A target detection method for low-altitude unmanned aerial vehicle aerial photograph image

Abstract

The application discloses a target detection method for low-altitude unmanned aerial vehicle aerial photography images, first, different levels of features output by a feature extraction network from aerial photography images are acquired and sent to a feature attention fusion module, global and local features of adjacent layers are fused to obtain multi-scale features, under the guidance of the feature attention fusion module, an adaptive sampling strategy is used to dynamically set a threshold for each true value target and filter out optimal positive samples, and finally, the multi-scale features are transmitted to a detection head, and the optimal positive samples filtered out by the adaptive sampling strategy are detected. By combining the adaptive sampling strategy and the feature attention fusion module, the model focuses on the key information of the image, the representation ability of the multi-scale features is improved, the problems of sample imbalance and inaccurate selection of difficult samples caused by a fixed threshold sampling strategy are effectively solved, and the accuracy and robustness of the model are improved without significantly increasing the size of the parameter quantity of the model.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision, and particularly relates to a target detection method for low-altitude unmanned aerial vehicle aerial image. BACKGROUND

[0002] The target detection technology for aerial image has a wide range of applications in many fields, such as city planning, disaster relief, military reconnaissance, and automatic driving. With the progress of technology, the aerial image target detection has made a great leap forward. Early algorithms mainly rely on manually designed feature descriptions, which improve the detection accuracy to a certain extent, but have weak generalization ability and a complex and time-consuming design process. In the machine learning stage, the aerial image target detection begins to introduce more complex classification algorithms to automatically learn features from training data, reducing the dependence on manual feature design, but still relies on feature engineering and has limited adaptability to complex scenes. In the deep learning stage, the convolutional neural network automatically extracts hierarchical features from the original image with strong feature learning ability, and the performance of the aerial image target detection algorithm based on deep learning is also significantly improved.

[0003] Most of the existing aerial image target detection algorithms based on deep learning use the maximum IoU sampling strategy to select positive and negative samples, but the fixed sampling threshold is not friendly to aerial image targets with large aspect ratio and area difference. In addition, the simple extraction method of multi-scale features easily introduces low-level noise into high-level features, thereby interfering with the detection of targets. If corresponding solutions can be introduced, the performance of the aerial image target detection model can be greatly improved. SUMMARY

[0004] The purpose of the present application is to provide a target detection method for low-altitude unmanned aerial vehicle aerial image. For the specific properties of the aerial image dataset, by introducing an adaptive sampling strategy and a feature attention fusion module, the problems of fixed sampling strategy threshold and insufficient multi-scale feature representation ability of existing aerial image target detection methods are solved, and the accuracy and robustness of the aerial image target detection model are improved.

[0005] The technical solution for achieving the purpose of the present application is as follows: a target detection method for low-altitude unmanned aerial vehicle aerial image, comprising the following steps:

[0006] Step S1, download the dataset Drone-Vehicle, which includes 28439 pairs of RGB-infrared image pairs taken by a low-altitude unmanned aerial vehicle, select all RGB images for data augmentation, construct an augmented dataset, and divide the augmented dataset into a training set and a test set according to a proportion, and proceed to step S2.

[0007] Step S2, a target detection network based on adaptive sampling and attention fusion is constructed, the target detection network is composed of a feature extraction network, a feature attention fusion module, an adaptive sampling strategy and a detection head, and step S3 is entered.

[0008] Step S3, the target detection network is trained by using the training set to obtain a trained target detection model, and the specific process is as follows.

[0009] Step S3-1, the feature extraction network extracts features from the training set images, and outputs multiple layers of original features with different resolutions and semantic information.

[0010] Step S3-2, one layer of the highest level is selected as the original feature, and two layers of multi-scale features are directly generated from the original feature, considering the particularity of the unmanned aerial vehicle image view angle and the morphological variability of the target object, in order to enhance the fine-grained information extraction of local features, the multi-scale feature with large size is selected as the starting point from the two layers, and the corresponding original feature is input into the feature attention fusion module from top to bottom, so as to obtain multi-layer multi-scale features.

[0011] Step S3-3, each layer of multi-scale features generated by the feature attention fusion module is input into the adaptive sampling strategy, considering the significant difference of the number distribution, aspect ratio and area size of the unmanned aerial vehicle image target object, the sampling threshold is dynamically set by combining the statistical characteristics and the two information of the true value target itself, and the optimal positive sample of each true value target is screened out.

