Unmanned aerial vehicle environment perception moving target detection method based on global attention mechanism

Abstract

The invention relates to an unmanned aerial vehicle environment perception moving target detection method based on a global attention mechanism. The method comprises the following steps: collecting a road condition environment image based on an unmanned aerial vehicle; inputting the road condition environment image into a pre-trained detection model, and identifying a moving target in the road condition environment image; the detection model is obtained by training a preset ConvGRU-YOLOv7-GA model on the basis of a preset data set. According to the invention, it can be ensured that the unmanned aerial vehicle can timely and accurately sense the moving target on the path in a complex dynamic environment, and powerful support is provided for vehicle safety and intelligent driving decision making.

Description

Technical Field

[0001] The present invention relates to the field of target detection technology, and in particular to a method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism. Background Art

[0002] With the development of control technology and communication technology, drones can complete more complex and critical tasks with their flexibility, portability and excellent perception capabilities. UAV-based target detection and tracking methods are widely used in fields such as material distribution and mountain rescue.

[0003] Mobile target detection is a crucial part of environmental perception technology. It is a branch of target detection in the field of computer vision. Unlike traditional static image target detection, mobile target detection needs to process dynamically changing targets in video streams, which increases the complexity of the time dimension. The accuracy, model size, and inference time of the mobile target detection model will directly determine the quality and efficiency of drone environmental perception.

[0004] Especially for drones, how to accurately and efficiently detect moving targets on the drone's path is extremely important for completing rescue missions. However, due to the limited computing power of drone terminal devices that cannot meet the actual application of engineering, the moving target detection algorithm must be lightweight enough to be deployed on the terminal device to achieve real-time moving target detection and recognition in drone environment perception.

[0005] Currently, mobile target detection in drone environment perception uses traditional two-stage deep learning target detection methods, such as RCNN and SPPNet. This method has low real-time detection performance, and the model is not lightweight enough to be directly deployed on the device, which increases the labor cost of subsequent pruning and distillation of the model.

[0006] A method for detecting a moving target of an unmanned aerial vehicle (UAV) based on an attention mechanism specifically comprises the following steps: step 1, inputting a real-time video stream and an image; step 2, a target detection algorithm in an anti-UAV system detects a UAV in a target detection area based on an attention mechanism; step 3, after the target detection algorithm based on the attention mechanism detects a UAV, the UAV tracking algorithm is activated, so that the UAV tracking algorithm tracks the UAV target detected by the detection algorithm in real time; step 4, when the UAV tracking algorithm loses tracking, the target detection algorithm based on the attention mechanism re-performs the detection process of the UAV in the target detection area, and when tracking is successful, outputs the relevant coordinates to its linked defense controller to perform locking and anti-UAV operations.

[0007] A method, device and drone system for detecting targets in drone aerial images, specifically comprising the following steps: step 1, performing multiple level-by-level dense connection feature extractions on the drone aerial images, respectively obtaining first, second, third and fourth level feature maps; step 2, performing feature fusion of a specific combination of the first, second, third and fourth level feature maps through a progressive feature pyramid, respectively obtaining first, second, third and fourth level fused feature maps; step 3, generating target initial frames for the first, second, third and fourth level fused feature maps respectively through a clustering parameter, detecting them in turn and placing the detection results into the drone aerial images, thereby obtaining small target information in the drone aerial images.

[0008] The prior art has the following defects:

[0009] 1. The traditional two-stage target detection algorithm (such as RCNN and SPPNet) is not real-time enough and cannot meet the needs of real-time detection of drones. In addition, the model is too large and not lightweight enough to be directly deployed on the limited hardware resources of the drone backend.

[0010] 2. The existing technology does not fully consider the correlation of moving targets in the time dimension and ignores the timing-related information between the current frame and the previous frame, making it difficult to effectively model the target motion trajectory.

