A multi-rotor unmanned aerial vehicle target detection method in a complex scene

Abstract

The application discloses a multi-rotor unmanned aerial vehicle target detection method in a complex scene, and relates to the technical field of computer vision and target detection. First, open-source multi-rotor unmanned aerial vehicle image data is used to construct a MUAV target training data set; second, a MUAV target detection network in a complex environment is constructed. Based on YOLOv11n, a residual attention module RAM is introduced into the YOLOv11n backbone network, and then the shallow fine-grained features of the MUAV target are strengthened. A learnable linear transformation gate unit layer is introduced into the neck network, the channel adaptive weighting mechanism of the GU layer is used to dynamically adjust the contribution of the shallow fine-grained features and the deep semantic features, and the semantic gap between the fused feature maps is effectively reduced. The improved YOLOv11n is called Im-YOLOv11n. Finally, the network performance is verified by using a test data set. The detection accuracy and other quantitative indicators of the Im-YOLOv11n are better than those of the models in the R-CNN series and the YOLO series. The Im-YOLOv11n has more stable detection performance in complex scenes such as insufficient light, complex background and occlusion.

Description

Technical Field

[0001] This invention relates to the field of computer vision and target detection technology, and in particular to a method for target detection of multi-rotor UAVs in complex scenarios. Background Technology

[0002] In the field of low-altitude security defense, real-time and accurate detection of multirotor unmanned aerial vehicle (MUAV) targets is a key technology for building a monitoring system. Because MUAV targets are typically extremely small in physical size and have weak texture and structural information, they often occupy very few pixels in the image during actual detection, making them highly susceptible to interference from complex cloud layers, building backgrounds, and varying lighting conditions. Existing target detection frameworks, especially one-stage detection algorithms such as YOLOv11n, while offering significant advantages in detection speed and multi-scale feature fusion, still have technical limitations when applied to MUAV targets in complex scenarios.

[0003] On the one hand, deep neural networks involve multiple downsampling operations during feature extraction. This reduces the effective feature information in the generated feature maps, and spatial information is easily compressed or even lost when passed to deeper layers, resulting in insufficient ability to capture shallow, fine-grained features. On the other hand, in the multi-scale feature fusion stage, YOLO networks typically use simple cascading or splicing methods to integrate shallow spatial information from the backbone network with deep semantic information from the neck network. However, due to significant differences between semantic features at different levels, direct fusion without a filtering mechanism can easily generate semantic gaps and even introduce redundant background noise, leading to unstable model learning gradients or detection errors. Therefore, enhancing the network's ability to express shallow key features of MUAV targets and optimizing the semantic inconsistency problem during cross-level feature fusion have become crucial issues that urgently need to be addressed to improve the accuracy of MUAV target detection in complex scenarios. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a target detection method for multi-rotor unmanned aerial vehicles in complex scenarios. The MUAV detection method obtained by the present invention has better performance, higher detection accuracy and robustness.

[0005] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0006] A target detection method for multi-rotor UAVs in complex scenarios, proposed according to the present invention, includes:

[0007] S1. Collect image data of multi-rotor unmanned aerial vehicles (MUAVs), and label and segment the MUAV image data to obtain labeled and segmented MUAV image data; construct a model training dataset for target detection of multi-rotor unmanned aerial vehicles based on the labeled and segmented MUAV image data. The model training dataset includes a training set and a validation set.

[0008] S2. Construct an improved YOLOv11n network, denoted as the Im-YOLOv11n network. The Im-YOLOv11n network includes a backbone network, a neck network, and a head network. The backbone network incorporates a residual attention module (RAM), and the neck network incorporates two linear transform gate unit (GU) layers.

[0009] The backbone network is used to extract shallow features and deep semantic information of MUAV from MUAV image data, and output the shallow features and deep semantic information of MUAV to the neck network;

[0010] The neck network is used to perform feature fusion processing on shallow features and deep semantic information of MUAV, generate multi-scale fused features and output them to the head network;

[0011] The head network is used to classify and predict the multi-scale fusion features output by the neck network and perform bounding box regression, outputting the category, location information and confidence of the MUAV target;

[0012] S3. Train the Im-YOLOv11n network based on the training set to obtain the trained Im-YOLOv11n network; optimize the trained Im-YOLOv11n network using the validation set to obtain the optimized Im-YOLOv11n network; use the optimized Im-YOLOv11n network to perform target detection on MUAV image data and output the target's category and location information.

