Unmanned aerial vehicle low-altitude visual angle target detection method and device based on deep learning

Abstract

The invention discloses an unmanned aerial vehicle low-altitude visual angle target detection method and device based on deep learning, and belongs to the technical field of computer vision. The method comprises the following steps: performing data enhancement on a collected target detection object image data set, and dividing the enhanced target detection object image data set into a training set, a verification set and a test set; an attention mechanism module is added in a backbone network of the YOLOV7-OBB model; the YOLOV7-OBB model is trained based on the training set, the YOLOV7-OBB model in the training process is evaluated based on the verification set, and the trained YOLOV7-OBB model is obtained; and inputting the test set into the trained YOLOV7-OBB model, and carrying out target detection to obtain a detection result. By enhancing a target detection object image data set and adding an attention mechanism module, the YOLOV7-OBB model fully learns key feature information of a target, and the detection precision and robustness are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and more specifically, to a method and device for detecting targets from a low-altitude perspective of an unmanned aerial vehicle based on deep learning. Background Art

[0002] With the rapid development and widespread application of drone technology, target detection from the perspective of drones at low altitudes has become an important industry application. However, target detection in low-altitude backgrounds faces many challenges, especially in the detection of few samples and rotating targets. Traditional target detection methods are difficult to effectively extract the feature information of targets in complex and changeable low-altitude environments, especially when there are few targets and there are rotation changes.

[0003] Traditional target detection methods rely on a large amount of sample data to extract target features, but in practical applications, considering the flight cost and environmental impact, obtaining enough low-altitude target sample data is a difficult task, which brings great difficulties to the training and optimization of target detection models. Therefore, how to increase the amount of data and thus improve the model effect has become an urgent problem to be solved. In the prior art, for the problem of few samples, the traditional method is to continue to amplify the original data through drone sampling. In recent years, it has become popular to increase the amount of training samples by rotating and flipping pictures. However, the traditional method of increasing drone sampling methods has high labor costs, and although image transformations such as rotation and flipping have certain effects, they still remain in the content of the picture itself, and the proportion of pictures with targets from a high-altitude perspective is small, resulting in limited improvement effects.

[0004] When a drone flies at a low altitude, the target often rotates. Traditional target detection methods usually use a rectangular bounding box aligned with the length and width axes of the image to frame the object. The target box cannot change with the rotation of the target, and thus cannot accurately reflect the actual shape and direction of the target. In the prior art, in recent years, a rotating target box detection method of YOLOV7-OBB has been proposed to address the problem of rotating target detection. This method achieves the prediction of the rotating box by adding a rotation angle parameter to the YOLOV7 network model. However, YOLOV7-OBB is prone to missed detection when there are many targets to be detected at a low altitude perspective.

[0005] After searching, the Chinese patent application, application number 202310136199.X, published on May 23, 2023, discloses a method for multi-pedestrian detection and posture estimation from the perspective of an ultra-low altitude drone. The method includes: using a drone to collect image data and preprocessing the data set; enhancing the data set; using the YOLOv5 algorithm to use the enhanced data to train the target detection network; using the trained target detection network to perform multi-pedestrian detection, and using a single-person posture estimation algorithm to perform posture estimation. This method can solve the problems of redundancy and inaccurate annotation of the data set, but the method does not take into account the situation where the target may rotate when the drone collects images, so it is not widely applicable. Summary of the invention

[0006] 1. Technical issues to be solved In view of the problem that traditional target detection methods in the prior art cannot effectively extract target feature information in complex low-altitude environments due to the small number of targets and the existence of rotational changes, the present invention provides a low-altitude perspective target detection method and device for unmanned aerial vehicles based on deep learning, which can enhance the target data set and improve the detection accuracy and robustness of rotating targets.

[0007] 2. Technical solution The purpose of the present invention is achieved through the following technical solutions.