[0011] 3. The existing technology is not adaptable enough to complex backgrounds. Background interference will reduce the accuracy of mobile target detection, and the generalization ability needs to be improved. Summary of the invention

[0012] The purpose of the present invention is to propose a method for detecting mobile targets in drone environment perception based on a global attention mechanism to solve the problems existing in the above-mentioned prior art. The present invention focuses on a method for detecting mobile targets in drone environment perception based on a global attention mechanism. By adding a model suitable for time series analysis to the drone environment perception target detection backbone network, it can make full use of time information and better learn the information of the movement of mobile targets in each frame of the video stream. The introduction of the global attention mechanism allows the model to pay more attention to the mobile targets in the picture and reduce the impact of complex backgrounds on the model. While improving the accuracy of mobile target detection, it not only improves the accuracy and generalization ability of mobile target detection, but also can effectively meet the needs of drones under limited terminal computing resources, providing an efficient and reliable mobile target detection solution for it.

[0013] To achieve the above object, the present invention provides the following solutions:

[0014] A method for detecting moving targets in UAV environment perception based on global attention mechanism, comprising:

[0015] Collect road environment images based on drones;

[0016] The road environment image is input into the detection model to identify the moving target in the road environment image; the detection model is obtained based on the training of a preset data set, the data set includes moving targets in different environmental scenarios, and the detection model is constructed based on the ConvGRU-YOLOv7-GA model.

[0017] Optionally, the moving targets include: moving pedestrians, vehicles, and cyclists.

[0018] Optionally, before training the preset ConvGRU-YOLOv7-GA model based on the preset data set, the following steps are included:

[0019] The preset data set is subjected to data preprocessing, including image scaling and image enhancement.

[0020] Optionally, the ConvGRU-YOLOv7-GA model is constructed by fusing a YOLOv7 sub-model, a ConvGRU module and a global attention module;

[0021] The YOLOv7 sub-model is used for moving target detection;

[0022] The ConvGRU module is used to obtain the spatial pattern and temporal dynamic characteristics of the moving target;

[0023] The global attention module is used to capture the dependency between any two positions in the input moving target feature map sequence, and capture the global structure of the input sequence by calculating the correlation between each position and all other positions.

[0024] Optionally, the backbone network of the YOLOv7 sub-model adopts a CSPDarknet53 network; the CSPDarknet53 architecture includes: a Darknet53 sub-network and a CSP sub-network;

[0025] The Darknet53 subnetwork includes: input layer, convolution layer, residual block, pooling layer, and output layer;

[0026] The input layer of the Darknet53 subnetwork is used to receive a fixed-size moving target image input;

[0027] The convolutional layer of the Darknet53 subnetwork is used to extract low-level features in moving images;

[0028] The residual block of the Darknet53 subnetwork is used to maintain information transmission;

[0029] The pooling layer of the Darknet53 subnetwork is used to reduce the dimension of the feature map;

[0030] The output layer of the Darknet53 subnetwork is used to output the mobile target recognition result;

[0031] The CSP sub-network is constructed based on a convolutional layer;

[0032] The convolutional layer in the CSP subnetwork is connected in parallel with the convolutional layer in the Darknet53 subnetwork, so as to perform deep feature extraction on the features output by the convolutional layer in the Darknet53 subnetwork;

[0033] The input of the CSP sub-network and the features output by the pooling layer of the Darknet53 sub-network are summed up and used as the input of the output layer of the Darknet53 sub-network.

[0034] Optionally, the fusion of the YOLOv7 sub-model, the ConvGRU module and the global attention module includes:

[0035] Embed the ConvGRU module in the CSPDarknet53 network and build the ConvGRU-YOLOv7 sub-model;

[0036] A global attention module is added to the ConvGRU-YOLOv7 model to construct the ConvGRU-YOLOv7-GA model.

[0037] Optionally, a ConvGRU module is added between the 3rd convolutional layer and the 4th convolutional layer in the CSPDarknet53 network;

[0038] The global attention module is added after the last residual block of the CSPDarknet53 network.