[0013] As a further optimization scheme for the multi-rotor UAV target detection method in complex scenarios described in this invention, the experimental dataset also includes a test set, and the detection performance of the optimized Im-YOLOv11n network is evaluated based on the test set.

[0014] As a further optimization of the multi-rotor UAV target detection method in complex scenarios described in this invention, a head network is used to perform classification prediction and bounding box regression on the multi-scale fusion features output by the neck network by decoupling the detection heads.

[0015] As a further optimization of the multi-rotor UAV target detection method in complex scenarios described in this invention, the Im-YOLOv11n network is based on the YOLOv11n network. The second Conv layer, the third C3K2 layer, the fourth Conv layer, and the fifth C3K2 layer in the YOLOv11n backbone network are replaced with the first to fourth RAMs with attention mechanisms in sequence; and two GU layers are introduced in the feature fusion path of the neck network.

[0016] This is a further optimization scheme for the multi-rotor UAV target detection method in complex scenarios described in this invention.

[0017] The backbone network consists of a convolutional Conv layer, four cascaded RAMs, a fast spatial pyramid pooling layer (SPPF), and a local cross-stage layer (C2PSA). The RAM is composed of a convolutional layer, a C3K2 module, and a CAM submodule cascaded together.

[0018] The RAM module is composed of a front-end residual structure and a rear-end CAM sub-module. The front-end residual structure includes a convolutional layer and a C3K2 module. In the RAM module, the feature map output from the previous layer of the backbone network is added element-wise to the feature map processed by the convolutional layer in the RAM module through skip connections to obtain a new feature map. The new feature map is then input to the CAM sub-module for dimension reweighting.

[0019] As a further optimization of the multi-rotor UAV target detection method in complex scenarios described in this invention, the backbone network, neck network, and head network comprise a total of twenty-two layers.

[0020] The backbone network has seven layers, consisting of layers one through seven connected sequentially; among them,

[0021] The first layer is a convolutional layer, the second layer is the first RAM, the third layer is the second RAM, the fourth layer is the third RAM, the fifth layer is the fourth RAM, the sixth layer is the SPPF layer, and the seventh layer is the C2PSA layer.

[0022] The neck network has 14 layers, including layers 8 to 21 connected sequentially. Among them, layer 8 is the first upsampling layer, layer 9 is the first concat layer, layer 10 is the first C3K2 layer, layer 11 is the second upsampling layer, layer 12 is the second concat layer, layer 13 is the second C3K2 layer, layer 14 is the first downsampling layer, layer 15 is the third concat layer, layer 16 is the first GU layer, layer 17 is the third C3K2 layer, layer 18 is the second downsampling layer, layer 19 is the fourth concat layer, layer 20 is the second GU layer, and layer 21 is the fourth C3K2 layer. Layer 7 is connected to layers 8 and 19 respectively, layer 4 is connected to layer 9, layer 3 is connected to layer 12, and layer 10 is connected to layer 15.

[0023] Header network: Layer 22 is the decoupled detection head, and Layer 22 is connected to Layer 13, Layer 17 and Layer 21 respectively.

[0024] As a further optimization of the multi-rotor UAV target detection method in complex scenarios described in this invention, the CAM submodule includes a channel attention component and a spatial attention component, wherein...

[0025] Channel Attention Component: The input feature map for the channel attention component is the feature map output from the previous layers of the backbone network. Global max pooling and average pooling are performed on the input feature map to obtain the first channel feature vector and the second channel feature vector. The first channel feature vector is sequentially input into a cascaded convolutional structure consisting of the first and second convolutional layers to obtain the first weight vector. The second channel feature vector is sequentially input into the cascaded convolutional structure consisting of the first and second convolutional layers to obtain the second weight vector. The first weight vector and the second weight vector are added element-wise, and the result is subjected to Sigmoid activation to obtain the channel attention weight vector. The channel attention weight vector is multiplied by the input feature map to obtain the channel attention feature map.

[0026] Spatial attention component: Average pooling and max pooling are performed on the channel attention feature maps along the channel dimension to obtain two feature maps of scale 1×W×H, where W and H represent the width and height of the feature maps, respectively; the two two-dimensional feature maps are concatenated along the channel dimension to obtain a concatenated feature map; the concatenated feature map is then subjected to Sigmoid activation to obtain a spatial attention weight map; the spatial attention weight map is then multiplied element-wise with the channel attention feature map to obtain the output of the CAM submodule.

