Unmanned aerial vehicle detection method and system based on improved YOLOv11n

Abstract

The invention discloses an unmanned aerial vehicle detection method and system based on improved YOLOv11n, and the method comprises the steps: constructing an unmanned aerial vehicle road vehicle data set, and dividing a training set and a verification set; the method comprises the following steps: constructing an improved YOLOv11n-SDEM unmanned aerial vehicle target detection network, wherein the improved YOLOv11n-SDEM unmanned aerial vehicle target detection network comprises a backbone network, a neck network, a target detection head and a loss function; the improved YOLOv11n-SDEM unmanned aerial vehicle target detection network is used as a detection model, the detection model is trained and verified, and a final detection model is obtained; and outputting a vehicle target detection result by using the final detection model and taking a road vehicle image or video frame aerially photographed by the unmanned aerial vehicle as input. According to the method provided by the invention, the detection accuracy of the small target of the aerial vehicle of the unmanned aerial vehicle can be improved, relatively high reasoning efficiency is kept, and the interpretability of the model is greatly improved on the premise of keeping the precision.

Description

Technical Field

[0001] The present invention relates to the technical field of computer vision target detection, and in particular to a drone vehicle detection method and system based on improved YOLOv11n. Background Art

[0002] Traffic object detection is a key computer vision technology in smart transportation scenarios and provides fundamental support for building intelligent traffic management systems. Efficient and accurate identification of vehicles and other targets in road environments can effectively improve road safety, optimize traffic scheduling efficiency, and reduce accident rates. With the rapid development of drone technology, drone monitoring platforms have demonstrated significant advantages in traffic regulation due to their low maintenance costs, wide coverage, and high flexibility. By flexibly adjusting flight altitude and observation angle, drones can dynamically track moving targets and avoid ground obstructions, providing a new solution for traffic monitoring.

[0003] However, vehicle detection from a drone perspective still faces many technical challenges, including the following:

[0004] (1) Small target recognition challenges: To obtain traffic flow information with a wider coverage area, drones typically use high-altitude, wide-angle shooting modes, which significantly reduces the pixel ratio of vehicle targets in the image and places higher demands on the fine-grained perception capabilities of the detection algorithm.

[0005] (2) Complex background interference: Aerial images often contain interference information from non-road areas such as buildings and vegetation, which can easily lead to target misidentification or missed detection, reducing the accuracy and reliability of the detection algorithm;

[0006] (3) Algorithm performance balance problem: Under the constraints of limited computing resources, there is a technical bottleneck in meeting the detection accuracy and real-time requirements at the same time. Especially in dense traffic scenarios, it is necessary to coordinate and optimize both accuracy and computing efficiency.

[0007] Existing technologies have limitations in addressing these issues. For example, while the traditional YOLO series of algorithms excels in real-time performance, they still need to improve their accuracy in detecting small objects and adapting to complex backgrounds. Furthermore, existing technologies lack robustness when dealing with dynamic environmental changes, limiting the widespread application of drone traffic monitoring systems. Summary of the Invention

[0008] To overcome the shortcomings and deficiencies of existing drone-based traffic detection technologies, the present invention leverages the advantages of drones to provide a drone vehicle detection method and system based on an improved YOLOv11n. This method addresses the problem of small vehicle target detection models in drone aerial videos and images being affected by false detections and missed detections when extracting global features, resulting in reduced vehicle detection accuracy. The proposed method can improve the accuracy of small vehicle target detection in drone aerial videos, maintain high inference efficiency, and significantly enhance the interpretability of the model while maintaining accuracy.