[0039] Optionally, the expression of the global attention module is:

[0040]

[0041]

[0042] Among them, F in Represents the input feature map, F mid It is the intermediate feature map generated by the global attention mechanism in processing the feature map. out Represents the final output feature map. channel represents the channel attention mechanism, M spatial Represents the spatial attention mechanism.

[0043] Optionally, training a preset ConvGRU-YOLOv7-GA model based on a preset data set includes:

[0044] The preset data set is input into the ConvGRU-YOLOv7-GA model, and the parameters of the ConvGRU-YOLOv7-GA model are fine-tuned end-to-end using the back propagation algorithm and optimizer.

[0045] The beneficial effects of the present invention are:

[0046] The present invention adopts the lightweight YOLOv7 as the basic model, and integrates the ConvGRU and GA modules to better utilize the timing and attention information, so the accuracy of mobile target detection is significantly improved. YOLOv7 has good lightweight performance, ConvGRU can model in the time dimension, and GA can focus on the target area. The combination of the three makes the model's detection effect on mobile targets much better than the simple lightweight model. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0048] Figure 1 A schematic diagram of a flow chart of a method for detecting a moving target in an unmanned aerial vehicle environment based on a global attention mechanism according to an embodiment of the present invention;

[0049] Figure 2 It is an overall block diagram of a method for detecting a moving target in an unmanned aerial vehicle environment based on a global attention mechanism according to an embodiment of the present invention;

[0050] Figure 3 A schematic diagram of the E-ELAN structure of an embodiment of the present invention;

[0051] Figure 4 This is a diagram of the ConvGRU embedding backbone network structure of an embodiment of the present invention. Figure 5 Schematic diagram of the ConvGRU structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0054] This embodiment proposes a method for detecting moving targets in a drone environment based on a global attention mechanism, which specifically includes:

[0055] Collect road environment images based on drones;

[0056] The road environment image is input into a pre-trained detection model to identify moving targets in the road environment image; the detection model is obtained by training a preset ConvGRU-YOLOv7-GA model based on a preset data set.

[0057] Specifically, in this embodiment, this embodiment proposes a method for detecting a moving target by UAV environment perception based on a global attention mechanism. The method for detecting a moving target by UAV environment perception based on a global attention mechanism is obtained by training a UAV environment perception moving target detection model using a training data set. The main steps and overall block diagram of the process are as follows: Figure 1 and Figure 2 shown.

[0058] Furthermore, the data set includes several different environmental scenes and the moving targets; wherein the moving targets include: moving pedestrians, vehicles, and cyclists.

[0059] Furthermore, before training the preset ConvGRU-YOLOv7-GA model based on the preset data set, the following steps are performed:

[0060] The preset data set is subjected to data preprocessing, including image scaling and image enhancement.

[0061] Specifically, in this embodiment, the public VSCrowd dataset is selected to train the ConvGRU-YOLOv7-GA mobile target detection model. The dataset contains 634 video sequences, of which 479 are used for model training and 137 are used for model verification, totaling 62,938 frames of images, with a resolution of 1920×1080 per frame. These video data cover various complex scenes, including moving pedestrians, vehicles, cyclists and other moving targets, providing rich data support for training models.

[0062] After data collection, data processing was also performed, including:

[0063] Image scaling: Scale the original 1920×1080 resolution image to 768×768 to reduce the input size and speed up model training and inference;

[0064] Data enhancement: The image is randomly rotated and vertically flipped with a flip probability of 0.5 to generate more samples with different angles and directions to increase data diversity.

[0065] Furthermore, the ConvGRU-YOLOv7-GA model is constructed by fusing the YOLOv7 sub-model, the ConvGRU module and the global attention module.

[0066] The YOLOv7 sub-model is responsible for implementing mobile target detection. The ConvGRU module is responsible for obtaining the spatial pattern and temporal dynamic characteristics of the mobile target. The global attention mechanism module captures the dependency between any two positions in the input mobile target feature map sequence. By calculating the correlation between each position and all other positions, the global structure of the input sequence is captured, so that the model can better understand the correlation of each frame of the input mobile target image.