[0012] Step S3-4, each layer of multi-scale features is transmitted into the detection head, and the optimal positive sample screened out by the adaptive sampling strategy is detected, and the detection result is output.

[0013] Step S3-5, the trained target detection model is obtained, and step S4 is entered.

[0014] Step S4, the test set is input into the trained target detection model, and the class and position of each target in the test set image are output to evaluate the precision of the trained model. Compared with the prior art, the advantages of the present application are:

[0015] 1) Since the aerial image data set has the characteristics of unbalanced quantity, large difference in target area ratio and large change in target aspect ratio, based on this, a new adaptive sampling strategy is proposed, the innovation of the strategy is to set a dynamic threshold for selecting samples by combining the statistical characteristics and the characteristics of the aerial image target, and the problems of sample imbalance and inaccurate selection of difficult samples caused by using a fixed threshold sampling strategy in aerial image target detection are effectively solved.

[0016] 2) Due to the particularity of the perspective of the unmanned aerial vehicle image, the target in different levels of features is more complex, the noise of the low-level features is introduced into the high-level features by using the simple feature addition mode of the FPN, based on this, the application introduces a feature attention fusion module, uses the EAU to obtain attention features, and adds the weighted average and the original features to obtain multi-scale features with strong representation ability, which not only enhances the model to extract the fine-grained information of the local features, but also retains the edge and detail features of the target, so that the target detection model can more effectively utilize the feature information under different resolutions, thereby improving the ability to detect small targets.

[0017] 3) The application innovatively combines the adaptive sampling strategy and the attention fusion module to construct a network with global and local sensitivity, the adaptive sampling strategy and the attention fusion module have complementary advantages in the detection of different categories of targets in aerial images, the combination of the two enhances the effect of each other, and under the guidance of the attention fusion module, the adaptive sampling strategy is more accurate in selecting positive samples, and the precision and robustness of the aerial image target detection are improved. BRIEF DESCRIPTION OF DRAWINGS

[0018] Fig. 1 The flowchart of the target detection method for low-altitude unmanned aerial vehicle aerial images of the application.

[0019] Fig. 2 The structural diagram of the adaptive sampling strategy designed by the application.

[0020] Fig. 3 The structure diagram of the attention fusion module designed by the application. DETAILED DESCRIPTION

[0021] The application provides a target detection method for low-altitude unmanned aerial vehicle aerial images, aiming to solve the problems of fixed sampling strategy threshold and insufficient multi-scale feature representation ability of the commonly used aerial image target detection method, improve the accuracy and robustness of the aerial image target detection model, and the following further detailed description of the embodiment of the application is made.

[0022] In combination Fig. 1 to Fig. 3 The target detection method for low-altitude unmanned aerial vehicle aerial images provided by the application comprises the following steps:

[0023] Step S1, download the data set Drone-Vehicle, the data set includes 28439 pairs of RGB-infrared image pairs taken by a low-altitude unmanned aerial vehicle, select all the RGB images for data enhancement, construct an enhanced data set, and divide the enhanced data set into a training set and a test set according to the proportion.

[0024] The data enhancement operation includes a random horizontal flip in a training process, the original image is cropped into a 712*840 sub-image during training, the data enhancement operation can avoid overfitting and improve the robustness of the model, and the enhanced data set also contains a training set and a test set, and the method proceeds to step S2.

[0025] In step S2, a target detection network for aerial images based on adaptive sampling and attention fusion is constructed, and the target detection network is composed of a feature extraction network, a feature attention fusion module, an adaptive sampling strategy, and a detection head.

[0026] The main part of the target detection network is a residual network ResNet, a feature attention fusion module is introduced based on the FPN pyramid network, the detection head includes a feature alignment module and a directional detection module, the feature alignment module generates high-quality anchor boxes through an anchor box refinement network, and aligns the anchor boxes with convolution features through an alignment convolution, the directional detection module uses an active rotation filter to generate direction-sensitive features, and the adaptive sampling strategy is responsible for screening the optimal positive sample for each true value target during network training.

[0027] In step S3, the target detection network is trained using the training set to obtain a trained target detection model, and the specific process is as follows.