[0009] The purpose of the present invention is achieved through the following technical solutions:

[0010] A drone vehicle detection method based on improved YOLOv11n, including:

[0011] S1 builds the UAV road vehicle dataset DRC Dataset and divides it into training set and validation set;

[0012] S2 constructs an improved YOLOv11n-SDEM drone vehicle target detection network, wherein the improved YOLOv11n-SDEM drone vehicle target detection network includes a backbone network, a neck network, a target detection head, and a loss function;

[0013] S3 uses the improved YOLOv11n-SDEM UAV vehicle target detection network as the detection model, and uses the training set and validation set to train and verify the detection model to obtain the final detection model;

[0014] S4 uses the final detection model, takes drone aerial images or video frames of road vehicles as input, and outputs vehicle target detection results.

[0015] Furthermore, the improved YOLOv11n-SDEM UAV vehicle target detection network includes a backbone network, a neck network and a target detection head;

[0016] In the backbone network, the SRConv module is used to replace the Conv modules of the first, third, fifth and seventh layers in the original backbone network; the C3K2-DBB module is used to replace the C3k2 module in the original backbone network;

[0017] In the neck network, a PAN-FPN structure is adopted, and the C3K2-DBB module replaces the C2f module in the original neck network;

[0018] The object detection head, specifically the dynamic detection head, adopts a decoupled structure, in which independent branches are used to predict category and position information respectively, and different loss functions are selected according to different tasks.

[0019] Furthermore, the SRConv module combines the SPD layer with the RFACon layer, specifically:

[0020] The SPD layer converts the spatial dimension of the input feature map into the channel dimension while ensuring that the information within the channel is preserved;

[0021] After the SPD layer, the RFAConv layer is used to further extract fine-grained small target features from the feature map output by the SPD layer. The receptive field attention mechanism is introduced to solve the sharing problem of convolution kernel parameters. By considering the global importance of each feature within the receptive field, the network's attention to targets of different scales is adaptively adjusted.

[0022] Furthermore, the backbone network also introduces a spatial pyramid pooling fast module to pool the feature map to a fixed size, and then further enhances the feature extraction capability through the C2PSA module, which sets the pyramid slice attention mechanism.

[0023] Furthermore, the dynamic detection head includes scale-aware attention, space-aware attention, and task-aware attention;

[0024] Scale-aware attention:

[0025] It is deployed only on the level L dimension, by scaling the feature pyramid to the same scale to form a 3D tensor, which is then used as the input of the dynamic detection head;

[0026] Spatial Perception Attention:

[0027] Deployed in the spatial S dimension, it focuses on the distinctive areas in the image by aggregating features;

[0028] Task-aware attention:

[0029] Deployed in the channel C dimension, feature channels are dynamically opened or closed according to the response of different convolution kernels to objects to support different tasks, including classification, bounding box regression and center point learning.

[0030] Furthermore, the improved YOLOv11n-SDEM UAV vehicle target detection network is trained, using the MPDIoU loss function for the bounding box loss and the EMA-SlideLoss loss function for the classification loss.

[0031] Furthermore, the C3k2-DBB module integrates the DBB diversified branch block, which increases complexity and improves feature representation during training, and is converted to a single branch during inference to maintain efficiency. After training, it is equivalent to being converted to a single convolutional layer for deployment.

[0032] Furthermore, the UAV road vehicle dataset DRC Dataset is constructed and divided into a training set and a validation set. Specifically:

[0033] Use a drone to hover at high altitude to shoot videos of road traffic scenes, then perform frame segmentation on the videos, build an automated labeling tool chain, label the vehicle targets in the images, and divide the labeled images into training and validation sets for model training and verification.

[0034] Furthermore, the training process is as follows: the training set is input into the improved YOLOv11n-SDEM UAV vehicle target detection network for training until the loss function converges. At the same time, the precision, recall rate and average precision of the validation set are used as indicators to evaluate the model performance and obtain the detection model.