[0027] Compared with the prior art, the present invention, employing the above technical solution, has the following technical effects:

[0028] (1) To address the problem that shallow fine-grained features of MUAV targets are easily lost in the existing technology, RAM is introduced into the backbone network to perform adaptive weighting processing on the shallow fine-grained feature map, thereby enhancing the ability to extract shallow fine-grained features of multi-rotor UAV targets.

[0029] (2) To address the semantic inconsistency problem in the fusion process of shallow fine-grained features and deep semantic features, a GU layer is introduced into the neck network to dynamically adjust the weights of shallow fine-grained features and deep semantic features, thereby reducing the semantic gap between shallow fine-grained features and deep semantic features and improving the fusion quality of shallow fine-grained features and deep semantic features.

[0030] (3) Through the above structural improvements of introducing RAM into the backbone network and GU layer into the neck network, the network can maintain stable feature expression ability in scenarios such as complex background, lighting changes and target occlusion, thereby improving the robustness of detection results;

[0031] (4) This invention improves the accuracy and stability of target detection for multi-rotor UAVs without significantly increasing computational complexity, while taking into account both detection performance and real-time requirements. Attached Figure Description

[0032] Figure 1 This is a block diagram of the Im-YOLOv11n network architecture proposed in this invention;

[0033] Figure 2 This is a schematic diagram of the residual attention module (RAM) structure;

[0034] Figure 3 This is a schematic diagram of the internal structure of the Convolutional Attention Module (CAM).

[0035] Figure 4 The images show the visual performance of YOLO architectures in MUAV detection on the DUT-Anti-UAV dataset under different lighting conditions. Among them, (a) shows the visual performance of the Im-YOLOv11n network in detecting MUAV targets, (b) shows the visual performance of the YOLOv11n network in detecting MUAV targets, (c) shows the visual performance of the YOLOv8n network in detecting MUAV targets, and (d) shows the visual performance of the YOLOv5n network in detecting MUAV targets.

[0036] Figure 5Visual representations of MUAV detection using YOLO architectures on the DUT-Anti-UAV dataset under conditions of dense targets, complex backgrounds, and occluded targets; where (a) is the visual representation of MUAV target detection using the Im-YOLOv11n network, (b) is the visual representation of MUAV target detection using the YOLOv11n network, (c) is the visual representation of MUAV target detection using the YOLOv8n network, and (d) is the visual representation of MUAV target detection using the YOLOv5n network. Detailed Implementation

[0037] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0038] A target detection method for multi-rotor UAVs in complex scenarios includes the following steps:

[0039] S1. Collect image data of multi-rotor unmanned aerial vehicles (MUAVs), and label and segment the MUAV image data to obtain labeled and segmented MUAV image data; construct a model training dataset for target detection of multi-rotor unmanned aerial vehicles based on the labeled and segmented MUAV image data. The model training dataset includes a training set and a validation set.

[0040] S2. Construct an improved YOLOv11n network, denoted as the Im-YOLOv11n network. The Im-YOLOv11n network includes a backbone network, a neck network, and a head network. The backbone network incorporates a residual attention module (RAM), and the neck network incorporates two linear transform gate unit (GU) layers.

[0041] The backbone network is used to extract shallow features and deep semantic information of MUAV from MUAV image data, and output the shallow features and deep semantic information of MUAV to the neck network;

[0042] The neck network is used to perform feature fusion processing on shallow features and deep semantic information of MUAV, generate multi-scale fused features and output them to the head network;

[0043] The head network is used to classify and predict the multi-scale fusion features output by the neck network and perform bounding box regression, outputting the category, location information and confidence of the MUAV target;

[0044] S3. Train the Im-YOLOv11n network based on the training set to obtain the trained Im-YOLOv11n network; optimize the trained Im-YOLOv11n network using the validation set to obtain the optimized Im-YOLOv11n network; use the optimized Im-YOLOv11n network to perform target detection on MUAV image data and output the target's category and location information.

[0045] The training dataset method for the multi-rotor UAV target detection model in step S1 is as follows: The DUT-Anti-UAV public dataset and data collected in complex scenes are used, and the image data input to the network is uniformly adjusted to 640×640×3. The existing data covers various lighting conditions (bright, dark, night), target scales (large, medium, small), and occlusion situations. The experimental platform for training the Im-YOLOv11n network is a Windows operating system, configured with a high-performance GPU (such as an NVIDIA GeForce RTX 4060 Ti), and uses the deep learning framework PyTorch.