[0028] In step S3-1, the feature extraction network extracts features from the training set images, and outputs multiple layers of original features with different resolutions and semantic information, and the method proceeds to step S3-2.

[0029] In step S3-2, one layer with the highest level is selected as the original feature, and two layers of multi-scale features are directly generated from the original feature, considering the particularity of the unmanned aerial vehicle image view and the variability of the target object shape, in order to enhance the extraction of fine-grained information of local features, the multi-scale feature with large size is selected as the starting point from the two layers, and the corresponding original feature is input into the feature attention fusion module from top to bottom, so as to obtain multiple layers of multi-scale features, and the specific process is as follows.

[0030] The highest layer of original features generates two layers of multi-scale features through 3*3 convolution, the feature with large size is selected as the second layer, the second layer of multi-scale features and the highest layer of original features are input into the feature attention fusion module to generate the third layer of multi-scale features, and the above process is repeated, and the lower layer of multi-scale features is gradually generated by using the upper layer of multi-scale features and the corresponding original features.

[0031] For the upper-level multi-scale features and their corresponding original features, after unifying their size according to the resolution of the original features, they are added together to obtain the fused features. The fused features are used as the input of EAU in the feature attention fusion module to obtain the local attention features and global attention features of the fused features. The two are added together to obtain the attention features. The attention features are then subjected to a weighted average of the unified upper-level multi-scale features and the original features using the Sigmoid function. Finally, the attention features are added to the original features to obtain the lower-level multi-scale features.

[0032] The fused features are used as input to the efficient attention unit (EAU) in the feature attention fusion module to obtain the local attention features and global attention features of the fused features. The attention features are obtained by adding the two together, as follows:

[0033] The upper-level multi-scale features are represented as feature X1, and the corresponding original features are represented as feature Y2. The resolution of X1 is half the resolution of Y2. To ensure the consistency of feature size, feature X1 is converted to the same size as feature Y2 through bilinear interpolation and named Y1.

[0034] Y1 and Y2 are added together to obtain the fused feature Y'. The fused feature Y' is used as the input to EAU to obtain the local attention feature Y of the fused feature Y'. L (Y') and global attention feature Y G (Y'), and then perform an addition operation on the two to obtain the attention feature of the fused feature;

[0035] Local attention feature Y of fusion features L The formula for calculating (Y') is as follows:

[0036] Y L (Y')=PConv(ReLU(PConv(Y'))

[0037] Here, PConv represents 1×1 convolution. In EAU, PConv is used for the first time to reduce the number of feature channels of the fused feature Y', and PConv is used a second time to restore the number of feature channels of the fused feature Y' to the original number of channels. In addition, EAU uses the non-linear function ReLU as the activation function.

[0038] Global attention feature Y of fused features G The formula for calculating (Y') is as follows:

[0039] Y G (Y')=PConv(ReLU(PConv(Avg(Y'))))+PConv(ReLU(PConv(Max(Y'))))

[0040] Wherein, Avg(Y') represents the global average pooling processing to the fusion feature Y', Max(Y') represents the global maximum pooling processing to the fusion feature Y', the first use of PConv in EAU is to reduce the feature channel of the fusion feature Y', and the second use of PConv is to restore the feature channel of the fusion feature Y' to the original channel number.

[0041] After obtaining the global attention feature and the local attention feature of the fusion feature, the two are added to output the attention feature Z, and the calculation formula of the attention feature Z is as follows:

[0042] Z=Y L (Y')+Y G (Y')

[0043] The attention feature Z is subjected to a Sigmoid function to perform weighted average on the unified size of the upper layer multi-scale feature and the original feature, and finally added to the original feature to obtain the lower layer multi-scale feature, as follows:

[0044] The attention feature Z is subjected to a Sigmoid function to perform weighted average on the unified size of the feature Y1 and the feature Y2, and the network determines its own weight through training, and the calculation formula of the weighted average feature Y3 is as follows:

[0045] Y3=(1-ε)Y1(1-Sigmoid(Z))+εY2Sigmoid(Z)

[0046] In the formula, ε represents a weighting coefficient obtained by training the target detection network, which is used to control the proportion of the feature Y1 and the feature Y2 in the weighted average.