[0035] A system for implementing the UAV vehicle target detection method includes:

[0036] Dataset processing module: used to obtain and annotate the UAV road vehicle dataset and divide it into training set, validation set and test set;

[0037] Model building module: used to build an improved YOLOv11n-SDEM drone vehicle target detection network;

[0038] Model training module: Used to train the improved YOLOv11n-SDEM UAV vehicle target detection network using the DRC Dataset training set according to the set model network training parameters, and output the training weight file after the training is completed; the training weight file is verified through the validation set, and the training weight file with the highest accuracy is selected as the optimal weight file to obtain the trained improved YOLOv11 UAV vehicle target detection network;

[0039] Target detection module: Input the road vehicle video frames or images taken by drones and the optimal weight files into the trained YOLOv11n-SDEM drone vehicle target detection network to perform road vehicle target detection and output the vehicle target detection results.

[0040] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0041] Introducing the SRConv module to replace the standard convolution module and enhance feature extraction capabilities;

[0042] Design the C3k2-DBB module to replace the original C3k2 module, optimize the network structure, and improve model performance;

[0043] Develop a dynamic detection head (Detect-Dyhead) to improve the accuracy of small target detection;

[0044] Introducing the exponential moving average sliding loss function (EMA-SlideLoss) and MPDIoU loss function to optimize loss calculation and enhance model robustness;

[0045] The DRC Dataset is used to train the YOLOv11n-SDEM drone vehicle target detection network and optimize the model parameters. In the inference phase, the images to be detected are input into the trained model to output high-precision vehicle detection results.

[0046] The present invention addresses the problems of low small target detection accuracy and large complex background interference in vehicle detection from the perspective of existing drones. It can achieve high-precision vehicle small target detection in complex environments such as urban traffic monitoring and highways, and has significant technical advantages and broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is the overall flow chart of the method of the present invention;

[0048] Figure 2 This is the YOLOv11n-SDEM UAV vehicle target detection network structure diagram proposed by the present invention;

[0049] Figure 3 This is the YOLOv11n-SDEM UAV vehicle target detection network training framework diagram of the present invention;

[0050] Figure 4 This is a diagram of the DBB module structure during model training of the present invention;

[0051] Figure 5 This is the YOLOv11n-SDEM UAV vehicle target detection network reasoning framework diagram of the present invention;

[0052] Figure 6 This is a diagram of the DBB module structure during model reasoning in the present invention;

[0053] Figure 7 1 is a structural diagram of the SPD layer of the present invention;

[0054] Figure 8 This is the RFAConv structure diagram of the present invention;

[0055] Figure 9 This is a schematic diagram of the four structures of the present invention being reparameterized into a single convolution:

[0056] Figure 10 This is a structural diagram of the C3k2-DBB module of the present invention;

[0057] Figure 11 It is a structural diagram of the dynamic detection head (Dyhead) of the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0059] Example

[0060] In the first aspect, this embodiment provides a drone vehicle detection method based on improved YOLOv11n, such as Figure 1 As shown, the following steps are included:

[0061] Step 1: Construct the DroneRoadVehicle Dataset (DRCDataset for short). Specifically, use a quadrotor drone to hover 230 meters above the road, with the onboard camera perpendicular to the ground to collect traffic scene videos. Use the drone to hover at high altitude to shoot road traffic scene videos, then perform frame segmentation on the video, build an automated annotation tool chain to mark vehicle targets, and divide the obtained images into training set and validation set, which are divided into training set and validation set in a ratio of 8:2.

[0062] Step 2: Build an improved YOLOv11n-SDEM drone vehicle target detection network, such as Figure 2 As shown, the improved YOLOv11n-SDEM UAV vehicle target detection network includes a backbone network, a neck network, a target detection head and a loss function. Specifically: the SRConv module is designed to replace the standard convolution module (Conv) in the original backbone network, and the feature extraction capability of small target vehicles is significantly enhanced by introducing the spatial attention mechanism. The C3k2-DBB module is designed to replace the C3k2 module in the original backbone network and the neck network, optimize the network structure, and improve the model performance. The dynamic detection head (Dyhead) is designed to improve the detection head and improve the accuracy of small target detection. The exponential moving average sliding loss (EMA-SlideLoss) and MPDIoU loss function are introduced to optimize the original loss function and enhance the robustness of the model in complex backgrounds;

[0063] Furthermore, the SRConv module is designed to improve the backbone network.