[0046] In step S2, an Im-YOLOv11n network is constructed. RAM is introduced into the YOLOv11n backbone network architecture, replacing the original combination of convolutional layers and C3K2 layers. RAM is used to enhance the backbone network's ability to extract shallow features from MUAV targets. A GU layer is introduced along the feature fusion path between the backbone and neck networks. The linear convolutional operation of the GU layer adaptively adjusts the feature weights, generating fused features with stronger semantic consistency and better performance.

[0047] The Im-YOLOv11n network constructed in this invention comprises a backbone network, a neck network, and a head network, and its specific hierarchical structure is as follows:

[0048] (a) Backbone network hierarchical structure: The backbone network has 7 layers. Its main task is to extract multi-scale features through downsampling and to enhance the ability to capture MUAV targets using RAM.

[0049] The first layer is a convolutional layer Conv 1, which is used to perform convolutional feature extraction processing on the input MUAV image data, extract the initial edge and texture information, and downsample the image resolution through a convolution operation with a stride of 2, outputting the first feature map to the second layer. The resolution of the first feature map is 320×320 and the number of channels is 16.

[0050] The second layer is the first RAM layer, RAM 1. This layer is used to further extract and downsample the first feature map, and to enhance the features through residual connections and attention mechanisms, outputting the second feature map to the third layer. The resolution of the second feature map is 160×160, and the number of channels is 32.

[0051] The third layer is the second RAM layer, RAM 2. This layer is used to extract and downsample the second feature map, and further enhance the shallow fine-grained feature representation of the target, outputting the third feature map to the fourth layer. The resolution of the third feature map is 80×80, and the number of channels is 64.

[0052] The fourth layer is the third RAM layer, RAM 3. This layer is used to extract features and downsample the third feature map, outputting the fourth feature map to the fifth layer. The resolution of the fourth feature map is 40×40, and the number of channels is 128.

[0053] The fifth layer is the fourth RAM layer, RAM 4. It is used to perform deep feature extraction and downsampling on the fourth feature map, and outputs the fifth feature map to the sixth layer. The fifth feature map is a deep semantic feature map with a resolution of 20×20 and 256 channels.

[0054] The sixth layer is SPPF. It is used to perform multi-scale pooling operations on the fifth feature map, obtain contextual information under different receptive fields and fuse them, and output the sixth feature map to the seventh layer. The resolution of the sixth feature map is 20×20 and the number of channels is 256.

[0055] The seventh layer is C2PSA, which is used to perform spatial attention enhancement processing on the sixth feature map to improve the response capability of key regions and output the seventh feature map to the neck network. The seventh feature map is the enhanced deep semantic feature map with a resolution of 20×20 and 256 channels.

[0056] (ii) Neck network hierarchical structure: The neck network consists of 14 layers, which achieve deep decoupling and fusion of shallow spatial information and deep semantic information through a bidirectional fusion path from top to bottom and from bottom to top.

[0057] The eighth layer is the first upsampling layer, Upsample 1, which is used to upsample the seventh feature map and output the eighth feature map to the ninth layer. The resolution of the eighth feature map is 40×40 and the number of channels is 256.

[0058] The ninth layer is the first concatenation layer Concat 1, which is used to concatenate the eighth feature map with the fourth feature map through channels, and output the ninth feature map to the tenth layer. The ninth feature map has 384 channels and a resolution of 40×40.

[0059] The tenth layer is the first C3K2 layer C3K2 A, which is used to extract features and compress channels from the ninth feature map, and output the tenth feature map to the eleventh layer. The tenth feature map is a fused feature map with a resolution of 40×40 and 128 channels.

[0060] The eleventh layer is the second upsampling layer, Upsample 2, which is used to upsample the tenth feature map and output the eleventh feature map to the twelfth layer. The resolution of the eleventh feature map is 80×80 and the number of channels is 128.

[0061] The twelfth layer is the second Concat layer, Concat 2, which is used to concatenate the eleventh feature map with the third feature map through channels, and output the twelfth feature map to the thirteenth layer. The twelfth feature map has 192 channels and a resolution of 80×80.

[0062] The thirteenth layer is the second C3K2 layer, C3K2 B, which is used to extract features and compress channels from the twelfth feature map and output the thirteenth feature map to the head network. The thirteenth feature map is a fine-grained feature map with a resolution of 80×80 and 64 channels.

[0063] The fourteenth layer is the first downsampling layer Conv 2, which is used to downsample the thirteenth feature map and output the fourteenth feature map to the fifteenth layer. The resolution of the fourteenth feature map is 40×40 and the number of channels is 64.