[0047] At the same time, in order to save as much original information as possible, the original feature Y2 and the weighted average feature Y3 are added again, and the calculation formula of the final multi-scale feature Y of the lower layer is as follows:

[0048] Y=(1-ε)Y1(1-Sigmoid(Z))+εY2Sigmoid(Z)+Y2

[0049] According to the characteristics of low-flying unmanned aerial vehicles and vehicle targets, a feature attention fusion module is introduced to emphasize the extraction and fusion of local detailed features, thereby improving the model performance and robustness, and the step S3-3 is entered.

[0050] In step S3-3, each layer of multi-scale feature generated by the feature attention fusion module is input into the adaptive sampling strategy. Considering the significant differences in the number distribution, aspect ratio and area size of the unmanned aerial vehicle image target objects, the adaptive sampling strategy combines the statistical characteristics of each ground truth target and its own characteristics to dynamically set the sampling threshold and select the optimal positive sample of each ground truth target, as follows:

[0051] The multi-layer multi-scale features generated by the feature attention fusion module are taken as the input of the adaptive sampling strategy. According to the characteristics of the data set Drone-Vehicle, the adaptive sampling strategy dynamically sets the first threshold value by using the statistical characteristics of each ground truth target during network training, dynamically sets the second threshold value by using the characteristics of each ground truth target, and finally sets different final threshold values for each ground truth target by combining the two threshold values, and the optimal positive samples are screened according to the final threshold values.

[0052] According to the characteristics of the data set Drone-Vehicle, the adaptive sampling strategy dynamically sets the first threshold value by using the statistical characteristics of each ground truth target during network training, and the specific implementation is as follows:

[0053] In the adaptive sampling strategy, it is assumed that the number of multi-scale feature output layers is L, and the optimal A anchor boxes are selected for each ground truth box at each layer, so there are A*L candidate boxes in total. The mean μ and the standard deviation σ of each ground truth are calculated according to the intersection over union (IOU), and the sum of the mean and the standard deviation is taken as the first threshold value t1 of each ground truth target based on the statistical characteristics, that is,

[0054] t1 = μ + σ

[0055] The mean μ reflects the matching degree of the anchor box and the ground truth box, and the higher the mean μ, the better the matching of the anchor box and the ground truth box. The standard deviation σ reflects which layer of multi-layer multi-scale features is used to detect the target. Adding the mean and the standard deviation as the IOU threshold value can automatically screen the positive samples on the corresponding layer features.

[0056] The adaptive sampling strategy dynamically sets the second threshold value by using the characteristics of each ground truth target, and finally sets different final threshold values for each ground truth target by combining the two threshold values, and the optimal positive samples are screened according to the final threshold values. The specific implementation is as follows:

[0057] Due to the characteristics of the aerial image data set, such as extremely unbalanced quantity, large difference in target area ratio, and large change in target aspect ratio, the characteristics of the ground truth target itself are taken as the second factor for selecting positive samples. By introducing the characteristics of the ground truth target, such as area, quantity, and aspect ratio, more information is considered when setting the threshold value, so that the detection is more accurate.

[0058] First, the proportion of each class is counted, and the proportion c of class i is calculated as follows: i

[0059]

[0060] In the formula, N i represents the number of class i, N k represents the number of class k, K represents the total number of classes of the ground truth target, and i = 1, 2, 3, …, K.​

[0061] The area and shape information of the single target is combined with the corresponding category proportion, and the second threshold t of the target j belonging to the category i is set based on the characteristics of the target itself 2,i,j As follows:

[0062]

[0063] Wherein, W0 represents the width of the input image, H0 represents the height of the input image, the aspect ratio of the ground truth target ratio = w / h, the ground truth target area area = w x h, w represents the width of the ground truth target, and h represents the height of the target.

[0064] When the exponential part of the exponential function is negative, the second threshold t 2,i,j Presents a decreasing trend, and the second threshold t 2,i,j The range is controlled within 0-1.

[0065] The threshold of the positive sample selection is monotonically decreasing with respect to the aspect ratio of the ground truth target, monotonically increasing with respect to the number of target category samples, and monotonically decreasing with respect to the area proportion of the target itself. The combination of area, number and aspect ratio enables the network to adaptively set the threshold of the positive sample.

[0066] Finally, the first threshold and the second threshold are multiplied to obtain the final threshold as follows:

[0067]

[0068] Wherein, t i,j Indicates the final threshold.