[0064] The original YOLOv11n backbone network may cause a certain amount of fine-grained information loss during convolution and pooling operations. In order to reduce the false detection and missed detection rates of small targets in drone images, the present invention combines the Space-to-Depth (SPD) layer with the Receptive Field Attention Convolution (RFAConv) to design the SRConv module, which replaces the standard convolution modules (Conv) in the 1st, 3rd, 5th and 7th layers of the original backbone network. The SRConv module combines the advantages of the SPD layer in retaining all information within the channel dimension and the ability of RFAConv to adaptively adjust the network's attention to targets of different scales, and can extract richer small target features and fine-grained information.

[0065] first, Figure 7 As shown in Figure 2, the SPD layer converts the spatial dimension of the input feature map into the channel dimension while ensuring that the information within the channel is preserved, thereby avoiding the information loss problem that may occur in traditional methods. Assuming that the size of the input image X is S×S×C1, in each step, the image will be divided into multiple sub-feature sequences. The specific process is shown in formulas (1) to (3):

[0066]

[0067]

[0068] In general, for any given (original) feature Figure X , a sub-graph is composed of all elements, where these elements can be separated by scale. Therefore, each sub-graph will be downsampled by the scale factor scale. Figure 3 In the example, given scale = 2, the four shapes generated are Then, these sub-feature maps are concatenated along the channel dimension to obtain Feature map. The SPD layer converts the feature Figure X (S,S,C1) is converted to Feature map.

[0069] After the SPD layer, the RFAConv module is used to further extract features. The RFAConv module effectively solves the problem of convolution kernel parameter sharing by introducing the receptive field attention mechanism (RFA), and fully considers the global importance of each feature in the receptive field. This mechanism enables the network to adaptively adjust the degree of attention to targets of different scales, thereby improving the detection performance of small targets and dense targets. The structure of RFAConv is as follows: Figure 4As shown. Assuming that the size of the input image is C×H×W, the network enhances performance by means of interactive receptive field features. In order to reduce the amount of computation and the number of parameters, the average pooling operation is first used to aggregate the global information of the receptive field features in the first branch, and the information is transferred through the 1×1 group convolution, and finally the attention map of size 9C×H×W is generated through the Softmax activation function. In the second branch, 3×3 group convolution is used for normalization, and nonlinear features are introduced through the ReLU activation function, and finally the receptive field spatial features of size 9C×H×W are obtained. Subsequently, the feature maps of the two branches are fused by element-by-element multiplication to reshape the weights and aggregate relevant information from different channels. After shape adjustment, a non-overlapping receptive field spatial feature map of size C×3H×3W is obtained. Finally, the convolution operation is used to adjust the length and width of the feature map to restore it to the same size as the input image. The specific calculation process is shown in formula (4):

[0070] RFA=Softmax(g (1×1) (Avgpool(X)))×ReLU(Norm(g (s×s) (X)))=A rfa ×F rfa (4)

[0071] Among them, s is the convolution kernel size, g (1×1) Represents a group convolution of size 1×1, Norm is normalization, AvgPool is the average pooling operation, ReLU and Softmax are activation functions, X is the input feature map, and RFA is the receptive field spatial feature F after conversion. rfa and attention map A rfa Multiply the result.

