Target detection system in unmanned aerial vehicle aerial scene, unmanned aerial vehicle aerial system

Abstract

The application discloses a kind of unmanned plane aerial scene target detection system, unmanned plane aerial system.For solving the technical problem that unmanned plane aerial in prior art cannot satisfy real-time and high precision simultaneously, in the neural network proposed by the application, including main network, neck network and head network, wherein the main network contains CBAM attention mechanism module and SiLU activation function, which helps to improve the detection accuracy, at the same time, the new main network also reduces the parameter amount of network model, greatly improves the detection speed;Combined with the characteristics of aerial task data set VisDrone, the prior frame size is optimized, and data enhancement methods such as Mosaic and Gaussian blur are used to further improve the detection effect.The application solves the precision and real-time problem in unmanned aerial system, realizes the technical effect of high-precision and high real-time target detection.The application is suitable for unmanned plane aerial field.

Description

Technical Field

[0001] This invention relates to a target detection system and an aerial photography system, specifically to a target detection system and an aerial photography system for drone aerial photography scenarios. Background Technology

[0002] With the widespread adoption of drone technology, remote monitoring using drone aerial photography has become increasingly common in fields such as environmental surveying, road traffic flow monitoring, security inspections, and epidemic prevention and control. Furthermore, employing deep learning algorithms to identify and label objects in images can significantly reduce manual workload and improve monitoring efficiency. Currently, two-stage detection methods are the mainstream approach in object detection, with the R-CNN series being the most representative. Compared to two-stage detection algorithms, single-stage detection methods can simultaneously predict object bounding boxes and categories, thus offering a significant speed advantage, but with lower accuracy. Representative solutions for single-stage detection include YOLO, SSD, and RetinaNet. Even so, the scale of network parameters and computational load are not ideal for the low-power embedded computing platforms that can be mounted on conventional drone platforms, making it difficult to meet real-time requirements.

[0003] While the more lightweight MobileNet and ShuffleNet series solutions can meet real-time performance requirements, they sacrifice accuracy and recall, making them less suitable for detecting small objects in multiple aerial images.

[0004] Furthermore, the inventors discovered significant differences between footage captured by drones in aerial photography scenarios and mainstream object detection datasets:

[0005] (1) Drones take aerial photos of complex urban scenes. The location and time of the aerial photos are different, which will lead to complex and varied background conditions of the ground images. For example, different weather and lighting conditions in the same scene will cause changes in the characteristics of the target, which can easily interfere with the detection of the target.

[0006] (2) In images taken by drones from a distance in the air, the targets are usually small and blurry, and there may be a lot of overlap or occlusion of the targets.

[0007] (3) The perspective is unique. Most of the images are taken by drones from a distance at different angles of attack, which is different from the images taken at close range from a horizontal perspective in the regular dataset. Therefore, the features in this dataset are quite different from the volume features of similar objects in the regular dataset.

[0008] These differences can lead to unsatisfactory target detection results, low accuracy, and difficulty in meeting the needs of practical applications.

[0009] Therefore, proposing a target detection task adapted to aerial photography scenarios that can meet both real-time requirements and ensure accuracy is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] To solve or alleviate some or all of the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0011] A target detection system for drone aerial photography includes an image acquisition system and a target detection device. The target detection device is configured to perform a target detection method based on images captured by the image acquisition system during drone flight. Specifically, the target detection method involves the target detection device receiving image information captured by the image acquisition system and feeding it into a neural network inference device. The neural network inference device is equipped with a neural network, which is used to infer the input signal based on the input signal and trained network parameters, obtain the inference result, and output it through an output device.

[0012] In one embodiment, the neural network includes a backbone network, a neck network, and a head network.

[0013] In one embodiment, the neck network and head network are the neck network and head network of the YOLOv5 neural network.

[0014] In one embodiment, the backbone network includes the following modules connected in sequence: Cov2d, B_1_3, B_2_3, B_3_4, B_4_3, B_4_2; wherein, the Cov2d module contains sequentially connected ordinary convolution, BN, and Hard-swish activation function operations; B_1_3 consists of three sequentially connected ordinary convolution, BN, SiLU, DWConv, and BN modules; B_2_3 consists of three sequentially connected ordinary convolution, BN, SiLU, CBAM, DWConv, and BN modules; and B_3_4 consists of four sequentially connected... The first part of the code is Cov2d, DWConv-H, and BN; the second part is B_4_3, which consists of three sequentially connected Cov2d, CBAM, DWConv-H, and BN; the third part is B_4_2, which consists of two sequentially connected Cov2d, CBAM, DWConv-H, and BN; where DWConv is a depthwise separable convolution; DWConv-H is a sequentially connected DWConv, BN, and Hard-Swish module; BN is batch normalization; and CBAM is the CBAM attention mechanism module, which includes the channel attention module CAM and the spatial attention module SAM.