[0069] By introducing the adaptive sampling strategy, a threshold is dynamically set for each ground truth target, improving the model performance and robustness. In particular, the adaptive sampling strategy improves the detection of targets with large area and aspect ratio compared to other categories in aerial images, and the optimal positive sample is selected according to the final threshold.

[0070] Go to step S3-4.

[0071] Step S3-4, pass each layer of multi-scale features into the detection head, and use the optimal positive sample selected by the adaptive sampling strategy for detection, output the detection result, and go to step S3-5.

[0072] Step S3-5, get the trained target detection model, and go to step S4.

[0073] Step S4, input the test set into the trained target detection model, and output the category and position of each target in the test set image to evaluate the precision of the trained model.

[0074] The target detection method for low-altitude unmanned aerial vehicle aerial photograph image provided by the application is realized based on a python language and a pytorch framework, model training is performed using a single TeslaV100-PCIE-32GB graphics card, and the inference process is performed using an NVIDIA GeForce RTX 3080.

[0075] In the experiment, the model training is optimized by adopting an SGD strategy, the momentum attenuation is 0.9, and the weight attenuation is 0.0001. The initial learning rate is 0.0025, and the learning rate is divided by 10 at the 24th round and the 33rd round. The training round of the Drone-Vehicle dataset is 36 rounds, and the size of each batch is 1. When the Drone-Vehicle dataset is used for training, the model training input is a small piece of image after cropping, the size is 712*840, and the IoU threshold value is set to 0.1 when the small piece of image prediction result is merged in the inference. In the adaptive sample selection strategy, the value of the optimal anchor frame number A of each layer is set to 9, and the value of the weight epsilon in the feature attention fusion module is set to 0.5.

[0076] In order to show the effectiveness of the application, some mainstream remote sensing image target detection algorithms are selected as comparison models, the method of the application respectively takes ResNet50 and ResNet101 as backbone networks in the experiment, the IoU threshold value of other models is set to 0.5, and the remaining parameters remain unchanged, wherein R50 represents the ResNet50 network, R101 represents the ResNet101 network, and the comparison experimental results are shown in Table 1.

[0077] Table 1 Comparison experimental results of the application and other remote sensing target detection algorithms under the Drone-Vehicle dataset

[0078] Model name mAP (%) ReDet (R50) 60.8 S 2 ANet(R50)]]> 61.7 [R 3 Det(R101)]]> 63.4 The present invention (R50) 68.6 The present invention (R101) 68.6

[0079] The experimental results prove the practicability and effectiveness of the method of the application.

Claims

1. A target detection method for low-altitude UAV aerial images, characterized in that, Includes the following steps: Step S1: Download the Drone-Vehicle dataset, which includes 28,439 pairs of RGB-infrared images taken by low-altitude drones. Select all RGB images for data augmentation to construct an augmented dataset. Divide the augmented dataset into training and testing sets according to the proportions and proceed to step S2. Step S2: Construct an aerial image target detection network based on adaptive sampling and attention fusion. The target detection network consists of four parts: a feature extraction network, a feature attention fusion module, an adaptive sampling strategy, and a detection head. Proceed to step S3. Step S3: Train the object detection network using the training set to obtain the trained object detection model, as detailed below: Step S3-1: The feature extraction network extracts features from the training set images and outputs multiple layers of original features, which have different resolutions and semantic information. Step S3-2: Select the highest level layer as the original feature, and directly generate two layers of multi-scale features from the original feature. Considering the special perspective of UAV images and the variability of target object shapes, in order to enhance the extraction of fine-grained information of local features, select the larger multi-scale feature from the two layers as the starting point, and input it together with the corresponding original feature from top to bottom into the feature attention fusion module to obtain multi-layer multi-scale features. Step S3-3: Input the multi-scale features generated by the feature attention fusion module into the adaptive sampling strategy. Considering the significant differences in the distribution of the number of target objects, aspect ratio, and area size in the UAV image, the sampling threshold is dynamically set by combining the statistical features and characteristics of each ground truth target to select the optimal positive sample for each ground truth target. Step S3-4: Input the multi-scale features of each layer into the detection head, use the optimal positive samples selected by the adaptive sampling strategy for detection, and output the detection results; Step S3-5: Obtain the trained object detection model, then proceed to step S4; Step S4: Input the test set into the trained object detection model and output the category and location of each object in the test set image to evaluate the accuracy of the trained model.