[0072] Further, such as Figure 10 As shown, the C3k2-DBB module is designed to improve the backbone network and neck network:

[0073] In YOLOv11n, due to the limitation of the single receptive field of the standard convolutional layer Conv, the C3k2 module is difficult to adapt to the scale changes of small vehicle targets, especially the drastic fluctuations in vehicle size caused by drones hovering high in the air to shoot road traffic scenes, which leads to the loss of some features of small drone vehicle targets. In order to enable the model to reduce the false detection and missed detection rates of small vehicle targets in drone images or video frames, the diversified branch block (DBB) is applied to the C3k2 module, and the DBB module is used to replace the standard convolution layer (Conv) in the Bottleneck module. The Bottleneck-DBB module is designed, and then the C3k2-DBB module is designed to replace the C3k2 module of the original backbone network and neck network. The C3k2-DBB module integrates the multi-scale feature fusion of the DBB module and enhances the nonlinear modeling capabilities, replaces the C3k2 module in the original backbone network and neck network, and extracts more fine-grained information of small target features. The DBB module structure is as follows Figure 5 and Figure 6 shown.

[0074] The Diversified Branch Block (DBB) increases complexity to improve feature representation during training and converts to a single branch during inference to maintain efficiency. After training, it can be equivalently converted to a single convolutional layer for deployment.

[0075] Training phase: DBB has a diversified branch structure, combining convolutional layers of different sizes and average pooling layers, which are arranged in a complex way and finally merge the outputs, such as Figure 5 shown.

[0076] Inference stage: Diversified branch structures will introduce too many parameters. In order to achieve efficient operation in the inference stage, they need to be reparameterized to compress the number of parameters. This example provides four methods to reparameterize the diversified branch structure into a single convolution, including: reparameterizing the convolution connection normalization layer into a single convolution, reparameterizing the continuous convolution into a single convolution, reparameterizing the average pooling into a single convolution, and reparameterizing the depth-wise splicing convolution into a single convolution. Figure 9 shown.

[0077] (1) The convolutional connection normalization layer is reparameterized into a single convolution

[0078] The convolution kernel is a fourth-order tensor with C input channels, D output channels and kernel size K, denoted as F∈R D ×C×K×K , the optional bias is denoted by b∈R D , such a convolution kernel is based on the feature map I∈R of C channels C×H×W As input, output D channel feature map O∈R D×H′×W′ , where H′ and W′ are determined by K, padding, and step size configuration. Copy the bias to REP(b)∈R D×H′×W′ , and add it to the convolution result. The convolution formula is expressed as follows:

[0079]

[0080] The output of the convolutional connection normalization layer can be expressed as follows:

[0081]

[0082] Among them, j is the index of the output channel, σ j is the cumulative channel standard deviation, μ j is the cumulative channel average, γ j is the scaling factor to be learned, β j Is the bias term to be learned. From the above formula, the convolution kernel F′ and bias b′ are used to represent the re-parameterized convolution.

[0083] (2) Continuous convolution is reparameterized into a single convolution

[0084] Assume that the convolution kernel shapes of the 1×1 layer and the K×K layer are D×C×1×1 and E×D×K×K respectively, where D can be any value. Fuse the two normalization layers into the two convolution layers to get F (1) ∈R D×C×1×1 , b (1) ∈R D , F (2) ∈R E ×D×K×K and b (2) ∈R E The output is:

[0085]

[0086] because It is a 1×1 convolution, which only performs linear combination within the channel without spatial aggregation, so its parameters can be linearly reorganized and merged into the K×K convolution, that is:

[0087]

[0088] Among them, TRANS(F (1) )∈R C×D×1×1 It is from F (1) The transposed tensor. The second term of Equation (6) is the convolution of the constant matrix, so the output is also a constant matrix.

[0089] Based on this, we can build

[0090]

[0091] Then it can be verified that:

[0092]

[0093] Thus:

[0094]

[0095] (3) Deep convolution is re-parameterized into a single convolution

[0096] When using depthwise concatenation to combine branches, when each branch contains only a single convolutional layer and all convolutional layers have the same parameter configuration, depthwise concatenation is equivalent to a convolution whose kernels are concatenated along the axis that distinguishes the output channels. Can be spliced ​​into Obviously there are:

[0097]

[0098] (4) Average pooling is reparameterized into a single convolution

[0099] For C channels, average pooling with kernel size K and stride s is equivalent to convolution with the same K and s. Such convolution F′∈R C×C×K×K It can be constructed by the following formula:

[0100]

[0101] Furthermore, a dynamic detection head (Detect-Dyhead) is designed to improve the target detection head:

[0102] In order to embed scale perception, spatial position perception and task perception into a target detection head at the same time, the original detection head of YOLOv11 is improved and the dynamic head mechanism (Dyhead) is introduced, such as Figure 11 As shown in the figure, a dynamic detection head (Detect-Dyhead) is designed, which includes three key components: scale-aware attention, space-aware attention, and task-aware attention.