[0015] In one type of embodiment, the neural network deployed on the neural network inference device is a neural network obtained after training on a dataset, wherein the dataset is subjected to one or more of the following operations:

[0016] 1) Set the probability of flipping vertically to 0;

[0017] 2) Mosaic processing, the specific steps are as follows: select four images; scale, flip and change the color gamut of the four images, then overlap a portion of the four images after the above operations and stitch them together to generate a large new image; then mark the target anchor box on the new image;

[0018] 3) Two-dimensional Gaussian blur processing, the specific steps are as follows: for the pixel to be processed, perform two-dimensional Gaussian blur processing using the pixel values ​​around the pixel, and then use the obtained value as the pixel value of the point.

[0019] 4) Gamma transform processing, the transformation formula is as follows:

[0020] S = cr γ

[0021] Where s is the grayscale output, c is the scaling factor, r is the input grayscale, with a value between 0 and 1, and γ is the grayscale scaling parameter, which controls the degree of scaling of grayscale transformation.

[0022] In one embodiment, the dataset is the VisDrone training set;

[0023] In a certain embodiment, the K-means clustering method is used to calculate the target bounding boxes of the VisDrone training set to obtain the target bounding boxes. Specifically, step (1) is to set 9 anchor boxes of random size as the initial prior boxes.

[0024] Step (2): Assign the anchor box sample of each target to the prior box that is most similar to it;

[0025] Step (3): Calculate the mean of the height and width of the prior box again and set this as the new prior box;

[0026] Step (4): Repeat steps (2) and (3) until the size of the prior box is fixed or the maximum number of iterations is reached.

[0027] A drone aerial photography system includes: a drone, and a target detection system mounted on the drone, wherein the target detection system is a target detection system for drone aerial photography scenarios as described in any of the preceding claims.

[0028] Some or all of the embodiments of the present invention have the following beneficial technical effects:

[0029] 1. In aerial photography scenarios, this invention can meet the requirements of real-time processing and timely processing of aerial photography data, as well as the requirements of accuracy and recall, thus solving the pain points in this field.

[0030] 2. In this invention, by optimizing the prior box size and adjusting the data augmentation strategy, the characteristics of the dataset are utilized to the greatest extent possible without increasing network complexity, thereby optimizing the network parameters and improving the accuracy of target detection.

[0031] Further beneficial effects will be described in the preferred embodiments.

[0032] The technical solutions / features disclosed above are intended to summarize the technical solutions and features described in the Detailed Embodiments section, and therefore the scope of the description may not be entirely the same. However, these new technical solutions disclosed in this section are also part of the numerous technical solutions disclosed in this invention document. The technical features disclosed in this section, together with the technical features disclosed in the subsequent Detailed Embodiments section and some contents in the drawings not explicitly described in the specification, disclose more technical solutions in a reasonable combination.

[0033] The technical solution formed by combining all the technical features disclosed at any position in this invention is used to support the summary of the technical solution, the modification of the patent document, and the disclosure of the technical solution. Attached Figure Description

[0034] Figure 1 This is a structural diagram of a drone aerial photography system;

[0035] Figure 2 This is a schematic diagram of the target detection device;

[0036] Figure 3 This is a schematic diagram of a neural network structure;

[0037] Figure 4 This is a schematic diagram of neural network training and deployment;

[0038] Figure 5 This is a schematic diagram of the backbone network structure;

[0039] Figure 6 This is a structural diagram of the CBAM attention mechanism module. Detailed Implementation

[0040] Since it is impossible to exhaustively describe all alternative solutions, the key points of the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Other technical solutions and details not disclosed in detail below generally belong to technical objectives or features that can be achieved by conventional means in the art, and due to space limitations, they will not be described in detail here.

[0041] Unless it refers to division, the " / " in any position in this invention represents logical "OR". The serial numbers "first", "second", etc., in any position in this invention are merely descriptive distinguishing marks and do not imply an absolute temporal or spatial order, nor do they imply that terms prefixed with such serial numbers necessarily refer to different things than the same terms prefixed with other modifiers.