2. The target detection method for low-altitude UAV aerial images according to claim 1, characterized in that, In step S3-2, the highest-level layer is selected as the original feature, and two layers of multi-scale features are directly generated from the original features. Considering the special perspective of UAV images and the variability of target object shapes, in order to enhance the extraction of fine-grained information of local features, a larger multi-scale feature is selected from the two layers as the starting point. It is input into the feature attention fusion module from top to bottom along with the corresponding original features to obtain multi-layer multi-scale features, as follows: The highest-level layer is selected as the original feature. The original feature is used to generate two multi-scale feature layers through 3×3 convolution. The feature with the larger size is selected as the second layer. The second multi-scale feature and the original feature are input into the feature attention fusion module to generate the third multi-scale feature layer. This process is repeated, using the upper multi-scale features and their corresponding original features to gradually generate the lower multi-scale features. For the upper multi-scale feature and its corresponding original feature, after unifying the size according to the resolution of the original feature, they are added together to obtain the fused feature. The fused feature is used as the input to the EAU in the feature attention fusion module to obtain the local attention feature and the global attention feature of the fused feature. The two are added together to obtain the attention feature. The attention features are weighted by the Sigmoid function to calculate the average of the upper-level multi-scale features after unification and the original features. Finally, the average is added to the original features to obtain the lower-level multi-scale features.

3. The target detection method for low-altitude UAV aerial images according to claim 2, characterized in that, For the upper-level multi-scale features and their corresponding original features, after unifying their size according to the resolution of the original features, they are added together to obtain the fused features. The fused features are used as input to the EAU in the feature attention fusion module to obtain the local attention features and global attention features of the fused features. The two are added together to obtain the attention features, as follows: The upper-level multi-scale features are represented as feature X1, and the corresponding original features are represented as feature Y2. The resolution of X1 is half the resolution of Y2. To ensure the consistency of feature size, feature X1 is converted to the same size as feature Y2 through bilinear interpolation and named Y1. Y1 and Y2 are added together to obtain the fused feature Y'. The fused feature Y' is used as the input to EAU to obtain the local attention feature Y of the fused feature Y'. L (Y') and global attention feature Y G (Y'), and then perform an addition operation on the two to obtain the attention feature of the fused feature; Local attention feature Y of fusion features L The formula for calculating (Y') is as follows: AND L (Y')=PConv(ReLU(PConv(Y')) Where PConv represents 1×1 convolution. In EAU, PConv is used for the first time to reduce the number of feature channels of the fused feature Y', and PConv is used for the second time to restore the number of feature channels of the fused feature Y' to the original number of channels. In addition, EAU uses the non-linear function ReLU as the activation function. Global attention feature Y of fused features G The formula for calculating (Y') is as follows: AND G (Y')=PConv(ReLU(PConv(Avg(Y'))))+PConv(ReLU(PConv(Max(Y')))) Where Avg(Y') represents global average pooling of the fused feature Y', and Max(Y') represents global max pooling of the fused feature Y'. In EAU, PConv is used for the first time to reduce the number of feature channels of the fused feature Y', and PConv is used for the second time to restore the number of feature channels of the fused feature Y' to the original number of channels. After obtaining the global attention features and local attention features of the fused features, the two are added together, and the formula for calculating the output attention feature Z is as follows: Z=Y L (Y')+Y G (AND').

4. The target detection method for low-altitude UAV aerial images according to claim 2, characterized in that, The attention features are weighted by the Sigmoid function, and the average of the uniformly sized upper-level multi-scale features and the original features is then added to the original features to obtain the lower-level multi-scale features, as follows: The attention feature Z is passed through a sigmoid function and then weighted and averaged with features Y1 and Y2 of uniform size. The network is trained to determine its own weights. The formula for calculating the weighted average feature Y3 is as follows: Y3=(1-ε)Y1(1-Sigmoid(Z))+εY2Sigmoid(Z) In the formula, ε represents the weighting coefficients obtained by training the object detection network, which are used to control the proportion of features Y1 and Y2 in the weighted average; Meanwhile, to preserve as much original information as possible, the original feature Y2 is added again to the weighted average feature Y3, and the final calculation formula for the lower-level multi-scale feature Y is as follows: Y=(1-ε)Y1(1-Sigmoid(Z))+εY2Sigmoid(Z)+Y2 To address the characteristics of low-flying drones and vehicle targets, a feature attention fusion module is introduced to emphasize the extraction and fusion of local detailed features, thereby improving model performance and robustness.