[0008] A method for detecting targets from a low altitude perspective of a UAV based on deep learning, comprising the following steps: Use drones to collect target detection object image datasets; Performing data enhancement on the target detection object image dataset, and dividing the enhanced target detection object image dataset into a training set, a validation set, and a test set; Improve the YOLOV7-OBB model; add an attention mechanism module to the backbone network of the YOLOV7-OBB model; Train the YOLOV7-OBB model based on the training set, evaluate the YOLOV7-OBB model during training based on the validation set, and obtain the trained YOLOV7-OBB model; The test set is input into the trained YOLOV7-OBB model for target detection to obtain the target detection result.

[0009] As a further improvement of the present invention, the backbone network of the YOLOV7-OBB model includes a CBS module and an ELAN module; an attention mechanism module is connected after the CBS module and the ELAN module.

[0010] As a further improvement of the present invention, the attention mechanism module includes a coordinate information embedding module and a coordinate attention generation module.

[0011] As a further improvement of the present invention, in the coordinate information embedding module, The attention mechanism module receives the feature map output by the previous convolutional layer as input; The attention mechanism module performs global pooling operations on the feature map in the horizontal and vertical directions respectively; The horizontal pooling results and the vertical pooling results are concatenated along the channel dimension to obtain a coordinate information tensor.

[0012] As a further improvement of the present invention, in the coordinate attention generation module, The attention mechanism module compresses and activates the coordinate information tensor to obtain key feature representations; The attention mechanism module restores the original number of channels of the feature map through splitting and independent convolution operations, and generates the first attention weight map in the horizontal direction and the second attention weight map in the vertical direction; The first attention weight map and the second attention weight map are respectively normalized by the Sigmoid activation function to obtain a first output value and a second output value; The first output value is used as the attention coefficient in the horizontal direction, and the second output value is used as the attention coefficient in the vertical direction.

[0013] As a further improvement of the present invention, matrix multiplication operations are performed on the feature map with the attention coefficient in the horizontal direction and the attention coefficient in the vertical direction respectively to obtain a feature map with an added attention mechanism.

[0014] As a further improvement of the present invention, the data enhancement of the target detection object image data set includes an oversampling operation and a random pasting operation.

[0015] As a further improvement of the present invention, the specific steps of the oversampling operation are: setting the target detection object image as a copy sample; setting the target detection object image data set as a total sample; copying the copy sample several times; setting the sampling rate, and controlling the copy frequency of the copy sample and the proportion of the copy sample to the total sample through the sampling rate.

[0016] As a further improvement of the present invention, the specific steps of the random pasting operation are: The target detection object in the target detection object image is randomly copied to any position in the image to generate a new target detection object image and annotation information; Crop the target detection object in the target detection object image according to the annotation information; The cropped target object is processed and pasted into the new target object image to generate a label.

[0017] A low-altitude target detection device for unmanned aerial vehicles based on deep learning, comprising: The data acquisition module uses drones to collect target detection object image data sets; A data processing module performs data enhancement on a target detection object image data set, and divides the enhanced target detection object image data set into a training set, a validation set, and a test set; Model improvement module, improves the YOLOV7-OBB model; in particular, adds an attention mechanism module to the backbone network of the YOLOV7-OBB model; Model training module, trains the YOLOV7-OBB model based on the training set, evaluates the YOLOV7-OBB model during training based on the validation set, and obtains the trained YOLOV7-OBB model; The target detection module inputs the test set into the trained YOLOV7-OBB model for target detection and obtains the target detection result.

[0018] 3. Beneficial effects Compared with the prior art, the advantages of the present invention are: (1) The present invention provides a method and device for low-altitude target detection from a drone based on deep learning. The method performs data enhancement on a target detection image dataset collected from a drone at a low altitude. This can not only increase the number of image datasets, but also effectively increase the frequency of appearance of small targets and their relative size in the target detection image through methods such as multi-angle image transformation, oversampling operations, and random pasting operations. This ensures that the YOLOV7-OBB model can fully learn the feature information of small targets during training, significantly enhancing the comprehensiveness and accuracy of feature extraction.

[0019] (2) The present invention provides a method and device for low-altitude target detection from a drone based on deep learning. Based on the YOLOV7-OBB model, the backbone network structure is optimized and an attention mechanism module is added to the backbone network. The attention mechanism module is used to dynamically adjust the YOLOV7-OBB model's attention to different areas of the input target detection image, so that the YOLOV7-OBB model can focus more on key features during the detection process, effectively suppress the interference of irrelevant information, and effectively improve the accuracy and robustness of target detection.