[0042] This invention describes various key points used to combine into various specific embodiments, which will be incorporated into various methods and products. In this invention, even if a key point is described only when introducing a method / product solution, it means that the corresponding product / method solution also explicitly includes that technical feature.

[0043] The description of the existence or inclusion of a step, module, or feature at any location in this invention does not imply that such existence is exclusive or unique. Those skilled in the art can obtain other embodiments by supplementing the technical solutions disclosed in this invention with other technical means. Based on the key points described in the specific embodiments of this invention, those skilled in the art can substitute, delete, add, combine, or rearrange certain technical features to obtain a technical solution that still follows the concept of this invention. These solutions that do not depart from the technical concept of this invention are also within the protection scope of this invention.

[0044] refer to Figure 1 The system showcases a drone aerial photography system 100. This system 100 includes a drone 102 and a target detection system 101 mounted on the drone 102. The target detection system 101 includes an image acquisition system 103 and a target detection device 104. The target detection device 104 is configured to perform a target detection method based on images captured by the image acquisition system 103 during the flight of the drone 102.

[0045] refer to Figure 2 The diagram illustrates the structure of the target detection device 104. The target detection device 104 receives input signals, such as image information captured by the target acquisition system 103, and sends them to the neural network inference device 201. This neural network inference device 201 can be a GPU, AI chip, or other device specifically designed to run neural networks. The neural network inference device 201 deploys a neural network 201-n, which is used to infer the input signals based on the input signals and trained network parameters, obtaining an inference result 202. The output device 203 acquires the inference result 202 and outputs it. The output device 203 can output the inference result in any reasonable manner, such as wireless communication or storage in the drone's local storage medium.

[0046] Figure 3The diagram shows the structure of the neural network 201-n deployed on the neural network inference device 201. The neural network 202-n includes a backbone network 301, a neck network 302, and a head network 303.

[0047] Figure 4 This displays a diagram illustrating network training and deployment. Figure 2 The neural network 201-n deployed in the neural network inference device 201 shown contains a large number of network configuration parameters, which need to be obtained through training. Therefore, Figure 4 The neural network training system 400 includes a neural network inference device 201', which has the same network topology as the aforementioned neural network inference device 201. The neural network inference device 201' receives batches of data samples from the dataset sequentially and performs an inference process based on the samples to obtain an inference result 202'. If the inference result 202' is inconsistent with the label of the data sample, the adjustment device 401 adjusts (403) the network configuration parameters in the neural network deployed by the neural network inference device 201' according to the error. The above steps are repeated until all training samples are used for training. The aforementioned adjustment process can be any known scheme, such as backpropagation of error, gradient descent, cosine annealing decay, etc.

[0048] After obtaining the trained network configuration parameters, the network configuration parameters in the neural network inference device 201' are mapped and deployed to the aforementioned neural network inference device 201. The network deployment device 201 thus obtained can be used in the aforementioned unmanned aerial vehicle target detection system 101.

[0049] For the dataset used before training, such as the VisDrone dataset, this invention employs one or more of the following processing steps, which are then used for the aforementioned actual training to address motion blur, complex backgrounds, and complex lighting conditions in aerial human-machine images, thereby improving recognition accuracy:

[0050] 1) Set the probability of flipping vertically to 0;

[0051] 2) Mosaic processing, the specific steps are as follows: select four images; scale, flip and change the color gamut of the four images, then overlap a portion of the four images after the above operations and stitch them together to generate a large new image; then mark the target anchor box on the new image;

[0052] 3) Two-dimensional Gaussian blur processing, the specific steps are as follows: for the pixel to be processed, perform two-dimensional Gaussian blur processing using the pixel values ​​around the pixel, and then use the obtained value as the pixel value of the point.

[0053] The probabilities of the two-dimensional Gaussian distribution used in two-dimensional Gaussian blurring are as follows:

[0054]

[0055] The Fourier transform spectrum of the Gaussian function is single-lobed, and the filter width is positively correlated with the smoothness, determined by the parameter σ. In this paper, the value of σ is set to be random from 1 to 100 during network training, and the probability of performing Gaussian blur operation is 0.3.

[0056] 4) Due to factors such as weather, lighting, and environmental conditions, the brightness of the same target can vary greatly during drone aerial photography. Therefore, the dataset has undergone gamma transformation processing, and the transformation formula is as follows:

[0057] S = cr γ

[0058] Where s is the grayscale output, c is the scaling factor, r is the input grayscale value (between 0 and 1), and γ is the grayscale scaling parameter, which is the main parameter controlling the degree of grayscale transformation scaling. When γ changes, the image will have the following effects: when γ < 1, the contrast of the low grayscale areas of the image will gradually increase; when γ = 1, the image remains unchanged; when γ > 1, the contrast of the high grayscale areas of the image will gradually increase. In this paper, the value of γ is set to be randomized from 1 to 5 during network training, and the probability of performing gamma transformation processing is 0.5.