5. A target detection method for low-altitude UAV aerial images according to claim 2, characterized in that, In step S3-3, the multi-scale features generated by the feature attention fusion module are input into the adaptive sampling strategy. Considering the significant differences in the distribution of the number of target objects, aspect ratio, and area size in UAV images, the sampling threshold is dynamically set by combining the statistical features and inherent characteristics of each ground truth target to select the optimal positive sample for each ground truth target, as follows: The multi-layer, multi-scale features generated by the feature attention fusion module are used as input to the adaptive sampling strategy. Based on the characteristics of the Drone-Vehicle dataset, the adaptive sampling strategy dynamically sets the first threshold by utilizing the statistical features of each ground truth target during network training, and dynamically sets the second threshold by utilizing the characteristics of each ground truth target. Finally, the two thresholds are combined to set different final thresholds for each ground truth target, and the optimal positive sample is selected based on the final threshold.

6. A target detection method for low-altitude UAV aerial images according to claim 5, characterized in that, The multi-layer, multi-scale features generated by the feature attention fusion module are used as input to the adaptive sampling strategy. Based on the characteristics of the Drone-Vehicle dataset, the adaptive sampling strategy dynamically sets the first threshold during network training using the statistical features of each ground truth target, as follows: In the adaptive sampling strategy, assuming the number of multi-scale feature output layers is L, the top A anchor boxes are selected for each ground truth box in each layer, resulting in a total of A×L candidate boxes. The mean μ and standard deviation σ of each ground truth box are calculated based on the intersection-over-union ratio (IoU). The sum of the mean and standard deviation is used as the first threshold t1 for each ground truth target based on statistical features. t1=μ+σ The mean μ reflects the matching degree between the anchor box and the ground truth box. The higher the mean μ, the better the matching between the anchor box and the ground truth box. The standard deviation σ reflects which layer in the multi-layer, multi-scale feature is used to detect the target. Adding the mean and standard deviation as the IoU threshold can automatically filter positive samples on the corresponding layer features.

7. A target detection method for low-altitude UAV aerial images according to claim 5, characterized in that, The second threshold is dynamically set using the characteristics of each ground truth objective. Finally, the two thresholds are combined to set a different final threshold for each ground truth objective. The optimal positive sample is then selected based on the final threshold, as follows: Because aerial image datasets are characterized by imbalance in quantity, large differences in target area ratio, and large variations in target aspect ratio, the inherent characteristics of the ground truth targets are used as a second factor in selecting positive samples. By introducing the inherent characteristics of the ground truth targets—area, quantity, and aspect ratio—more information is considered when setting the threshold, making the detection more accurate. First, we calculate the percentage of each category, with category i having a percentage c. i The calculation method is as follows: In the formula, N i N represents the number of items in category i. k Indicates the number of categories k, where K represents the total number of categories for the true objective; i = 1, 2, 3, ..., K; The area and shape information of a single target are combined with the corresponding category proportion. The second threshold t for target j belonging to category i is set based on its own characteristics. 2,i,j As shown below: Where W0 represents the width of the input image, H0 represents the height of the input image, the aspect ratio of the ground truth target is ratio = w / h, and the area of ​​the ground truth target is area = w × h, where w represents the width of the ground truth target and h represents the height of the ground truth target. When the exponent of the exponential function is negative, the second threshold t 2,i,j It shows a decreasing trend, the second threshold t 2,i,j The range is controlled within 0 to 1; The threshold for selecting positive samples decreases monotonically with respect to the aspect ratio of the true target, increases monotonically with respect to the number of target category samples, and decreases monotonically with respect to the area ratio of the target itself. The combination of area, number, and aspect ratio enables the network to adaptively set the threshold for positive samples. Finally, multiplying the first threshold and the second threshold yields the final threshold as follows: Among them, t i,j Indicates the final threshold; The optimal positive sample is selected based on the final threshold.

Images