[0020] (3) The present invention provides a method and device for detecting low-altitude target from a drone based on deep learning. By constructing a YOLOV7-OBB model, a good effect can be obtained for detecting low-altitude target from a drone using only a lightweight detector. Using a detector with a simpler structure in the process of detecting low-altitude target from a drone can greatly reduce the resources required for calculation, improve the detection speed of the YOLOV7-OBB model, and reduce the memory usage required in the detection process, thus avoiding unnecessary waste and effectively utilizing limited computing resources. This enables the practical application of the YOLOV7-OBB model in detecting low-altitude target from a drone, and has strong practicality and wide applicability. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a method according to an embodiment of the present invention; Figure 2 A schematic diagram of the structure of the YOLOV7-OBB model currently used in an embodiment of the present invention; Figure 3 A schematic diagram of the structure of the improved YOLOV7-OBB model according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the structure of the attention mechanism module according to an embodiment of the present invention; Figure 5 It is a schematic diagram of the CBS module structure of an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0023] Example like Figure 1 As shown, a method for low-altitude target detection from a drone based on deep learning is provided in this embodiment, and the method includes the following steps: using a drone to collect a target detection object image dataset; performing data enhancement on the target detection object image dataset, and dividing the enhanced target detection object image dataset into a training set, a validation set, and a test set; improving the YOLOV7-OBB model; wherein, an attention mechanism module is added to the backbone network of the YOLOV7-OBB model; the YOLOV7-OBB model is trained based on the training set, and the YOLOV7-OBB model during the training process is evaluated based on the validation set to obtain a trained YOLOV7-OBB model; the test set is input into the trained YOLOV7-OBB model for target detection to obtain a target detection result.

[0024] Specifically in this embodiment, a target detection object image data set is collected using a drone. It should be noted that the target detection object image data taken by a drone is less than the target detection object image data taken from a ground perspective, and the amount of data taken in some scenes is even less, such as car accidents, ships, etc., which leads to insufficient target detection object image data to support detection model training. Therefore, it is necessary to perform data enhancement on the target detection object image data set.

[0025] In this embodiment, data enhancement of the target detection object image data set includes image multi-angle transformation, oversampling operation and random pasting operation.

[0026] In this embodiment, the multi-angle transformation of the image used is the existing technology, including image geometric transformation, color space transformation and noise addition.

[0027] Specifically, image geometric transformations include flipping (horizontally flipping or vertically flipping to simulate observations from different perspectives), rotation (rotating at any angle to adapt to changes in the target detection object image in different directions), scaling (enlarging or reducing the image in proportion to simulate the presentation of the target detection object image at different distances), and cropping (capturing a specific area in the image to focus on the key information in the target detection object image).

[0028] Color space transformation focuses on adjusting the color attributes of the target detection object image, including color transformation (such as converting a color target detection object image into a grayscale target detection object image, or applying pseudo-color processing to the target detection object image to enhance the visual effect), brightness adjustment (simulating the target detection object image under different lighting conditions by increasing or decreasing the brightness of the target detection object image), contrast (improving the visual clarity of the target detection object image by enhancing or weakening the difference between the light and dark parts of the target detection object image) and saturation (adjusting the color saturation of the target detection object image by increasing or decreasing the vividness of the color of the target detection object image).

[0029] Adding noise, such as Gaussian noise and salt and pepper noise, to simulate the interference that the target detection object image may experience during the transmission or acquisition process, thereby improving the detection model's tolerance to noise; blur processing, such as mean blur and Gaussian blur, to simulate the blurring effect of the target detection image caused by motion or focal length change, helps the detection model maintain stable recognition capabilities in blurred target detection object images; sharpening processing, by enhancing the edges and details of the target detection object image, improves the clarity of the target detection object image, and helps the detection model capture more detailed features.

[0030] Therefore, in this embodiment, the target detection object image data set is preliminarily enhanced through multi-angle transformation of the image.