[0067] Furthermore, the backbone network of the YOLOv7 sub-model adopts the CSPDarknet53 network; the CSPDarknet53 architecture includes: a Darknet53 sub-network and a CSP sub-network;

[0068] The Darknet53 subnetwork consists of an input layer, a convolutional layer, a residual block, a pooling layer, and an output layer. The input layer accepts a fixed-size moving target image input. The convolutional layer is responsible for extracting low-level features from the moving image. The residual block is used to maintain information transmission and solve the gradient vanishing problem in deep neural networks. The pooling layer is used to reduce the dimension of the feature map and reduce the amount of calculation. The output layer outputs the moving target recognition result. The CSP subnetwork consists of a convolutional layer, which is connected in parallel with the convolutional layer in the Darknet53 network. It is responsible for deep feature extraction of the features output by the convolutional layer in the Darknet53 network. The input of the CSP subnetwork is summed with the features output by the pooling layer of the Darknet53 subnetwork as the input of the Darknet53 subnetwork output layer.

[0069] The CSP network divides the processing of feature maps into two parallel paths, one part is calculated through convolution, and the other part is directly passed, and then the results of the two paths are merged.

[0070] Furthermore, the fusion of the YOLOv7 sub-model, the ConvGRU module and the global attention module includes:

[0071] Embed the ConvGRU module in the CSPDarknet53 network and build the ConvGRU-YOLOv7 sub-model;

[0072] A global attention module is added to the ConvGRU-YOLOv7 model to construct the ConvGRU-YOLOv7-GA model.

[0073] Furthermore, a ConvGRU module is added between the 3rd convolutional layer and the 4th convolutional layer in the CSPDarknet53 network; Figure 5 As shown, ConvGRU includes an update gate, a reset gate, a candidate memory gate, and three convolutional layers, wherein the output of the update gate is connected to the convolutional layer, the output of the reset gate is connected to the convolutional layer, and the output of the candidate memory gate is connected to the convolutional layer;

[0074] The global attention module is added after the last residual block of the CSPDarknet53 network.

[0075] Specifically, in this embodiment, based on the YOLOv7 real-time target detection algorithm, a ConvGRU-YOLOv7-GA model integrating ConvGRU timing modeling and global attention mechanism is constructed.

[0076] The backbone network of YOLOv7 adopts the CSPDarknet53 architecture, which is an improved convolutional neural network based on Darknet53. The CSP (Cross Stage Partial) module divides the processing of feature maps into two parallel paths, one part is calculated by convolution, and the other part is directly transmitted, and then the two results are merged. This branch structure design improves the efficiency of gradient information flow in the network, enhances the receptive field, and thus improves the detection performance. In order to capture the temporal correlation of targets in video sequences, the ConvGRU module is embedded in a specific layer of CSPDarknet53. ConvGRU is a combination of gated recurrent units and convolutional neural networks. It can well model temporal data such as video frames and use historical information to help predict moving targets in the current frame. In addition, the global attention (GA) mechanism is introduced at the end of the network, so that the model can adaptively allocate attention weights to different regions, capture the semantic information of the image as a whole, focus on the area where the moving target is located, and suppress the interference of complex background.

[0077] The model architecture balances accuracy and computational complexity, and ConvGRU and GA are added in different stages of feature extraction and target detection, giving full play to the advantages of both. Finally, the entire ConvGRU-YOLOv7-GA model was trained end-to-end using a large amount of labeled data, and the accuracy and other indicators were evaluated and continuously optimized on the validation set.

[0078] The model architecture balances accuracy and computational complexity, and ConvGRU and GA are added in different stages of feature extraction and target detection, giving full play to the advantages of both. Finally, the entire ConvGRU-YOLOv7-GA model was trained end-to-end using a large amount of labeled data, and the accuracy and other indicators were evaluated and continuously optimized on the validation set.