[0064] The fifteenth layer is the third Concat layer, Concat 3, which is used to concatenate the fourteenth and tenth feature maps through channels, and output the fifteenth feature map to the sixteenth layer. The fifteenth feature map is a cross-scale fusion feature map with 192 channels and a resolution of 40×40.

[0065] The sixteenth layer is the first GU layer, GU 1, which is used to perform channel weighting and feature adaptive adjustment processing on the fifteenth feature map, and output the sixteenth feature map to the seventeenth layer. The sixteenth feature map is a weighted fusion feature map with 192 channels and a resolution of 40×40.

[0066] The seventeenth layer is the third C3K2 layer (C3K2 C), which is used to extract features and compress channels from the sixteenth feature map, and output the seventeenth feature map to the head network. The resolution of the seventeenth feature map is 40×40, and the number of channels is 128.

[0067] The eighteenth layer is the second downsampling layer Conv 3, which is used to downsample the seventeenth feature map and output the eighteenth feature map to the nineteenth layer. The resolution of the eighteenth feature map is 20×20 and the number of channels is 128.

[0068] The nineteenth layer is the fourth concat layer (Concat 4). This layer is used to concatenate the eighteenth feature map with the seventh feature map through channels, outputting the nineteenth feature map up to the twentieth layer. The nineteenth feature map has 384 channels and a resolution of 20×20.

[0069] The twentieth layer is the second GU layer, GU 2, which is used to perform channel weighting and semantic filtering on the nineteenth feature map and output the twentieth feature map to the twenty-first layer. The twentieth feature map is a high-order semantic feature map with 384 channels and a resolution of 20×20.

[0070] The 21st layer is the fourth C3K2 layer, C3K2 D, which is used to extract features and compress channels from the 20th feature map and output the 21st feature map to the head network. The resolution of the 21st feature map is 20×20 and the number of channels is 256.

[0071] (III) Head Network Hierarchical Structure: The twenty-second layer is the decoupled detection head Detect, which is used to receive the thirteenth feature map, the seventeenth feature map and the twenty-first feature map, which correspond to the target feature representation at different scales. The decoupled detection head performs classification prediction and bounding box regression respectively, and finally outputs the category information, local coordinates and confidence of the MUAV target.

[0072] The Im-YOLOv11n network is used to detect MUAV targets, and its performance is validated using a test dataset. Non-maximum suppression is applied to the observed data to eliminate overlapping detection boxes, and the final detection results are output. Precision, recall, and average precision (AP) are used to quantitatively evaluate and analyze the detection results of the Im-YOLOv11n network.

[0073] In practice, this invention is divided into a training phase and a testing phase.

[0074] The first step in training Im-YOLOv11n is to construct the training dataset for the Im-YOLOv11n network. The DUT-Anti-UAV public dataset is used as the basic data source. This dataset contains 10,109 high-resolution images, covering 35 different types of MUAV targets, with samples covering various lighting conditions, target scales, occlusion situations, and complex backgrounds. The dataset is divided into a training set (approximately 5,243 images), a validation set (approximately 2,621 images), and a test set (approximately 2,245 images). Before inputting the images into the network, the pixel dimensions of the input images are uniformly adjusted to 640×640×3. The experimental environment for training the Im-YOLOv11n network is a Windows operating system configured with a high-performance GPU (such as an NVIDIA GeForce RTX 4060 Ti with 8GB of VRAM). The software environment includes Python 3.12 and CUDA 12.6, and the deep learning framework PyTorch 2.6.0 is used for network construction and training.

[0075] The next step is to construct the Im-YOLOv11n network. For example... Figure 1As shown, the Im-YOLOv11n network mainly consists of three parts: a backbone network, a neck network, and a head network. This invention focuses on reconstructing the backbone network and optimizing the feature fusion path. In the backbone network, a convolutional layer, four cascaded RAMs, an SPPF layer, and a C2PSA layer are sequentially arranged. The RAM structure is as follows... Figure 2 As shown, the RAM consists of a convolutional layer, a C3K2 module, and a CAM submodule cascaded together, with a SkipConnection path. This path directly traverses the "Conv layer + C3K2 layer" structure with the RAM's input feature map, and adds it element-wise to the output of the C3K2 layer. The RAM utilizes the CAM submodule to simultaneously focus on key features in both the channel and spatial domains, enhancing the backbone network's ability to represent shallow features of MUAV targets. The specific structure of the CAM submodule is as follows... Figure 3 As shown, the algorithm includes channel attention and spatial attention components. In the channel attention component, the input feature map is subjected to global max pooling and average pooling, resulting in two one-dimensional vectors. These two vectors are forward-propagated to a cascaded convolutional structure. The first 1×1 convolutional layer compresses the number of channels from C to C / 16 to learn global channel dependencies. A second 1×1 convolutional layer then remaps the number of channels back to C to generate importance response weights. In the spatial attention component, the channel-weighted feature map is subjected to average pooling and max pooling along the channel dimension. After concatenation, it is processed by a 7×7 convolutional layer with a stride of 1 and padding of 3, and then activated by a sigmoid function to generate spatial weights.

