Unmanned aerial vehicle target detection method based on improved YOLOv11n

Abstract

The invention belongs to the technical field of computer vision and target detection. In particular to an unmanned aerial vehicle target detection method based on improved YOLOv11n. The method comprises the steps of collecting a small unmanned aerial vehicle image data set under different heights, angles and weather conditions, and then preprocessing the data set to form a target data set; the method comprises the following steps: firstly, preprocessing a data set, then performing improvement by taking YOLOv11n as a basic algorithm, constructing an improved YOLOv11n detection model, then training the improved YOLOv11n model by using the preprocessed data set, and storing a weight file after training is completed; and finally, inputting a to-be-detected aerial unmanned aerial vehicle image into the trained and improved YOLOv11n model for detection, and outputting the position and category information of the unmanned aerial vehicle. According to the method, an SFPC module, an SPN structure and a DECG lightweight sharing detail enhancement detection head are constructed and introduced, and a loss function in an original model is replaced by an Inner-EIoU function. The method has high accuracy in a complex background and a long-distance target detection task, and meanwhile, light weight and real-time performance are kept.

Description

Technical Field

[0001] This invention relates to the field of deep learning technology, specifically to a UAV target detection method based on an improved YOLOv11n. Background Technology

[0002] Airspace monitoring systems first utilize computer vision systems to detect unmanned aerial vehicle (UAV) targets, and then guide monitoring equipment or interception systems to perform operations based on the detection results. Therefore, UAV detection is a crucial link in the entire airspace management process, and its accuracy and speed directly affect the efficiency of airspace security. However, aerial UAV targets are typically small in size, fly at high altitudes, and are easily obstructed. This presents certain challenges for UAV identification. Rapidly and accurately identifying aerial UAVs in complex environments such as urban areas and mountainous forests is an urgent issue in the development of airspace safety monitoring systems.

[0003] In the field of low-altitude safety protection, traditional methods rely primarily on manual monitoring, specifically on security personnel visually observing surveillance footage, using binoculars, or conducting ground patrols to screen for drones in the air. These methods have significant limitations: firstly, the efficiency of manual observation is limited by personnel's energy and visual range, making 24-hour continuous monitoring difficult, and blind spots are easily created when dealing with large airspaces; secondly, the accuracy of manual judgment is greatly affected by environmental factors and subjective experience, easily leading to misjudgments or omissions of distant, small-sized drone targets, failing to meet the requirements of modern airspace management for real-time performance, coverage, and identification accuracy.

[0004] With the advancement of artificial intelligence and machine learning technologies, deep learning-based drone detection technology has emerged. This technology provides a more efficient and accurate means of drone detection, enabling real-time analysis of image data without disrupting normal airspace operations and resisting a certain degree of environmental interference. This is crucial for improving airspace monitoring efficiency and ensuring low-altitude safety. Through intelligent detection systems, security personnel can better understand drone flight trajectories, thereby minimizing the safety risks posed by unauthorized drone flights and promoting more refined and intelligent airspace management. However, existing technologies still have shortcomings in terms of adaptability to complex environments, real-time detection performance, and recognition accuracy. Therefore, we propose an improved drone target detection method based on YOLO11n to address these issues. Summary of the Invention

[0005] (a) Technical problems to be solved

[0006] To address the problems existing in the prior art, this invention provides a UAV target detection method based on an improved YOLOv11n. This solves the problems mentioned in the background section.

[0007] (II) Technical Solution

[0008] The UAV target detection method of the present invention includes the following steps:

[0009] Step S1: Acquire drone target images, preprocess the drone images, and obtain drone image datasets;

[0010] Step S2: Construct an improved UAV detection model based on YOLOv11n; specific improvements include:

[0011] S2.1: Construct an SFPC module to replace the original YOLOv11n SPPF structure, adopt a multi-branch convolutional structure, and configure different dilation rates for each branch;

[0012] S2.2: An SPN feature fusion structure was designed to enhance edge details and semantic expression;

[0013] S2.3: A lightweight shared detail enhancement detection head (DECG) was designed, which combines GroupNormalization with a shared parameter mechanism.

[0014] S2.4: Use Inner-EIoU as the bounding box regression loss function.

[0015] Step S3: Train the improved YOLOv11n model using the preprocessed image set;

[0016] Step S4: Use the trained model to detect drone images in the air and evaluate the model.