[0031] In this embodiment, the specific steps of the oversampling operation are: setting the target detection object image as the copy sample, and setting the target detection object image data set as the total sample. The copy sample is copied several times. In this embodiment, the name of the copied copy needs to be modified to ensure the uniqueness in the training set. In order to avoid excessive duplicate samples caused by oversampling, in this embodiment, the sampling rate rate is set, and the sampling rate rate is used to control the copy frequency of the copy sample and the proportion of the copy sample to the total sample, thereby helping the subsequent detection model to better learn the target features.

[0032] In this embodiment, the sampling rate is expressed as:

[0033] Among them, N a represents the replicated sample, N b In this embodiment, the sampling rate rate is less than 0.5.

[0034] In this embodiment, the specific steps of the random pasting operation are: randomly copying the target detection object in the target detection object image to any position in the image to generate a new target detection object image and annotation information. In this process, the target detection object in the target detection object image is first cut out according to the annotation information. The cut out target detection object is processed, including performing a series of random transformations (such as scaling, folding, rotating, etc.), and then the transformed target detection object is pasted into the new target detection object image to generate a corresponding label.

[0035] It is worth noting that in this embodiment, four methods for copy and paste target detection are provided, specifically including: (1) Single target-single image copy and paste: Select a target detection object in the original target detection object image, and then copy and paste it multiple times at random positions in the image.

[0036] Set the total number of targets in the original sample to =m, the number of copy and paste times is n, then the total number of targets after copying is = n + m Since the operation is performed on the same target detection image, the number of samples N image Remain unchanged.

[0037] (2) Single target-multiple image copy and paste: Select a target detection object in the original target detection image, and then copy and paste it multiple times at random positions in the remaining images.

[0038] Set the total number of targets in the original sample to =m, the sample size is N image, take one of the images as the source image and select a target detection object, the number of remaining images is N image_new= N image_new−1, the number of copies and pastes is n (in each remaining image), then the total number of objects after copying is =m+ n × N image_new.

[0039] (3) Multi-target-single-image copy and paste: Different target detection objects are selected from multiple original target detection object images, and then the multi-target detection objects are copied and pasted multiple times at random positions in the image.

[0040] Set the total number of targets in the original sample to =m, the sample size is N image , select K different target detection objects from M original target detection object images, and then copy and paste n times in the same target detection object image, then the total number of targets after copying is N targetcopied=m+ K × n .

[0041] (4) Multi-target-multi-image copy and paste: Select different target detection objects from multiple original target detection object graphics, and then copy and paste the multiple target detection objects multiple times at random positions in different target detection object images.

[0042] Set the total number of targets in the original sample to =m, select K different target detection objects from M original target detection object images, and then copy and paste each target detection object image n times in N target detection object images (assuming that each target detection object image accepts target detection objects from different original target detection object images), then the total number of targets after copying is N targettotal=m+ K × N × n .

[0043] Therefore, in this embodiment, by performing data enhancement on the target detection object image data set, the data set is expanded in a targeted manner to simulate different scenarios, thereby enhancing data diversity and avoiding overfitting in subsequent model training. At the same time, data enhancement on the target detection object image data set collected from the low-altitude perspective of the drone can not only increase the number of image data sets, but also effectively increase the frequency of small targets and their relative size in the target detection object image through methods such as multi-angle image transformation, oversampling operations, and random pasting operations, thereby ensuring that the YOLOV7-OBB model can fully learn the feature information of small targets during the training process, significantly enhancing the comprehensiveness and accuracy of feature extraction.