[0076] Channel attention component: Input feature map of the channel attention component The input feature map is derived from the output of the previous layer in the backbone network; global max pooling and average pooling are performed on the input feature map respectively to obtain the first channel feature vector. Second channel feature vector First channel feature vector The inputs are sequentially fed into a cascaded convolutional structure consisting of the first and second convolutional layers to obtain the first weight vector and the second channel feature vector. The first and second convolutional layers are sequentially input into a cascaded convolutional structure to obtain the second weight vector. The first and second weight vectors are then added element-wise, and the result is subjected to sigmoid activation to obtain the channel attention weight vector. The channel attention feature map is obtained by multiplying the channel attention weight vector with the input feature map. ;

[0077] Spatial attention component: Channel attention feature map Average pooling and max pooling are performed along the channel dimension to obtain two feature maps with a scale of 1×W×H. and Where W and H represent the width and height of the feature map, respectively; the two two-dimensional feature maps and The splicing is performed along the channel dimension to obtain the spliced ​​feature map. The concatenated feature map is passed through a 7×7 convolutional layer and then activated by Sigmoid to obtain the spatial attention weight map. Spatial attention weight map With channel attention feature map Perform element-wise multiplication to obtain the output of the CAM submodule. .

[0078] In the construction of the neck network, this invention achieves deep fusion of multi-scale features through a bidirectional fusion path of top-down and bottom-up. First, the deep semantic information output from the backbone network is progressively passed through upsampling layers and fused with the corresponding shallow spatial details in the backbone network via concatenation. To overcome the semantic differences caused by direct concatenation of features from different levels, GU1 and GU2 layers are introduced at layers 16 and 20, respectively. The GU layer utilizes a learnable weight mechanism built from its internal 1×1 convolutional layers to dynamically weight the concatenated feature maps, reducing the semantic gap between deep and shallow features and suppressing redundant information caused by excessively large receptive field spans. The features processed by the GU layer then enter the C3K2 module for further feature extraction, thereby improving the model's stability in detecting small MUAV targets in complex scenes.

[0079] During the training phase of Im-YOLOv11n, the training hyperparameters are first set. The batch size is set to 16, the initial learning rate is set to 0.001, and a learning rate decay mechanism and an early stopping strategy (patience value set to 50) are used to prevent overfitting. The constructed Im-YOLOv11n network is then iteratively trained using the training set.

[0080] This implementation example verifies the effectiveness of the Im-YOLOv11n method through experiments. Regarding evaluation metrics, this invention uses precision (P), recall (R), and average precision (AP) as core performance evaluation indicators. Specifically, P and R are used to evaluate the model's accuracy and capture capability for targets, respectively, while AP@0.5 serves as a global indicator comprehensively measuring the model's average precision performance at different recall levels.

[0081] In the comparative analysis of different detection methods, the experiment selected two-stage architecture networks such as Faster R-CNN, Mask R-CNN, and Cascade R-CNN, and one-stage architecture networks such as YOLOv5n, YOLOv8n, and YOLOv11n as benchmarks. The quantitative evaluation results are shown in Table 1. The analysis results show that Im-YOLOv11n has the highest values ​​in all evaluation indicators, with an AP@0.5 value of 0.954. Compared with the two-stage algorithm that relies on the candidate region generation mechanism, this invention, as a one-stage detection framework, effectively avoids information loss caused by candidate box screening when processing targets with extremely small pixel proportions through end-to-end prediction, significantly reducing the false negative rate, especially showing stronger robustness under extreme conditions such as low light. Compared with the evolution of the YOLO series, Im-YOLOv11n achieves a 2.8 percentage point improvement in average accuracy (AP@0.5) compared to YOLOv11n. This demonstrates that the proposed reconstruction of the backbone network to enhance shallow fine-grained feature representation and the optimization of feature fusion path by combining dynamic weighting mechanism are effective ways to improve the accuracy of MUAV target detection in complex scenarios.