[0017] Furthermore, in step S1 of the present invention, the preprocessing specifically involves: labeling the UAV images and dividing the UAV images into a training set, a validation set, and a test set in a ratio of 7:2:1.

[0018] Further, in step S2.1 of the present invention, the SFPC module specifically involves: dividing the UAV image feature map into a multi-branch feature stream, inputting the multi-branch feature streams into convolutional layers with different dilation rates for convolution operations, integrating the multi-branch feature streams after the convolution operation through a 1×1 convolution channel, and outputting the integrated result; designing convolution kernels with different dilation rates, and each dilation rate convolution path is constructed based on a single shared convolutional layer, with low dilation rate convolution focusing on capturing local details of the UAV, and high dilation rate convolution expanding the receptive field to incorporate global contextual information.

[0019] Furthermore, in step S2.2 of the present invention, the SPN feature fusion structure is specifically as follows: the structure uses the SPD-Conv module to perform fine-grained modeling of shallow P2 features, and at the same time integrates CSP and RFB-s mechanisms to enhance the expressive power of features.

[0020] Furthermore, in step S2.3 of this invention, the DECG module mainly consists of two cascaded 3×3 convolutional layers and GN (Group Normalization), which act on the multi-scale feature map in the form of shared parameters, thereby improving edge modeling capabilities and reducing computational overhead.

[0021] Furthermore, in step S2.4 of the present invention, Inner-EIoU is specifically defined as shown in formula (1).

[0022] (1)

[0023] In equation (1), and Let b represent the width and height of the minimum bounding rectangles of the anchor box and the target box, respectively. gt 'b' and 'a' represent the target box and anchor box, respectively. , ) represents the center coordinates of the target bounding box and its inner auxiliary bounding box, w gt and h gt These represent the width and height of the target bounding box, respectively, while the width and height of the anchor bounding box are denoted as w and h, respectively.

[0024] Furthermore, in step S3 of this invention, the improved YOLOv11n model is iteratively trained using the training set. During training, the model's performance changes are monitored in real time using the validation set, and a suitable optimizer is used to update the model parameters. Simultaneously, a set anchor box intersection-over-union ratio (IoU) is used as a standard to guide the model's learning of UAV target bounding boxes, ensuring that the model can effectively learn the features of small aerial UAV targets during training, until the model's performance on the validation set stabilizes, thus completing the model training process.

[0025] Furthermore, in step S4 of this invention, the evaluation specifically includes metrics such as precision, recall, average precision (AP), mean average precision (mAP), number of parameters, billion floating-point operations per second (GF1OPS), and frame rate (FPS). The specific calculation formulas are as follows:

[0026] (2)

[0027] (3)

[0028] (4)

[0029] (5)

[0030] (6)

[0031] (7)

[0032] (8)

[0033] Where TP is the number of images in which the improved YOLOv11n model correctly detected targets, FP is the number of images in which the improved YOLOv11n model incorrectly detected non-drone targets, FN is the number of images in which the improved YOLOv11n model failed to detect drone targets, Precision is the accuracy, Recall is the recall, t is the IoU (Intersection over Union) threshold, and T is the detection time for a single image.

[0034] (III) Beneficial Effects

[0035] Compared with existing technologies, this invention provides a UAV target detection method based on an improved YOLOv11n, which has the following beneficial effects:

[0036] 1. This invention replaces the original SPPF structure with an enhanced SFPC module in the YOLO11n backbone network. By dynamically adjusting the receptive field size through multi-dilation rate shared convolution and decomposing local and global features, the detection capability of target UAVs is enhanced, and the accuracy of the model in capturing low-altitude small UAVs is improved;

[0037] 2. In this invention, a fine-grained feature fusion network SPN is designed and introduced into the neck network of YOLOv11n. This network retains rich edge details and positional information of the target, effectively highlighting the characteristics of the UAV target, while suppressing background noise interference, thus further improving the detection accuracy of UAVs in occluded and complex scenes.

[0038] 3. By combining the DECG detection head, the model can more accurately recover the feature edges and positioning information of UAV targets, thereby improving the robustness and accuracy of detection;

[0039] 4. This invention uses Inner-EIoU as the bounding box regression loss function to enhance the target localization effect while maintaining regression accuracy and alleviate the prediction degradation problem in occluded scenes. Attached Figure Description

[0040] Figure 1This is an overall flowchart of a UAV target detection method based on an improved YOLOv11n according to the present invention;

[0041] Figure 2 This is a structural diagram of an improved model for a UAV target detection method based on the improved YOLOv11n of the present invention.