[0044] It should be noted that, in this embodiment, the YOLOV7-OBB model is a rotating target detection model based on the YOLOV7 model, which inherits the high efficiency and real-time performance of the YOLO series of algorithms and can accurately capture tilted or rotated target detection objects. Specifically, the network structure of the YOLOV7-OBB model is basically the same as that of the YOLOV7 model. The difference is that the YOLOV7-OBB model adds a one-dimensional rotation angle θ to the output dimension, and uses the KLD loss function for optimization, thereby achieving accurate detection of rotating target detection objects. Figure 2 As shown, the existing YOLOV7-OBB model structure includes an input module, a CBS module (Conv convolution layer, BN normalization layer and Sigmoid activation function), an ELAN module (Efficient Long-DistanceAttention Network), an MP module (downsampling module) and a REP module (reparameterization module). It should be noted that in this embodiment, the MP module includes two branch structures, the upper branch structure passes through the maximum pooling module and the CBS module, and the lower branch structure passes through two CBS modules. The two branch structures are spliced ​​together to form the output of the MP module. In this embodiment, the structure of the REP module is different during training and reasoning. During training, the REP module has three branch structures, the upper branch structure includes a convolution layer and a normalization layer with a convolution kernel of 1×1, and the main branch structure includes a convolution layer and a normalization layer with a convolution kernel of 3×3, and the lower branch has only one normalization layer. During reasoning, the REP module only retains the main branch structure with a convolution kernel of 3×3.

[0045] Considering that the target detection object occupies a smaller proportion of the image than the ground perspective and has limited features under the low-altitude perspective of the drone, the YOLOV7-OBB model performs poorly in small target detection and is prone to missed detection and false detection. Therefore, in this embodiment, two improvements are made on the basis of the YOLOV7-OBB model. On the one hand, the YOLOV7-OBB model's understanding of global features is enhanced by adding an attention mechanism module (CA), so that the YOLOV7-OBB model can autonomously learn to focus on key areas. On the other hand, in the CBS module, 1×1 convolution and dropout modules are used to increase nonlinearity and avoid overfitting of the YOLOV7-OBB model.

[0046] like Figure 3 As shown, in this embodiment, the YOLOV7-OBB model is improved, and an attention mechanism module (Coordinate Attention, CA) is added to the backbone network of the YOLOV7-OBB model to enhance the feature extraction capability of the YOLOV7-OBB model.

[0047] The backbone network of the YOLOV7-OBB model includes a CBS module and an ELAN module. The specific structure of the backbone network of the YOLOV7-OBB model is: CBS module, CBS module, CBS module, ELAN module, MP-1 module, ELAN module, MP-1 module, ELAN module, MP-1 module, and ELAN module connected in sequence.

[0048] In this embodiment, the attention mechanism module CA is connected after the CBS module and the ELAN module respectively. The specific structure of the backbone network of the improved YOLOV7-OBB model is: CBS module, CBS module, CBS module, CA module, ELAN module, MP-1 module, ELAN module, CA module, MP-1 module, ELAN module, CA module, MP-1 module, and ELAN module connected in sequence.

[0049] It is worth noting that, in this embodiment, Figure 4 As shown, the attention mechanism module CA includes a coordinate information embedding module and a coordinate attention generation module.

[0050] In the coordinate information embedding module, the attention mechanism module CA receives the feature map (C×H×W) output by the previous convolutional layer as input. Subsequently, in order to capture the important spatial information in the feature map, the attention mechanism module CA performs global pooling operations on the feature map in the horizontal direction (C×H×1) and the vertical direction (C×1×W). The horizontal pooling results and the vertical pooling results are concatenated along the channel dimension to obtain the coordinate information tensor (C×1×(W+H)).

[0051] In the coordinate attention generation module, the attention mechanism module CA compresses and activates the coordinate information tensor (C×1×(W+H)) to obtain the key feature representation. Furthermore, the attention mechanism module CA restores the original number of channels of the feature map through splitting and independent convolution operations, and generates a first attention weight map in the horizontal direction (C×H×1) and a second attention weight map in the vertical direction (C×1×W). The first attention weight map and the second attention weight map are normalized by the Sigmoid activation function to obtain the first output value and the second output value. In this embodiment, the first output value and the second output value are limited to the interval (0,1). The first output value is used as the attention coefficient in the horizontal direction, and the second output value is used as the attention coefficient in the vertical direction. Different attention coefficients directly reflect the importance of different positions in the feature map.

[0052] Finally, in order to integrate the attention mechanism into the feature map, the attention mechanism module CA performs matrix multiplication operations on the feature map with the attention coefficient in the horizontal direction and the attention coefficient in the vertical direction, respectively, to obtain a feature map with added attention mechanism. This process is to perform spatial weighting on the feature map, emphasizing important areas while weakening the influence of secondary areas. In this way, the attention mechanism module CA effectively improves the YOLOV7-OBB model's ability to capture key features, thereby improving overall performance.