[0103] (1) Scale-aware attention

[0104] Deployed only on the level (L) dimension, a 3D tensor is formed by scaling the feature pyramid to the same scale. This is then used as input to the dynamic header.

[0105] (2) Spatial Perception Attention

[0106] Deployed in the spatial (S) dimension (i.e., height × width), it focuses on distinctive regions in the image by aggregating features.

[0107] (3) Task-aware attention

[0108] Deployed in the channel (C) dimension, it dynamically opens or closes feature channels to support different tasks (such as classification, bounding box regression, and center point learning) based on the response of different convolution kernels to objects.

[0109] The dynamic head applies these attention mechanisms on different dimensions of the feature tensor so that the performance of each dimension can complement each other. The formula of its core attention mechanism can be expressed as:

[0110]

[0111] in, Represents the feature tensor This function is further decomposed into 3 consecutive attentions, each focusing on a different dimension (level, space and channel) of the feature tensor:

[0112]

[0113] Formulas (13) and (14) show that the dynamic head dynamically adjusts the focus and processing of features based on the semantic level, spatial position and task-related channels of different dimensions, thereby significantly improving the detection accuracy. The structure of the Dyhead block is as follows Figure 8 shown.

[0114] Furthermore, the EMASlideLoss loss function and MPDIoU loss function are introduced to improve the original loss function:

[0115] (1) For the bounding box loss, the MPDIoU loss function is used. MPDIoU is a loss function commonly used in target detection and instance analysis tasks. It can effectively improve the MPDIoU loss function compared to the CIoU loss function. It has advantages such as stronger robustness and more accurate bounding box regression. It greatly improves the positioning ability of small target detection in drones and effectively improves the performance of drone target detection algorithms.

[0116] L box =1-MPDIoU (15)

[0117]

[0118] Where IoU is the intersection over union (IoU) of the predicted bounding box and the ground-truth bounding box, d1 and d2 are the Euclidean distances between the diagonals of the predicted and ground-truth bounding boxes, and h and w are the height and width of the bounding box. The introduction of the MPDIoU loss function not only considers bounding box overlap but also their relative position and shape, improving bounding box alignment. Compared to the traditional IoU loss function, it achieves higher accuracy and quality in bounding box evaluation.

[0119] (2) For classification loss, the improved exponential average sliding loss (EMA-Slideloss) loss function is used. EMA-SlideLoss is a category classification loss function that can effectively deal with sample imbalance, reduce instability, oscillation, and gradient explosion in model training, and make model training more stable.

[0120] For SlideLoss, its classification loss function L SlideLoss It can be expressed by the following formula:

[0121]

[0122] Among them, μ refers to the average loU value of all bounding boxes. Negative samples are taken for samples smaller than μ, and positive samples are taken for samples larger than μ. However, this operation will often cause samples near the boundary to suffer greater losses due to unclear classification. In order to enable the model to make more full use of these boundary samples to train the network, the samples are first divided into positive samples and negative samples by the parameter μ, and then the sliding loss SlideLoss is used to strengthen the samples at the boundary. Subsequently, based on the sliding loss SlideLoss, the sliding average EMA idea is introduced for optimization, that is, the exponential moving average is used to dynamically adjust the value of μ, and adaptive adjustments are made according to the adjusted value. As shown in the following formula:

[0123]

[0124] Where d is the decay factor, which is set to 0.99 here, t is the number of iterations, and tau is the total number of iterations, which is set to 10,000 here.