[0042] Figure 3 This is a structural diagram of the SFPC module of an unmanned aerial vehicle target detection method based on an improved YOLOv11n according to the present invention.

[0043] Figure 4 This is a structural diagram of the DECG module of an unmanned aerial vehicle (UAV) target detection method based on an improved YOLOv11n according to the present invention.

[0044] Figure 5 This is a schematic diagram of the SPN principle of an unmanned aerial vehicle target detection method based on an improved YOLOv11n according to the present invention.

[0045] Figure 6 The image shows the detection results of a UAV target detection method based on an improved YOLOv11n according to the present invention. Detailed Implementation

[0046] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0047] The UAV image detection method described in this embodiment has the following process: Figure 1 The steps shown are as follows:

[0048] Step S1: Acquire drone images and preprocess them to obtain a drone image dataset. Specifically, the drone images are labeled using the DUT Anti-UAV public dataset and self-collected drone images. The category label is "uav". The training set, validation set and test set are divided in a ratio of 7:2:1. The input image resolution is uniformly adjusted to 640×640 to adapt to the model input requirements.

[0049] Step S2, constructing an improved YOLOv11n UAV detection model, specifically:

[0050] The improved model structure is as follows: Figure 2As shown, the improved model uses the YOLO11n model as the base model. The YOLO11n model includes a backbone network, a neck network, and a head network. The SPPF structure in the YOLO11n backbone network is replaced with an SFPC module; the feature fusion stage in the YOLO11n neck network is introduced into SPN; the original detection head is replaced with DECG; and Inner-EIoU is used as the bounding box regression loss function.

[0051] The SFPC module described above has the following structure: Figure 3 As shown, it is divided into three parts: The first part is a multi-branch convolutional layer with three parallel convolutional branches and dilation rates of 1, 3, and 5, respectively. The low dilation rate branch captures local details of the drone, while the high dilation rate branch incorporates the global context. The second part is a parameter-sharing layer, which uses a single shared convolutional kernel as the basis for each branch to reduce the number of parameters. The third part is a feature fusion layer, which performs channel compression and fusion on the multi-branch outputs through 1×1 convolution. The specific formula for calculating the 3×3 dilated convolution is as follows:

[0052] (1)

[0053] in, The output of the 3×3 dilated convolution operation is X, where X is the input feature map. Here is the weight matrix used for the 3×3 dilated convolution operation, and d is the dilation rate of the 3×3 dilated convolution. The image shows the characteristic response of the UAV under different void ratios.

[0054] (2)

[0055] in, This is an intermediate feature map concatenated along the channel dimension from the output features of multi-scale dilated convolution, used to fuse UAV feature responses from different receptive fields. For feature splicing module, , , These are the results of three groups after multi-scale dilated convolution. It is a 1×1 convolution;

[0056] (3)

[0057] in, This is the final multi-scale fused feature map.

[0058] The low-dilation convolution kernel is specifically:

[0059] (4)

[0060] in, To use a 3×3 convolution operation with a dilation rate of 1, where d represents different dilation coefficients, Let X be the local detail feature component of X, where X is the input feature map. The local feature response image of the UAV extracted using low dilation rate represents the modeling result of the UAV structural information in input X at the current scale;

[0061] The aforementioned medium dilation rate convolution kernel is specifically:

[0062] (5)

[0063] in, This is the UAV mesoscale feature enhancement map extracted after a 3×3 convolution operation with a dilation rate of 3. It is a 3×3 dilated convolution. Let X be the mesoscale feature component of X, where X is the input feature map. The feature components of the UAV extracted by the medium dilation rate are used to capture the context information of the UAV and its surrounding environment at medium range.

[0064] The high dilation rate convolution kernel is specifically:

[0065] (6)

[0066] in, This is a large-scale feature enhancement map of the UAV obtained after a 3×3 convolution operation with a dilation rate of 5. It is a 3×3 dilated convolution. Let X be the large-scale feature component of X, where X is the input feature map. The feature components of the drone extracted with high dilation rate are used to capture the global features of drones at long distances and with slight occlusion.