[0053] It is worth noting that, in this embodiment, Figure 5 As shown in the figure, in the CBS module, a 1×1 convolution layer and a dropout layer are used to increase nonlinearity and avoid overfitting of the YOLOV7-OBB model.

[0054] Specifically, the 1×1 convolution layer does not change the spatial dimension of the input feature map, and mainly acts on the number of output channels, including dimensionality reduction and increasing nonlinearity. By adjusting the number of output channels, the 1×1 convolution layer can achieve dimensionality reduction of the feature map, thereby reducing the number of parameters and calculations of the YOLOV7-OBB model and improving computational efficiency. At the same time, the 1×1 convolution layer can also perform linear combination and nonlinear mapping on the channel dimension, increase the nonlinear ability of the YOLOV7-OBB model, and enhance the expressiveness of the YOLOV7-OBB model.

[0055] The purpose of the Dropout layer is to prevent the YOLOV7-OBB model from overfitting. The Dropout layer randomly selects a certain proportion of neurons in each training batch and sets their output to zero, thereby forcing the YOLOV7-OBB model not to rely on specific neurons during training. This operation of randomly discarding neurons reduces the co-adaptability between neurons and increases the generalization ability of the YOLOV7-OBB model.

[0056] Therefore, in this embodiment, based on the YOLOV7-OBB model, the backbone network structure is optimized, and an attention mechanism module CA is added to the backbone network. The attention mechanism module CA dynamically adjusts the attention of the YOLOV7-OBB model to different areas of the input target detection image, so that the YOLOV7-OBB model can focus more on key features during the detection process, effectively suppress the interference of irrelevant information, and effectively improve the accuracy and robustness of target detection.

[0057] In this embodiment, after improving the YOLOV7-OBB model, the YOLOV7-OBB model is trained based on the training set, and the YOLOV7-OBB model during the training process is evaluated based on the validation set to obtain a trained YOLOV7-OBB model, and the test set is input into the trained YOLOV7-OBB model for target detection to obtain a target detection result.

[0058] Therefore, the present embodiment provides a method for detecting low-altitude target from a drone based on deep learning. By constructing the YOLOV7-OBB model, better results can be obtained for detecting low-altitude target from a drone using only lightweight detectors. Using detectors with simpler structures in the process of detecting low-altitude target from a drone can greatly reduce the resources required for calculation, improve the detection speed of the YOLOV7-OBB model, and reduce the video memory occupancy required in the detection process, thus avoiding unnecessary waste and effectively utilizing limited computing resources. This enables the practical application of the YOLOV7-OBB model in detecting low-altitude target from a drone, and has strong practicality and wide applicability.

[0059] The present embodiment also provides a low-altitude target detection device for unmanned aerial vehicles based on deep learning, including a data acquisition module, a data processing module, a model improvement module, a model training module and a target detection module. The data acquisition module uses a drone to collect a target detection object image data set. The data processing module performs data enhancement on the target detection object image data set, and divides the enhanced target detection object image data set into a training set, a verification set and a test set. The model improvement module improves the YOLOV7-OBB model; wherein an attention mechanism module is added to the backbone network of the YOLOV7-OBB model. The model training module trains the YOLOV7-OBB model based on the training set, and evaluates the YOLOV7-OBB model in the training process based on the verification set to obtain a trained YOLOV7-OBB model. The target detection module inputs the test set into the trained YOLOV7-OBB model for target detection to obtain a target detection result. The present embodiment provides a drone low-altitude perspective target detection device based on deep learning that can implement any method of the drone low-altitude perspective target detection method based on deep learning, and the specific working process of a drone low-altitude perspective target detection device based on deep learning can refer to the corresponding process in the drone low-altitude perspective target detection method embodiment based on deep learning. The method and device provided in this embodiment can be implemented in other ways. For example, the device embodiment described above is only schematic; for example, the division of a module is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point, the mutual connection or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, or it can be an electrical, mechanical or other form of connection.