[0082] Table 1. Quantitative evaluation of MUAV detection results of Im-YOLOv11n and the comparison network.

[0083]

[0084] To more intuitively demonstrate the detection performance of the proposed Im-YOLOv11n, the MUAV target detection visual results of Im-YOLOv11n, YOLOv8n, and YOLOv5n were compared, as shown in the figure. Figure 4 and Figure 5 As shown.

[0085] Figure 4 These are visual representations of the YOLO architecture's MUAV detection performance on the DUT-Anti-UAV dataset under different lighting conditions; among them, Figure 4 (a), (b), (c), and (d) in the figure correspond to the visual results of the Im-YOLOv11n network, YOLOv11n network, YOLOv8n network, and YOLOv5n network in detecting MUAV targets, respectively. Figure 4 Visualized detection results for each network under different lighting conditions are presented. Figure 4It can be seen that under good lighting conditions, all networks can accurately identify MUAV targets, but Im-YOLOv11n's confidence level is significantly higher than the other three network architectures. When the lighting is insufficient, the detection confidence levels of all networks decrease, while Im-YOLOv11n's confidence level decreases less. When the lighting conditions further deteriorate to darkness, YOLOv8n and YOLOv5n exhibit significant false positives, YOLOv11n remains relatively stable, and Im-YOLOv11n still maintains the highest detection confidence level.

[0086] Figure 5 Visual representations of YOLO architectures performing MUAV detection on the DUT-Anti-UAV dataset under conditions of dense targets, complex backgrounds, and occluded targets; among them, Figure 5 (a), (b), (c), and (d) in the figure correspond to the visual results of the Im-YOLOv11n network, YOLOv11n network, YOLOv8n network, and YOLOv5n network in detecting MUAV targets, respectively. Figure 4 Visualized detection results for each network under different lighting conditions are presented. Figure 5 As can be seen, in the aforementioned complex scenarios, Im-YOLOv11n still achieved higher detection confidence and fewer false negatives and missed detections. Figure 4 , Figure 5 The confidence level corresponding to the blue detection box is difficult to clearly identify due to display size limitations. Therefore, in Figure 4 , 5 The corresponding confidence levels were manually labeled. All labeled values ​​were directly read from the model output and are strictly consistent with the original detection results. There was no human modification or estimation; they were used only to enhance the visualization effect.

[0087] To further objectively evaluate the specific contributions of each module in the proposed improved network Im-YOLOv11n to performance enhancement, ablation experiments were conducted, and the results are shown in Table 2. Analysis of Table 2 shows that after introducing RAM (Integrated Residual Structure and Convolutional Attention Mechanism) into the basic YOLOv11n network, the average precision improved from 0.926 to 0.935. This verifies that RAM can effectively enhance the network's ability to extract detailed features of small targets. Simultaneously, introducing the GU layer into the fusion path of the neck network also slightly improved the detection performance. This is because GU reduces the semantic inconsistency between shallow and deep semantic features, improves the stability of deep semantic representations, and thus enhances the network's target detection performance. When RAM and GU layers work synergistically, the MUAV target detection performance reaches its optimal level. This is because RAM provides richer shallow features and key information, while GU reduces the semantic gap between shallow and deep abstract features. The combined advantages of both significantly improve the network's detection performance for MUAV targets under various complex conditions.

[0088] Table 2 Comparison of Ablation Experiment Performance Results

[0089]

[0090] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A target detection method for multi-rotor UAVs in complex scenarios, characterized in that, include: S1. Collect image data of multi-rotor unmanned aerial vehicles (MUAVs), and label and classify the MUAV image data to obtain labeled and classified MUAV image data; A model training dataset for target detection of multi-rotor UAVs is constructed based on the labeled and segmented MUAV image data. The model training dataset includes a training set and a validation set. S2. Construct an improved YOLOv11n network, denoted as the Im-YOLOv11n network. The Im-YOLOv11n network includes a backbone network, a neck network, and a head network. The backbone network incorporates a residual attention module (RAM), and the neck network incorporates two linear transform gate unit (GU) layers. The backbone network is used to extract shallow features and deep semantic information of MUAV from MUAV image data, and output the shallow features and deep semantic information of MUAV to the neck network; The neck network is used to perform feature fusion processing on shallow features and deep semantic information of MUAV, generate multi-scale fused features and output them to the head network; The head network is used to classify and predict the multi-scale fusion features output by the neck network and perform bounding box regression, outputting the category, location information and confidence of the MUAV target; S3. Train the Im-YOLOv11n network based on the training set to obtain the trained Im-YOLOv11n network; optimize the trained Im-YOLOv11n network using the validation set to obtain the optimized Im-YOLOv11n network; use the optimized Im-YOLOv11n network to perform target detection on MUAV image data and output the target's category and location information.

2. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The experimental dataset also includes a test set, which is used to evaluate the detection performance of the optimized Im-YOLOv11n network.

3. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The head network is used to fuse multi-scale features output by the neck network, and classification prediction and bounding box regression are performed separately by decoupling the detection heads.

4. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The im-YOLOv11n network is based on the YOLOv11n network. The second Conv layer, the third C3K2 layer, the fourth Conv layer, and the fifth C3K2 layer in the YOLOv11n backbone network are replaced with the first to fourth RAMs with attention mechanisms in sequence; and two GU layers are introduced in the feature fusion path of the neck network.

5. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The backbone network consists of a convolutional Conv layer, four cascaded RAMs, a fast spatial pyramid pooling layer (SPPF), and a local cross-stage layer (C2PSA). The RAM is composed of a convolutional layer, a C3K2 module, and a CAM submodule cascaded together. The RAM module is composed of a front-end residual structure and a back-end CAM sub-module. The front-end residual structure includes a convolutional layer and a C3K2 module. In the RAM module, the feature map output from the previous layer of the backbone network is added element-wise to the feature map processed by the convolutional layer in the RAM module through skip connections to obtain a new feature map. The new feature map is then input into the CAM submodule for dimension reweighting.

6. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The backbone network, neck network, and head network have a total of twenty-two layers, among which... The backbone network has seven layers, consisting of layers one through seven connected sequentially; among them, The first layer is a convolutional layer, the second layer is the first RAM, the third layer is the second RAM, the fourth layer is the third RAM, the fifth layer is the fourth RAM, the sixth layer is the SPPF layer, and the seventh layer is the C2PSA layer. The neck network has 14 layers, including layers 8 to 21 connected sequentially. Among them, layer 8 is the first upsampling layer, layer 9 is the first concat layer, layer 10 is the first C3K2 layer, layer 11 is the second upsampling layer, layer 12 is the second concat layer, layer 13 is the second C3K2 layer, layer 14 is the first downsampling layer, layer 15 is the third concat layer, layer 16 is the first GU layer, layer 17 is the third C3K2 layer, layer 18 is the second downsampling layer, layer 19 is the fourth concat layer, layer 20 is the second GU layer, and layer 21 is the fourth C3K2 layer. Layer 7 is connected to layers 8 and 19 respectively, layer 4 is connected to layer 9, layer 3 is connected to layer 12, and layer 10 is connected to layer 15. Header network: Layer 22 is the decoupled detection head, and Layer 22 is connected to Layer 13, Layer 17 and Layer 21 respectively.

7. The target detection method for multi-rotor UAVs in complex scenarios according to claim 1, characterized in that, The CAM submodule includes a channel attention component and a spatial attention component, wherein, Channel Attention Component: The input feature map for the channel attention component is the feature map output from the previous layers of the backbone network. Global max pooling and average pooling are performed on the input feature map to obtain the first channel feature vector and the second channel feature vector. The first channel feature vector is sequentially input into a cascaded convolutional structure consisting of the first and second convolutional layers to obtain the first weight vector. The second channel feature vector is sequentially input into the cascaded convolutional structure consisting of the first and second convolutional layers to obtain the second weight vector. The first weight vector and the second weight vector are added element-wise, and the result is subjected to Sigmoid activation to obtain the channel attention weight vector. The channel attention weight vector is multiplied by the input feature map to obtain the channel attention feature map. Spatial attention component: Average pooling and max pooling are performed on the channel attention feature maps along the channel dimension to obtain two feature maps of scale 1×W×H, where W and H represent the width and height of the feature maps, respectively; the two two-dimensional feature maps are concatenated along the channel dimension to obtain a concatenated feature map; the concatenated feature map is then subjected to Sigmoid activation to obtain a spatial attention weight map; the spatial attention weight map is then multiplied element-wise with the channel attention feature map to obtain the output of the CAM submodule.

Images