[0067] The SPDConv modeling of shallow P2 features is as follows:

[0068] (7)

[0069] in, for The enhanced feature map obtained after SPDConv fusion For use of space to depth transformation processing , X represents the shallow P2 feature map output by the YOLO11n backbone network, and X represents the input feature map. It is a 3×3 non-stride convolution. This is a shallow enhancement feature map.

[0070] The SFPC module, as described above, addresses the shortcomings of existing SPPF modules, which primarily employ max pooling to extract global features. While this approach can integrate multi-scale information to some extent, the fixed pooling kernel and lack of parameter sharing make it difficult to adapt to the complex variations in UAV images across different airspace environments. Furthermore, SPPF-extracted features suffer from detail loss and strong background interference, leading to technical bottlenecks such as missed target detections in subsequent detection tasks. Therefore, this implementation improves upon the SPPF module by designing the SFPC module, specifically as follows:

[0071] 1) Multi-branch convolution design: Three parallel convolution branches are set up with dilation rates of 1, 3, and 5, respectively, for different drone distance scenarios: the d=1 branch captures local details of the drone at close range, the d=3 branch captures the drone and its surrounding environment at medium range, and the d=5 branch captures global features of the drone at a long distance with slight occlusion. The convolution method is shown in the following formula:

[0072] (8)

[0073] in, To use a 3×3 multi-scale dilated convolution operation with a dilation rate of d, where d represents different dilation coefficients, The representation of the structural information modeling result for input X at the current scale is shown for UAV feature response images extracted using different void ratios. This operation can expand the receptive field without increasing the number of parameters, enhancing the perception of UAV targets at different scales.

[0074] 2) Parameter sharing mechanism: A single shared convolutional kernel is used as the basic convolutional kernel for the three branches. Different scale feature extraction is achieved by adjusting the dilation rate, avoiding the need to design convolutional kernels for each branch, greatly reducing the number of parameters, and maintaining the consistency of feature extraction in each branch.

[0075] 3) Feature fusion optimization: Channel compression is performed on multi-branch spliced ​​features through 1×1 convolution, reducing the number of channels to the original number of input channels to avoid a surge in feature dimensions. At the same time, effective features are strengthened through convolution weight learning to suppress background interference.

[0076] The structure of the DECG detection head is as follows: Figure 4 As shown, the detection head consists of two cascaded 3×3 convolutional layers and a GN, which act on multi-scale feature maps in the form of shared parameters. While maintaining lightweight design, it enhances the ability to model edge structures and texture details. Finally, it combines the Inner-EIoU loss function to optimize bounding box regression and improve target localization accuracy. The specific definition of Inner-EIoU is shown in formula (9).

[0077] (9)

[0078] In equation (9), and Let b represent the width and height of the minimum bounding rectangles of the anchor box and the target box, respectively. gt 'b' and 'a' represent the target box and anchor box, respectively. , ) represents the center coordinates of the target bounding box and its inner auxiliary bounding box, w gt and h gt These represent the width and height of the target bounding box, respectively, while the width and height of the anchor bounding box are denoted as w and h, respectively.

[0079] The SPN structure described is as follows Figure 5 As shown, the SPD-Conv module is used to perform fine-grained modeling of shallow P2 features to enhance their ability to perceive targets. Simultaneously, the CSP and RFB-s mechanisms are integrated to enhance the expressive power of the features. SPDConv employs spatial-to-depth transformation and non-stretch convolution to achieve shallow feature enhancement. First, from a given input feature map of size C×H×W... The SpaceToDepth operation transforms the feature map into a size of 4C×H / 2×W / 2. Then, use 3×3 non-stride convolution pairs The process yields an enhanced feature map of size C×H / 2×W / 2. .

[0080] Step S3: Divide the labeled UAV image set into training, validation, and test sets according to a set ratio. Iteratively train the improved YOLOv11n model using the training set. During training, monitor model performance changes in real time using the validation set and update model parameters with an appropriate optimizer. Simultaneously, using a set anchor box intersection-over-union ratio (IoU) as a standard, guide the model's learning of UAV target bounding boxes, ensuring the model effectively learns the features of small aerial UAV targets during training, until the model's performance on the validation set stabilizes, completing the model training process.