[0059] After the above processing, at least the following technical effects can be achieved:

[0060] 1) It helps improve the algorithm's ability to detect small targets in complex backgrounds and improves the detection accuracy in diverse backgrounds of images captured in UAV aerial photography application scenarios;

[0061] 2) Adapt to images that are unevenly lit, night scenes, or blurry due to the relative motion between the drone and ground objects during drone shooting;

[0062] 3) It helps improve the algorithm's target detection capability under different light and dark conditions.

[0063] refer to Figure 5 The paper presents the network topology of a 201-n neural network. This 201-n neural network includes the following structure: an input module, a backbone network, a neck network, and a head network; wherein the input module includes at least data augmentation, image transformation, and adaptive anchor boxes.

[0064] In this invention, the backbone network comprises the following modules connected in sequence: Cov2d, B1_3, B2_3, B3_4, B4_3, B4_2. The conv2d module contains sequentially connected ordinary convolutions, batch normalization (BN), and hard-swish activation function operations. The convolutions in the Bneck modules (B1_3, B2_3, B3_4, B4_3, B4_2) are depthwise separable convolutions (DWConv) and DWConv-H using the hard-swish activation function.

[0065] Specifically, B_1_3 consists of three sequentially connected modules: ordinary convolution, BN, SiLU, DWConv, BN, SiLU, ordinary convolution, and BN; B_2_3 consists of three sequentially connected modules: ordinary convolution, BN, SiLU, CBAM, DWConv, BN, SiLU, CBAM, ordinary convolution, and BN; B_3_4 consists of four sequentially connected modules: Conv2d, DWConv-H, ordinary convolution, and BN; B_4_3 consists of three sequentially connected modules: Conv2d, DWConv-H, CBAM, ordinary convolution, and BN; B_4_2 consists of two sequentially connected modules: Conv2d, DWConv-H, CBAM, ordinary convolution, and BN; and DWConv is a depthwise separable convolution, DWConv-H is a sequentially connected module: DWConv, BN, and Hard-Swish, and BN is batch normalization.

[0066] This backbone network has the following main advantages:

[0067] 1) The linear bottleneck-shaped residual inversion structure effectively expands the input features. Unlike traditional residual structures with many channels at both ends and few in the middle, the residual inversion structure has many channels in the middle part. This module first uses 1×1 Conv convolution for dimensionality increase, followed by 3×3 DWConv convolution, while traditional residual structures such as ResNet use general Conv convolution in this part.

[0068] 2) Introducing an attention module mechanism. SENet uses squeezing and excitation operations to suppress and emphasize information, thereby obtaining the weights of feature channels to determine their importance. The specific process is: global pooling, dimensionality reduction through two fully connected layers, and then output through a loss function.

[0069] The residual inversion module and the SE attention module are applicable to most convolutional neural networks and can effectively improve the detection accuracy of the network with almost no impact on the network size.

[0070] 3) Employs Depthwise Separable Convolution (DWconv). Unlike typical convolutional blocks, the DWconv convolution in the Bneck module is a combination of depthwise convolution and pointwise convolution, which can significantly reduce the computational cost and parameter count of the network.

[0071] In this invention, the neck network and the head network are the neck network and head network of the YOLOv5 network. This part is a structure known to those skilled in the art and will not be described in detail here.

[0072] Benefiting from DWconv convolution, the number of parameters and the size of the neural network after the improvement are much smaller than the original YOLOv5 network. At the same time, the improvement of using CBAM attention mechanism, SiLU activation function and replacing pooling layer avoids redundant and ineffective convolution and pooling operations, which also improves the detection accuracy of the network, reduces the amount of computation, and improves speed and stability.

[0073] Compared to the original network, this invention significantly reduces the amount of computation and parameters, and its lightweight design makes it suitable for deployment on embedded computing platforms for drones with limited computing power.

[0074] Further, refer to Figure 6 The paper discloses the aforementioned CBAM attention mechanism structure diagram. The CBAM attention mechanism includes two sub-modules, CAM and SAM. It determines the attention region by evaluating the combined importance of both the image's channels and spatial orientation, in order to suppress irrelevant background information or emphasize the information of the target to be detected, which can effectively improve the accuracy of target detection.