[0125] In summary, the classification loss L cls for:

[0126]

[0127] Step 3: Use the DRC Dataset to train and verify the improved YOLOv11 drone vehicle detection model and optimize the model parameters;

[0128] The training set of the UAV road vehicle dataset constructed in Step 1 is fed into the YOLOv11n-SDEM UAV vehicle target detection network for training until the loss function converges. The precision, recall, and mean average precision (mAP) of the validation set are used as metrics to evaluate the model's performance.

[0129] Precision refers to the proportion of samples that are actually positive among those predicted by the model. The calculation formula is:

[0130]

[0131] Recall refers to the proportion of samples that are actually positive that are correctly predicted by the model as positive. The calculation formula is:

[0132]

[0133] Among them, TP (True Positive) represents the true class, TN (True Negative) represents the true negative class, FP (False Positive) represents the false positive class, and FN (False Negative) represents the false negative class.

[0134] With Recall as the horizontal coordinate and Precision as the vertical coordinate, the average precision AP is the area enclosed by the Precision-Recall curve and the coordinate axis. The average AP of each target category in the same data set is called mAP. The calculation method of mAP is shown in formula (11):

[0135]

[0136] Where r is the rth type of target detected.

[0137] Step 4: Input the road vehicle video frames or images taken by the drone into the trained detection model to output high-precision vehicle detection results.

[0138] In this embodiment, step 4 is specifically as follows:

[0139] Import the weight file trained in step 3 into the detection program, add the prepared test set data path to the program and run the program to obtain the actual detection effect diagram of the YOLOv11n-SDEM drone vehicle target detection model for drone aerial photography vehicle targets.

[0140] Secondly, this example also provides a small target detection system for drone aerial photography based on the improved YOLOv11n algorithm, including:

[0141] Dataset processing module: Construct the UAV road vehicle dataset DRC Dataset, use the drone to hover at high altitude to shoot road traffic scene videos, then perform frame segmentation processing on the video, build an automated annotation tool chain to mark vehicle targets, and divide the obtained images into training and validation sets, with the training and validation sets divided in an 8:2 ratio;

[0142] Model building module: configured to build the YOLOv11n-SDEM drone vehicle target detection model;

[0143] Model training module: used to train the constructed YOLOv11n-SDEM UAV vehicle target detection model using the training set of the dataset DRC Dataset according to the set model network training parameters, and output the training weight file after the training is completed; the training weight file is verified through the validation set, and the training weight file with the highest accuracy is selected as the optimal weight file to obtain the trained YOLOv11n-SDEM UAV vehicle target detection model;

[0144] Target detection module: Input the road vehicle video frames or images taken by the drone and the optimal weight file into the trained YOLOv11n-SDEM drone vehicle target detection model to perform road vehicle target detection and output the vehicle target detection results.

[0145] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the embodiments. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be considered as equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A drone vehicle detection method based on improved YOLOv11n, characterized in that: include: S1 builds the UAV road vehicle dataset DRC Dataset and divides it into training set and validation set; S2 constructs an improved YOLOv11n-SDEM drone vehicle target detection network, wherein the improved YOLOv11n-SDEM drone vehicle target detection network includes a backbone network, a neck network, a target detection head, and a loss function; S3 uses the improved YOLOv11n-SDEM UAV vehicle target detection network as the detection model, and uses the training set and validation set to train and verify the detection model to obtain the final detection model; S4 uses the final detection model, takes drone aerial images or video frames of road vehicles as input, and outputs vehicle target detection results.

2. The UAV vehicle detection method according to claim 1, characterized in that: The improved YOLOv11n-SDEM UAV vehicle target detection network includes a backbone network, a neck network and a target detection head; In the backbone network, the SRConv module is used to replace the Conv modules of the first, third, fifth and seventh layers in the original backbone network; Use C3K2-DBB modules to replace the C3k2 modules in the original backbone network; In the neck network, a PAN-FPN structure is adopted, and the C3K2-DBB module replaces the C2f module in the original neck network; The object detection head, specifically the dynamic detection head, adopts a decoupled structure, in which independent branches are used to predict category and position information respectively, and different loss functions are selected according to different tasks.