[0081] Step S4: After training is complete, the improved YOLOv11n model is evaluated using a validation set and a test set. The test set is input into the trained improved model for detection, and the detection results are output as follows: Figure 6 As shown. The evaluation metrics used include precision, recall, average precision (AP), mean average precision (mAP), number of parameters, billion floating-point operations per second (GF1OPS), and frame rate (FPS), specifically:

[0082] (10)

[0083] (11)

[0084] (12)

[0085] (13)

[0086] (14)

[0087] (15)

[0088] (16)

[0089] Where TP is the number of images in which the improved YOLOv11n model correctly detected targets, FP is the number of images in which the improved YOLOv11n model incorrectly detected non-drone targets, FN is the number of images in which the improved YOLOv11n model failed to detect drone targets, Precision is the accuracy, Recall is the recall, t is the IoU (Intersection over Union) threshold, and T is the detection time for a single image.

[0090] The above provides a detailed description of the aerial UAV image detection method proposed in this invention. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. An unmanned aerial vehicle target detection method based on improved YOLOv11n, characterized in that, It comprises the following steps: Step S1: obtaining a UAV target image, preprocessing the UAV image, and obtaining a UAV image dataset; Step S2: constructing an improved UAV detection model based on YOLOv11n; the specific improvements include: S2.1: constructing an SFPC module to replace the original SPPF structure of YOLOv11n; S2.2: designing an SPN feature fusion structure for enhancing edge details and semantic expression; S2.3: designing a lightweight shared detail enhancement detection head DECG, which combines Group Normalization and a shared parameter mechanism; and S2.4: using Inner-EIoU as the boundary box regression loss function; Step S3: training the improved YOLOv11n model using the preprocessed image set; Step S4: detecting aerial UAV images using the trained model and evaluating the model.

2. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The preprocessing is labeling the UAV image and dividing it into a training set, a validation set, and a test set according to a ratio of 7:2:

1.

3. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The SFPC module specifically divides the UAV image feature map into multiple branch feature streams, inputs the multiple branch feature streams into convolution layers with different dilation rates for convolution operation, integrates the multiple branch feature streams after the convolution operation through 1x1 convolution channels, and outputs the integration result; different convolution kernels with different dilation rates are designed, and each dilation rate convolution path is constructed based on a single shared convolution layer, with low dilation rate convolution focusing on capturing local detail features of the UAV and high dilation rate convolution expanding the receptive field to include global context information.

4. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The SPN structure uses SPDConv to model the shallow P2 features in a fine-grained manner, while integrating the CSP and RFB-s mechanisms to enhance the feature expression capability.

5. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The lightweight shared detail enhancement detection head DECG module includes two 3x3 convolution layers and Group Normalization, which act on multi-scale feature maps in a shared parameter form, improve the edge modeling capability, and reduce the computational overhead.

6. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The loss function formula of the Inner-EIoU function is: （1） In formula (1), and respectively represent the width and height of the minimum enclosing rectangle of the anchor box and the target box, b gt and b respectively represent the target box and the anchor box. , ) are the center point coordinates of the target box and its internal auxiliary box, w gt and h gt respectively represent the width and height of the target box, and the width and height of the anchor box are respectively denoted as w and h.

7. The unmanned aerial vehicle target detection method based on improved YOLOv11n according to claim 1, characterized in that, The model training is iterative training of the improved YOLOv11n model using the training set, real-time monitoring of the performance change of the model through the validation set during the training process, updating of the model parameters using a suitable optimizer, and guiding the model to learn the UAV target boundary box based on the set anchor box intersection-over-union ratio to ensure that the model effectively learns the features of the aerial target UAV during the training process until the performance of the model on the validation set tends to be stable, completing the model training process.

8. The UAV target detection method based on the improved YOLOv11n according to claim 1, wherein the evaluation is specifically the precision (Precision), recall (Recall), average precision (AP), mean average precision (mAP), parameter quantity (Params), billion floating-point operations per second (GFlOPS), and frame rate (FPS) indicators, and the specific calculation formula is as follows: （2） （3） （4） （5） （6） （7） （8） wherein TP is the number of images in which the improved YOLOv11n model correctly detects the target, FP is the number of images in which the improved YOLOv11n model incorrectly detects a non-unmanned aerial vehicle target, FN is the number of images in which the improved YOLOv11n model does not detect an unmanned aerial vehicle target, Precision is the accuracy, Recall is the recall rate, t is an IoU (intersection over union) threshold, and T is the detection time of a single image.

Images