[0075] This invention selects the CBAM module. Compared with the prior art, the CBAM module adds a spatial attention mechanism on the basis of the channel attention mechanism, which is better than the SE module in terms of the accuracy of image attention processing. This is because if only the channel weight is considered, the weight of pixels at different coordinates in the same channel will be different, which makes it impossible to accurately determine the importance of the pixel.

[0076] refer to Figure 6 The input feature map is processed by the CAM and SAM modules to evaluate the channel weights and obtain the attention map. The input feature map is multiplied by the output attention map to optimize the features, highlighting the features of the image, namely the location and content of the target to be detected. This makes the network pay more attention to the target detected in the attention region, thereby improving the target recognition ability of the algorithm.

[0077] In the channel attention module, the input feature map (H×W×C) is processed by global max pooling and global average pooling based on length and width, respectively, to extract and compress spatial information, generating two different feature maps of size 1×1×C. These are then fed into a two-layer neural network with one hidden layer for computation. The parameters in this network are shared. The first layer has C / r (channel reduction rate, r = 16) neurons, which are then activated by ReLU before passing through a second layer with C neurons. Finally, the output feature maps are element-wise summed and concatenated, then activated by a Hard-sigmoid function to generate the final feature map (1×1×C) input to the spatial attention module.

[0078] In the spatial attention module, the feature maps output by the CAM module are compressed and merged using channel-based max pooling and average pooling to obtain two feature maps of size H×W×1. Then, a concat operation is performed, followed by a 7×7 convolutional kernel operation to reduce the dimensionality and obtain the spatial attention feature map (H×W×1).

[0079] CBAM's global max pooling and global average pooling operations in the channel attention module complement each other, effectively extracting compressed information. The spatial attention module uses 7×7 convolutions instead of the traditional multiple 3×3 convolutions, effectively increasing the receptive field and obtaining better spatial information. Therefore, replacing the SE attention module with the CBAM module allows the network to find and accurately detect targets more quickly and effectively. This attention mechanism can address the low accuracy problem encountered when detecting aerial images with complex backgrounds and overlapping occlusions of small targets by UAVs.

[0080] Furthermore, the SiLU activation function used in this invention is shown in the following specific expression.

[0081] SiLU = x·sigmoid(x)

[0082]

[0083] SiLU′=SiLU+(I-SiLU)·sigmoid(x)

[0084] Compared to existing technologies, the aforementioned backbone network based on this activation function has better stability, allowing the neural network to converge faster and achieve better training results. At the same time, the small number of parameters added to SiLU can effectively improve the accuracy of the neural network.

[0085] Furthermore, as mentioned earlier, there are significant differences between drone aerial photography footage and mainstream object detection datasets. To address these issues, this invention proposes a method using K-means clustering to calculate the object bounding boxes in the VisDrone training set, specifically as follows:

[0086] Step (1): Set 9 anchor boxes of random size as the initial prior boxes;

[0087] Step (2): Assign the anchor box sample of each target to the prior box that is most similar to it;

[0088] Step (3): Calculate the mean of the height and width of the prior box again and set this as the new prior box;

[0089] Step (4): Repeat steps (2) and (3) until the size of the prior box is fixed or the maximum number of iterations is reached.

[0090] In a preferred embodiment, the small-scale prior boxes used to detect small targets have sizes of (2,6), (4,13), and (8,21), the medium-scale prior boxes have sizes of (8,10), (14,33), and (17,16), and the large-scale prior boxes have sizes of (24,51), (32,27), and (55,68). In this invention, the improved prior boxes enable the trained network to more accurately detect a large number of small targets in UAV aerial images.

[0091] In summary, compared with the YOLOv5-large network, the technical solution of this invention achieves a 148% improvement in detection efficiency (FPS) at the cost of a 17.4% reduction in mAP, and with slightly better detection performance than YOLOv5s, the network size is only 60% of that of YOLOv5s.

[0092] Although the invention has been described with reference to specific features and embodiments, various modifications, combinations, and substitutions can be made therein without departing from the invention. The scope of protection of this invention is not limited to the specific embodiments of processes, machines, manufactures, material compositions, apparatuses, methods, and steps described in the specification, and these methods and modules may also be implemented in one or more related, interdependent, cooperative, or upstream / downstream products or methods.