[0060] This embodiment also provides a computer device. A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method for detecting low-altitude target from a drone based on deep learning when executing the computer program.

[0061] This embodiment also provides a computer-readable storage medium. A computer-readable storage medium, on which a computer program is stored, and when the computer program is executed by a processor, a method for detecting low-altitude targets from a drone based on deep learning described in this embodiment is executed. Among them, the computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by or in combination with an instruction execution system, device or device; the program code contained on the computer-readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wires, optical cables, RF, etc., or any suitable combination of the above.

[0062] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited thereto, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all belong to the protection scope of the present invention. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate the name, and do not indicate any specific order.

Claims

1. A method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning, characterized in that: The following steps are involved: Use drones to collect target detection object image datasets; Performing data enhancement on the target detection object image dataset, and dividing the enhanced target detection object image dataset into a training set, a validation set, and a test set; Improve the YOLOV7-OBB model; add an attention mechanism module to the backbone network of the YOLOV7-OBB model; Train the YOLOV7-OBB model based on the training set, evaluate the YOLOV7-OBB model during training based on the validation set, and obtain the trained YOLOV7-OBB model; The test set is input into the trained YOLOV7-OBB model for target detection to obtain the target detection result.

2. According to the method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning in claim 1, it is characterized in that: The backbone network of the YOLOV7-OBB model includes a CBS module and an ELAN module; an attention mechanism module is connected after the CBS module and the ELAN module.

3. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 2, characterized in that: The attention mechanism module includes a coordinate information embedding module and a coordinate attention generation module.

4. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 3, characterized in that: In the coordinate information embedding module, The attention mechanism module receives the feature map output by the previous convolutional layer as input; The attention mechanism module performs global pooling operations on the feature map in the horizontal and vertical directions respectively; The horizontal pooling results and the vertical pooling results are concatenated along the channel dimension to obtain a coordinate information tensor.

5. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 4, characterized in that: In the coordinate attention generation module, The attention mechanism module compresses and activates the coordinate information tensor to obtain key feature representations; The attention mechanism module restores the original number of channels of the feature map through splitting and independent convolution operations, and generates the first attention weight map in the horizontal direction and the second attention weight map in the vertical direction; The first attention weight map and the second attention weight map are respectively normalized by the Sigmoid activation function to obtain a first output value and a second output value; The first output value is used as the attention coefficient in the horizontal direction, and the second output value is used as the attention coefficient in the vertical direction.

6. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 5, characterized in that: Perform matrix multiplication operations on the feature map with the attention coefficient in the horizontal direction and the attention coefficient in the vertical direction respectively to obtain the feature map with added attention mechanism.

7. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 1, characterized in that: The data enhancement of the target detection object image data set includes an oversampling operation and a random pasting operation.

8. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 7, characterized in that: The specific steps of the oversampling operation are: setting the target detection object image as the copy sample; setting the target detection object image data set as the total sample; copying the copy sample several times; setting the sampling rate, and controlling the copy frequency of the copy sample and the proportion of the copy sample to the total sample through the sampling rate.

9. The method for detecting targets from a low altitude perspective of an unmanned aerial vehicle based on deep learning according to claim 7, characterized in that: The specific steps of the random pasting operation are: The target detection object in the target detection object image is randomly copied to any position in the image to generate a new target detection object image and annotation information; Crop the target detection object in the target detection object image according to the annotation information; The cropped target object is processed and pasted into the new target object image to generate a label.

10. A low-altitude target detection device for unmanned aerial vehicles based on deep learning, characterized in that: include: The data acquisition module uses drones to collect target detection object image data sets; A data processing module performs data enhancement on a target detection object image data set, and divides the enhanced target detection object image data set into a training set, a validation set, and a test set; Model improvement module, improves the YOLOV7-OBB model; in particular, adds an attention mechanism module to the backbone network of the YOLOV7-OBB model; Model training module, trains the YOLOV7-OBB model based on the training set, evaluates the YOLOV7-OBB model during training based on the validation set, and obtains the trained YOLOV7-OBB model; The target detection module inputs the test set into the trained YOLOV7-OBB model for target detection and obtains the target detection result.

Images