number; finally, the Sigmoid activation function is used to generate feature weights, and the output is the weighted sum of the input features and the enhanced features; Step 5-3: The enhanced feature map is subjected to 1×1 convolution to adjust the number of channels and activated by the ReLU function Step 5-4: For the enhanced feature map F i On the one hand, a 3×3 pooling kernel and a maximum pooling operation with a stride of 2 are used to generate a downsampled feature map. Expressed as: On the other hand, the feature map size is doubled by bicubic interpolation upsampling operation to obtain the upsampled feature map Expressed as: Finally, the downsampled and upsampled feature maps are spliced ​​in the channel dimension to obtain: Step 5-5: After the downsampled and upsampled feature maps are concatenated in the channel dimension, a depth-wise separable convolution is performed on the concatenated feature maps to adjust the number of channels. Step 5-6: Select a feature map and generate dynamic weights through the convolution layer and Sigmoid activation function: W=σ(Conv weight (F1)) Merge another feature map by weighted fusion, expressed as: F fused =W·F1+(1-W)·F2 Steps 5-7: The adaptively fused feature map is subjected to 1×1 convolution to adjust the number of channels, and then batch normalization and SiLU activation function operations are performed in sequence.

7. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: Step 6: Step 6-1: The FADH (Feature Alignment Dynamic Head) module receives the fused feature map F from the DAFI module final ; Step 6-2: Decompose the feature map into classification feature map F through the task decomposition module (TAP) cls and regression feature map F reg : F cls ,F reg =TAP(F final ) Step 6-3: Regression feature map F reg Apply dynamic alignment convolution (DAC) to generate the offset map O and mask map M: O,M=DAC(F reg ) Step 6-4: Use the offset map O and the mask map M to dynamically align the regression feature map Freg: Step 6-5: Align the regression feature map and classification feature map F cls Splicing to generate the final detection feature map F detect : Step 6-6: Detect feature map F through convolutional layer detect Processing is performed to generate bounding box predictions and category predictions: B,C=Conv detect (F detect ) Among them, B represents the bounding box prediction and C represents the category prediction.

8. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 7 includes: Step 7-1: The En-IoU (Enhanced-IoU) loss function module receives the bounding box prediction B and category prediction C from the FADH module, as well as the corresponding ground truth annotations B* and C*; Step 7-2: Calculate the inner intersection-over-union (InnerIoU) between the predicted bounding box and the true bounding box, expressed as: Among them, A inter is the intersection area of ​​the predicted bounding box and the true bounding box, A gt is the area of ​​the ground-truth bounding box; Step 7-3: Calculate the aspect ratio difference between the predicted bounding box and the true bounding box, expressed as: Among them, w pred and h pred is the width and height of the predicted bounding box, w gt and h gt are the width and height of the ground-truth bounding box; Step 7-4: Combine the internal intersection-over-union ratio and the aspect ratio difference term to calculate the En-IoU loss, expressed as: En-IoU=1-Inner IoU+λ·Δ AR Among them, λ is a weight parameter used to balance the contributions of the two items; Step 7-5: Combine the En-IoU loss with the classification loss to form the final total loss function, expressed as: Total Loss=Cls Loss+En-IoU Loss.

9. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 8 comprises: Step 8-1: Divide the data-augmented drone aerial image dataset into a training set and a validation set; Step 8-2: Use the training set to train the PE-YOLO network model, update the model parameters through back propagation, and minimize the total loss function; Step 8-3: During the training process, regularly use the validation set to evaluate the model performance and adjust hyperparameters such as the learning rate to optimize the training process; Step 8-4: After training is completed, save the trained model weights for subsequent detection tasks.

10. The multi-target traffic perception and intelligent detection method based on drone aerial photography according to claim 1 is characterized in that The following steps are involved: The step 9 includes: Step 9-1: Input the drone aerial image to be detected into the trained PE-YOLO network model and output the initial position and category information of all targets in the image; Step 9-2: Perform post-processing on the detection results, filter out overlapping detection boxes through non-maximum suppression (NMS), retain the best detection results, and then extract the detected vehicle target image and perform standardization preprocessing; Step 9-3: Use a pre-trained deep neural network as a feature extractor, pass the standardized pre-processed image blocks into the network, and extract high-dimensional vectors through multi-layer convolution modules; Step 10: Step 10-1: In the object association of adjacent frame sequences, a strategy combining feature pyramid matching and motion trajectory prediction is adopted, including the following steps: (1) Construct a three-layer feature representation of the target feature vector, which includes the bottom-level spatial detail features, the middle-level semantic and position information features, and the high-level global semantic features; (2) Predict the target position through Kalman filtering, perform feature similarity search in the predicted area, and obtain candidate matching targets; (3) For candidate matching targets, calculate their cosine similarity at each level of the feature pyramid and obtain the final matching score through weighted fusion. The weight coefficient is dynamically adjusted according to the target size. The matching score is calculated as follows: Among them, w i is the weight coefficient of the i-th layer feature, l i is the cosine similarity of the i-th layer feature; At the same time, combined with position information to assist matching, the distance between the current position and the position of the previous frame is calculated. If the distance exceeds a certain threshold, it is considered not the same target; based on the matching score and motion consistency check results, the final target association result is output; Step 10-2: When a new vehicle target is detected and successfully matched in a subsequent frame, a new track is initialized, a unique ID is assigned, and the initial position and timestamp information is recorded. The track status is updated based on the vehicle target position and timestamp information of the current frame. When the vehicle target is out of the field of view or has not been matched for a long time, the track is terminated. Step 11: Step 11-1: Determine the camera's intrinsic and extrinsic parameters; use these parameters to convert the vehicle's target position in the image coordinate system to its actual position in the world coordinate system; calculate the vehicle's displacement in three-dimensional space based on the change in the vehicle's target world coordinate position in consecutive frames: Step 11-2: Calculate the time interval Δt between frames using the timestamp information of the current frame and the previous frame; calculate the vehicle's speed using the displacement and time interval: speed=Δd / Δt And update the calculated speed information to the trajectory state; Step 11-3: Integrate the location, category information, and real-time speed of the vehicle target into a structured output format. The output content includes vehicle ID, location, category, speed, and other information.

Images