[0093] Therefore, the specification and drawings should be simply regarded as a description of some embodiments of the technical solutions defined by the appended claims, and thus the appended claims should be interpreted in accordance with the principle of the greatest reasonable interpretation, and are intended to cover as much as possible all modifications, variations, combinations or equivalents within the scope of the invention, while avoiding unreasonable interpretations.

[0094] The technical features mentioned in the appended claims may have alternative technical features, or the order of certain technical processes or material organization may be rearranged. Those skilled in the art, upon learning of this invention, will readily conceive of these alternative means, or alter the order of the technical processes or material organization, and then employ substantially the same means to solve substantially the same technical problems and achieve substantially the same technical effects. Therefore, even if the claims explicitly define the aforementioned means and / or order, these modifications, alterations, and substitutions should all fall within the scope of protection of the claims based on the principle of equivalents.

[0095] The method steps or modules described in the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the steps and components of each embodiment have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application or design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered outside the scope of protection claimed by this invention.

Claims

1. A target detection system (101) for drone aerial photography scenarios, characterized in that: It includes an image acquisition system (103) and a target detection device (104), the target detection device being configured to perform a target detection method based on images captured by the image acquisition system during the flight of the UAV; The target detection method is as follows: the target detection device receives the image information captured by the target acquisition system and sends it to the neural network inference device (201). The neural network inference device is equipped with a neural network, which is used to infer the input signal based on the input signal and the trained network parameters, obtain the inference result, and output it through the output device (203). The neural network comprises a backbone network (301), a neck network (302), and a head network (303); among which, The backbone network includes the following modules connected in sequence: Cov2d, B_1_3, B_2_3, B_3_4, B_4_3, B_4_2; wherein, the Conv2d module contains ordinary convolution, BN, and Hard-swish activation function operations connected in sequence. B_1_3 consists of three sequentially connected convolutional operations: ordinary convolution, BN, SiLU, DWConv, BN, SiLU, ordinary convolution, and BN. B_2_3 consists of three sequentially connected convolutional operations: ordinary convolution, BN, SiLU, DWConv, BN, SiLU, CBAM, ordinary convolution, and BN. B_3_4 consists of four sequentially connected Conv2d, DWConv-H, ordinary convolution, and BN; B_4_3 consists of three sequentially connected Conv2d, DWConv-H, CBAM, ordinary convolution, and BN; B_4_2 consists of two sequentially connected Conv2d, DWConv-H, CBAM, ordinary convolution, and BN; where DWConv is a depthwise separable convolution; DWConv-H is a sequentially connected DWConv, BN, and Hard-Swish module; BN is batch normalization; and CBAM is the CBAM attention mechanism module, which includes channel attention and spatial attention modules. The neck network and head network mentioned are the neck network and head network of the YOLOv5 neural network; The neural network deployed on the neural network inference device is a neural network obtained after training on a dataset, wherein the dataset is subjected to the following operations: 1) Set the probability of vertical flipping to 0; 2) Mosaic processing: Select four images; scale, flip, and change the color gamut of the four images; then overlap a portion of the four images after the above operations and stitch them together to generate a large new image; then mark the target anchor box on the new image; 3) Two-dimensional Gaussian blur processing: For the pixel to be processed, perform two-dimensional Gaussian blur processing using the pixel values ​​around the pixel, and then use the resulting value as the pixel value of that point; 4) Gamma transform processing: The transformation formula is as follows: s=cr γ Where s is the grayscale output, c is the scaling factor, r is the input grayscale, its value is between 0 and 1, and γ is the grayscale scaling parameter, which controls the degree of grayscale transformation scaling. The dataset is the VisDrone training set. The target bounding boxes of the VisDrone training set are calculated using the K-means clustering method to obtain the target bounding boxes. Specifically, the steps are as follows: (1) Set 9 anchor boxes of random size as the initial prior boxes; (2) Assign each target's anchor box sample to the prior box that is most similar to it; (3) Calculate the mean of the height and width of the prior box again and set this as the new prior box; (4) Repeat steps (2) and (3) until the size of the prior box is fixed or the maximum number of iterations is reached.

2. The target detection system in a drone aerial photography scene according to claim 1, characterized in that: The small-scale prior box sizes used to test small targets are (2,6), (4,13), and (8,21), the medium-scale prior box sizes are (8,10), (14,33), and (17,16), and the large-scale prior box sizes are (24,51), (32,27), and (55,68).

3. A drone aerial photography system, characterized in that, The drone aerial photography system includes: a drone (102), and a target detection system (101) mounted on the drone, wherein the target detection system is the target detection system in the drone aerial photography scene as described in any one of claims 1 to 2.

Images