[0079] E-ELAN Module:

[0080] For the original YOLOv7 model, the E-ELAN module is used. The E-ELAN module uses the ideas of extended cardinality, shuffle cardinality, cross connection, and merge cardinality. This structural design can enhance the learning ability of the network without destroying the original gradient path. The E-ELAN module structure is as follows Figure 3 shown.

[0081] The YOLOv7 model uses CSPDarknet53 as its backbone network architecture. The E-ELAN module is an extension module in the YOLOv7 model, which is used to enhance the fusion and learning capabilities of features. The E-ELAN module is used in conjunction with CSPDarknet53 in YOLOv7, allowing the model to process and fuse features more efficiently when extracting features. CSPDarknet53 is responsible for providing deep feature extraction capabilities, while the E-ELAN module further optimizes the expression and fusion of these features, thereby improving the accuracy and detection effect of the model.

[0082] exist Figure 3 In order to further improve the receptive field of the E-ELAN module, the original 3×3 convolution kernel size was changed to 5×5.

[0083] ConvGRU module:

[0084] ConvGRU is embedded into the feature extraction stage of the YOLOv7 model, and the output of ConvGRU is used as the input of YOLOv7. Specifically, a ConvGRU layer is added between the 3rd and 4th layers in the YOLOv7 backbone network CSPDarknet53 network to capture the temporal information of the feature map. In this way, the model can use the information of the previous frames to better predict the target of the current frame. The ConvGRU embedded backbone network structure is as follows Figure 4shown.

[0085] The advantage of this method is that it can improve the accuracy of target detection by using the time series information processed by ConvGRU without changing the original architecture of YOLOv7. At the same time, since ConvGRU is an end-to-end model, the back propagation algorithm can be used directly to train the entire model without complex model fusion or feature selection. Combining ConvGRU and YOLOv7 can achieve continuous and stable detection of targets in the video.

[0086] Attention Module:

[0087] The global attention mechanism is introduced based on the built ConvGRU-YOLOv7 model. Specifically, a global attention module is added after the last residual block of CSPDarknet53. In this way, the global attention module can help the model focus on the important parts of the high-level features. The global attention module is implemented by the following formula:

[0088]

[0089]

[0090] Among them, F in Represents the input feature map, F mid It is the intermediate feature map generated when the global attention mechanism processes the feature map. out Represents the final output feature map. channel represents the channel attention mechanism, M spatial Represents the spatial attention mechanism.

[0091] Furthermore, training the preset ConvGRU-YOLOv7-GA model based on the preset data set includes:

[0092] The parameters of the ConvGRU-YOLOv7-GA model are fine-tuned end-to-end using the back-propagation algorithm and optimizer.

[0093] Specifically, in this embodiment,

[0094] The training process of the ConvGRU-YOLOv7-GA model is divided into the following steps:

[0095] Prepare a training data set. This example uses the public VSCrowd data set, which contains 634 video sequences and 62,938 frames of images, covering various complex scenes and multiple moving targets. The original images are preprocessed by scaling and data enhancement.

[0096] Build the ConvGRU-YOLOv7-GA network architecture:

[0097] 1) The backbone network uses the lightweight CSPDarknet53 to improve the receptive field and gradient propagation efficiency;

[0098] 2) Introduce the ConvGRU module at a specific position in the backbone network to capture the timing information of the video frame sequence;

[0099] 3). Add a global attention (GA) module at the end of the network to adaptively assign attention weights and focus on the moving target area.

[0100] Set training hyperparameters, such as learning rate strategy, optimizer, loss function, etc., and initialize the model.

[0101] Build a data loader, sample the training data in batches, and feed it into the model for training in a loop.

[0102] The parameters of the entire ConvGRU-YOLOv7-GA model are fine-tuned end-to-end using the back-propagation algorithm and optimizer.

[0103] Evaluate model performance on the validation set, including indicators such as accuracy, recall, and F1 score, and continuously optimize training strategies and hyperparameters to ultimately achieve satisfactory performance.