3. The UAV vehicle detection method according to claim 2, characterized in that: The SRConv module combines the SPD layer with the RFACon layer, specifically: The SPD layer converts the spatial dimension of the input feature map into the channel dimension while ensuring that the information within the channel is preserved; After the SPD layer, the RFAConv layer is used to further extract fine-grained small target features from the feature map output by the SPD layer. The receptive field attention mechanism is introduced to solve the sharing problem of convolution kernel parameters. By considering the global importance of each feature within the receptive field, the network's attention to targets of different scales is adaptively adjusted.

4. The method for detecting unmanned aerial vehicle (UAV) according to claim 1, wherein: The backbone network also introduces a fast spatial pyramid pooling module to pool the feature map to a fixed size, and then further enhances the feature extraction capability through the C2PSA module, which sets the pyramid slice attention mechanism.

5. The UAV vehicle detection method according to claim 2, characterized in that: The dynamic detection head includes scale-aware attention, space-aware attention and task-aware attention; Scale-aware attention: It is deployed only on the level L dimension, by scaling the feature pyramid to the same scale to form a 3D tensor, which is then used as the input of the dynamic detection head; Spatial Perception Attention: Deployed in the spatial S dimension, it focuses on the distinctive areas in the image by aggregating features; Task-aware attention: Deployed in the channel C dimension, feature channels are dynamically opened or closed according to the response of different convolution kernels to objects to support different tasks, including classification, bounding box regression and center point learning.

6. The method for detecting unmanned aerial vehicle (UAV) according to claim 1, wherein: Train the improved YOLOv11n-SDEM drone vehicle target detection network, use the MPDIoU loss function for the bounding box loss, and use the EMA-SlideLoss loss function for the classification loss.

7. The method for detecting unmanned aerial vehicle (UAV) according to claim 2, wherein: The C3k2-DBB module integrates the DBB diversified branch block, which increases complexity and improves feature representation during training, and is converted to a single branch during inference to maintain efficiency. After training, it is equivalent to being converted to a single convolutional layer for deployment.

8. The method for detecting a drone vehicle according to any one of claims 1 to 7, wherein: The UAV road vehicle dataset DRC Dataset is constructed and divided into a training set and a validation set. Specifically: Use a drone to hover at high altitude to shoot videos of road traffic scenes, then perform frame segmentation on the videos, build an automated labeling tool chain, label the vehicle targets in the images, and divide the labeled images into training and validation sets for model training and verification.

9. The method for detecting a drone vehicle according to claim 8, wherein: The training process is as follows: the training set is input into the improved YOLOv11n-SDEM UAV vehicle target detection network for training until the loss function converges. At the same time, the precision, recall rate and average precision of the validation set are used as indicators to evaluate the model performance and obtain the detection model.

10. A system for implementing the method for detecting unmanned aerial vehicle targets according to any one of claims 1 to 9, characterized in that: include: Dataset processing module: used to obtain and annotate the UAV road vehicle dataset and divide it into training set, validation set and test set; Model building module: used to build an improved YOLOv11n-SDEM drone vehicle target detection network; Model training module: Used to train the improved YOLOv11n-SDEM UAV vehicle target detection network using the DRC Dataset training set according to the set model network training parameters, and output the training weight file after the training is completed; the training weight file is verified through the validation set, and the training weight file with the highest accuracy is selected as the optimal weight file to obtain the trained improved YOLOv11 UAV vehicle target detection model; Target detection module: Input the road vehicle video frames or images taken by drones and the optimal weight files into the trained YOLOv11n-SDEM drone vehicle target detection network to perform road vehicle target detection and output the vehicle target detection results.

Images