[0104] The advantages of this training process are that it uses an end-to-end training method, without the need for complex feature fusion, thus reducing engineering complexity. At the same time, the training process combines temporal modeling and attention mechanisms to balance the accuracy and speed of mobile target detection. In addition, it is also lightweight in design to ensure that the model can be efficiently deployed and inferred in a drone environment.

[0105] ConvGRU-YOLOv7-GA Model Deployment

[0106] After training, the ConvGRU-YOLOv7-GA model is deployed on the backend computing platform of the drone for online reasoning and moving target detection in real-time video streams.

[0107] Specifically, the model is compiled into highly optimized machine code and stored in the storage medium of the on-board computing platform. When the drone is driving, the on-board camera collects road condition video in real time, and the video stream is input into the deployed detection model for online reasoning. The model can quickly and accurately detect and identify moving targets such as pedestrians, vehicles, and cyclists in the video, providing key input for the vehicle's intelligent decision-making system. Since YOLOv7 itself has a good lightweight design, and the introduction of ConvGRU and GA modules improves the timing modeling and attention mechanism, the ConvGRU-YOLOv7-GA model can achieve efficient reasoning with limited hardware resources, meeting the real-time and mobility requirements of drone scenarios.

[0108] This model deployment solution ensures that the UAV can timely and accurately perceive moving targets on the path in complex dynamic environments, providing strong support for vehicle safety and intelligent driving decisions.

[0109] In this embodiment, a ConvGRU module is embedded between the 3rd and 4th layers in the YOLOv7 backbone network CSPDarknet53 architecture. The purpose of this innovation is to capture the dynamic change information of the target in the video sequence in time, and use the historical information of the previous frames to help predict the moving target in the current frame, thereby improving the detection accuracy. After the introduction of ConvGRU, the model can not only learn the appearance characteristics of the target, but also model the target's motion pattern in time, making it more robust for target tracking and detection.

[0110] The global attention (GA) module is introduced after the last residual block of CSPDarknet53. The purpose of this innovation is to enable the model to adaptively allocate attention weights to different regions, capture the semantic information of the image as a whole, focus on the region where the moving target is located, and suppress the interference of complex background. The GA module generates an attention weight map through learning, weights the important regions (i.e. the moving target region) and downweights the background region, which is conducive to extracting more effective and robust target feature representation.

[0111] End-to-end training combines the ConvGRU-YOLOv7-GA mobile target detection model with ConvGRU temporal modeling and global attention (GA) mechanism. This innovation seamlessly integrates the above two modules into the YOLOv7 architecture, building a detection model that has both temporal modeling capabilities and the ability to adaptively focus on target areas of interest. Compared with traditional two-stage detection algorithms, this model is trained and predicted in an end-to-end manner, reducing the engineering complexity of the system.

[0112] Compared with the simple lightweight YOLOv5 model, this implementation directly uses the lightweight YOLOv7 as the basic model, and integrates the ConvGRU and GA modules to better utilize the timing and attention information, so the accuracy of mobile target detection is significantly improved. YOLOv7 already has good lightweight performance, ConvGRU can model in the time dimension, and GA can focus on the target area. The combination of the three makes the model's detection effect on mobile targets much better than the simple lightweight model.

[0113] Compared with the two-stage detection method of moving targets, this embodiment adopts an end-to-end single-stage detection architecture, which has faster detection speed and stronger real-time performance, which is very suitable for drone autonomous driving scenarios with extremely high real-time requirements. The two-stage detection requires generating candidate areas first and then classifying them, which is cumbersome: the end-to-end method directly outputs the results in one step, which is simple and efficient.

[0114] By integrating time series modeling and attention mechanism, the model of this embodiment has better adaptability to complex dynamic backgrounds and stronger detection generalization ability. The motion trajectory of a moving target is often accompanied by complex background interference. ConvGRU can capture the target motion pattern, and GA can suppress background noise. The synergy of the two greatly improves the adaptability of the model to various environments.

[0115] Despite the introduction of ConvGRU and GA modules, thanks to the lightweight design of YOLOv7 itself, the computational complexity of this implementation model is still moderate, and it can be efficiently deployed and inferred under the limited computing resources of the drone, fully meeting the real-time and mobile application requirements.

[0116] The embodiments described above are only descriptions of the preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Without departing from the design spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should all fall within the protection scope determined by the claims of the present invention.

Claims

1. A method for detecting mobile targets in UAV environment perception based on global attention mechanism, characterized in that: include: Collect road environment images based on drones; Inputting the road environment image into a detection model to identify a moving target in the road environment image; The detection model is obtained by training based on a preset data set, wherein the data set includes moving targets in different environmental scenarios, and the detection model is constructed based on a ConvGRU-YOLOv7-GA model.

2. The method for detecting mobile targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 1 is characterized in that: The moving targets include: moving pedestrians, vehicles, and cyclists.

3. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 1 is characterized in that: Before training the preset ConvGRU-YOLOv7-GA model based on the preset dataset, the following steps are included: The preset data set is subjected to data preprocessing, including image scaling and image enhancement.

4. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 1 is characterized in that: The ConvGRU-YOLOv7-GA model is constructed by integrating the YOLOv7 sub-model, the ConvGRU module and the global attention module; The YOLOv7 sub-model is used for moving target detection; The ConvGRU module is used to obtain the spatial pattern and temporal dynamic characteristics of the moving target; The global attention module is used to capture the dependency between any two positions in the input moving target feature map sequence, and capture the global structure of the input sequence by calculating the correlation between each position and all other positions.

5. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 4 is characterized in that: The backbone network of the YOLOv7 sub-model adopts the CSPDarknet53 network; The CSPDarknet53 architecture includes: a Darknet53 subnetwork and a CSP subnetwork; The Darknet53 subnetwork includes: input layer, convolution layer, residual block, pooling layer, and output layer; The input layer of the Darknet53 subnetwork is used to receive a fixed-size moving target image input; The convolutional layer of the Darknet53 subnetwork is used to extract low-level features in moving images; The residual block of the Darknet53 subnetwork is used to maintain information transmission; The pooling layer of the Darknet53 subnetwork is used to reduce the dimension of the feature map; The output layer of the Darknet53 subnetwork is used to output the mobile target recognition result; The CSP sub-network is constructed based on a convolutional layer; The convolutional layer in the CSP subnetwork is connected in parallel with the convolutional layer in the Darknet53 subnetwork, so as to perform deep feature extraction on the features output by the convolutional layer in the Darknet53 subnetwork; The input of the CSP sub-network and the features output by the pooling layer of the Darknet53 sub-network are summed up and used as the input of the output layer of the Darknet53 sub-network.

6. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 5 is characterized in that: The fusion of the YOLOv7 sub-model, the ConvGRU module and the global attention module includes: Embed the ConvGRU module in the CSPDarknet53 network and build the ConvGRU-YOLOv7 sub-model; A global attention module is added to the ConvGRU-YOLOv7 model to construct the ConvGRU-YOLOv7-GA model.

7. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 5 is characterized in that: Adding a ConvGRU module between the 3rd convolutional layer and the 4th convolutional layer in the CSPDarknet53 network; The global attention module is added after the last residual block of the CSPDarknet53 network.

8. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 4 is characterized in that: The expression of the global attention module is: Among them, F in Represents the input feature map, F mid It is the intermediate feature map generated by the global attention mechanism in processing the feature map. out Represents the final output feature map, M channel represents the channel attention mechanism, M spatial Represents the spatial attention mechanism.

9. The method for detecting moving targets in an unmanned aerial vehicle environment based on a global attention mechanism according to claim 1, characterized in that: Training the preset ConvGRU-YOLOv7-GA model based on the preset dataset includes: The preset data set is input into the ConvGRU-YOLOv7-GA model, and the parameters of the ConvGRU-YOLOv7-GA model are fine-tuned end-to-end using the back propagation algorithm and optimizer.

Images