Aerial photography vehicle re-identification method and system based on any angle detection

Abstract

The invention provides an aerial vehicle re-identification method and system based on any angle detection, and the method comprises the steps: inputting a to-be-identified aerial image after preprocessing into a trained vehicle re-identification model, and obtaining a vehicle re-identification result; wherein the vehicle re-identification model comprises a position code fusion module, a multi-head attention module and a re-identification module based on any angle detection; the position code fusion module is used for carrying out vehicle detection on the input image and extracting preliminary vehicle features; the multi-head attention module is used for capturing multi-scale features of the preliminary vehicle features in different postures and orientations; and the re-identification module based on any angle detection is used for determining a vehicle any angle detection frame based on the multi-scale features, and performing cosine similarity calculation on the interested vehicle and the target vehicle in the vehicle any angle detection frame to obtain a vehicle re-identification result. According to the invention, angle detection is carried out on the multi-attitude and multi-orientation vehicle, so that the accuracy of re-identification of vehicles with different attitudes is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle re-identification, and in particular to a method and system for aerial vehicle re-identification based on arbitrary angle detection. Background Art

[0002] Re-identifying vehicle targets in aerial scene images has broad application prospects. However, this process faces many challenges. Among them, the vehicle targets themselves are small and their postures are diverse, making it very difficult to accurately capture the detailed features of the vehicle. For example, the vehicle posture will change accordingly due to changes in the shooting angle; the vehicle's turning during driving will also cause its posture to change; and when the shooting angle is high, the pixel area occupied by the vehicle target in the image will be smaller. These situations increase the difficulty of accurately identifying vehicle targets in aerial scene images, thereby hindering the re-identification of vehicle targets.

[0003] Based on these problems, traditional detection and recognition networks focus on extracting obvious features, and can distinguish and identify vehicle features such as color, general outline, and partial body structure to a certain extent. However, the traditional detection and recognition process lacks sufficient sensitivity and adaptability to changes in the characteristic direction and posture of vehicles caused by various factors in actual scenarios. There is a high rate of false detection and missed detection, which makes it difficult to meet the complex and changing needs of traffic monitoring. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a method and system for aerial vehicle re-identification based on arbitrary angle detection, which performs angle detection on vehicles with multiple postures and directions, and introduces a channel attention mechanism that integrates position information to enhance the attention weight on vehicle targets; at the same time, an axially compressed multi-head attention fusion process is designed to enhance the global semantic features; finally, the final vehicle re-identification is performed on vehicle targets at any angle based on posture correction and cosine similarity, thereby improving the accuracy of vehicle re-identification with different postures.

[0005] In order to achieve the above object, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides an aerial vehicle re-identification method based on arbitrary angle detection, comprising:

[0007] Obtain the aerial image to be identified for preprocessing;

[0008] The pre-processed aerial image to be identified is input into the trained vehicle re-identification model to obtain the vehicle re-identification result;

[0009] Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

[0010] Preferably, the preprocessing includes: using data enhancement methods such as randomly flipping images horizontally, randomly flipping images vertically, and randomly rotating images to unify the size of the aerial images to be identified.

[0011] Preferably, the position encoding fusion module is used to detect vehicles on the input image and extract preliminary vehicle features, specifically including:

[0012] The position encoding fusion module includes a feature extraction and encoding unit, a feature fusion and enhancement unit, and a channel attention fusion unit;

[0013] The aerial image to be identified is input into the feature extraction and encoding unit, a multi-scale pyramid feature map is extracted, and a multi-scale feature with position enhancement is obtained based on the added target vehicle position encoding information;

[0014] The multi-scale pyramid feature map is adjusted to a uniform size and input into the feature fusion and enhancement unit together with the position-enhanced multi-scale feature to obtain a multi-scale feature map with position encoding enhancement;

[0015] The multi-scale feature map enhanced by position coding is input into the channel attention fusion unit. Channel association is performed based on global average pooling, two 1*1 convolution operations, and S function to obtain the channel attention weight features. Then, it is merged with the multi-scale feature map enhanced by position coding to obtain the position coding fused channel attention feature map, i.e., the preliminary vehicle features.

[0016] Preferably, after the channel attention weight feature is obtained, it is merged with the multi-scale feature map enhanced by position coding, specifically: the channel attention weight feature is divided into 4 layers, and multiplied with the pooled multi-scale feature map enhanced by position coding at the channel level to obtain the position coding fused channel attention feature map.

[0017] Preferably, the multi-head attention module is used to capture multi-scale features of preliminary vehicle features in different postures and orientations, specifically including:

[0018] The multi-head attention module includes a 3*3 convolutional layer, three posture compression channels and a multi-head attention unit;

[0019] The position encoding fusion channel attention feature map is input into the 3*3 convolution layer to obtain Q features, K features and V features, which are respectively input into three posture compression channels to extract posture orientation features in the horizontal axis, channel direction and vertical axis, and input into the multi-head attention unit to obtain a fusion feature map converged on one axis;

[0020] The position encoding fusion channel attention feature map and the fusion feature are multiplied and fused to obtain the multi-scale feature.

[0021] Preferably, the Q feature, K feature and V feature are input into three posture compression channels respectively to extract posture orientation features in the horizontal axis, channel direction and vertical axis, specifically including:

[0022] On the horizontal axis, the Q, K, and V feature maps of size H×W×C are aggregated according to the mean values ​​of Q, K, and V in the vertical dimension in the horizontal direction to obtain {Q h ,K h ,V h};

[0023] On the vertical axis, the Q, K, and V feature maps of size H×W×C are aggregated according to the mean values ​​of Q, K, and V along the horizontal dimension in the vertical axis direction to obtain {Q v ,K v ,V v};

[0024] In the channel direction, Q, K, and V with a size of H×W×C are aggregated according to the mean values ​​of Q, K, and V in the horizontal and vertical dimensions in the channel direction to obtain {Q c ,K c ,V c}.

[0025] Preferably, determining the vehicle arbitrary angle detection frame based on multi-scale features specifically includes:

[0026] The re-identification module based on arbitrary angle detection includes two branches. The first branch is used to obtain a heat map of the input image after the input multi-scale features are sequentially subjected to deconvolution, convolution and S-function units, and preliminarily obtain the axis and long and short side information of the arbitrary angle detection frame according to the heat map result; the second branch is used to generate a prediction of five information (x, y, w, h, θ) in the arbitrary angle detection result after the input multi-scale features pass through two convolution units, where (x, y) is the center point, and (w, h, θ) is the width, height and angle of the horizontal detection frame;

[0027] According to (x, y, w, H, θ) and the axis and long and short side information obtained by the first branch, the final vehicle arbitrary angle detection frame is determined.

[0028] In a second aspect, the present invention provides an aerial vehicle re-identification system based on arbitrary angle detection, comprising:

[0029] A data acquisition module is used to acquire the aerial images to be identified for preprocessing;

[0030] The re-identification module is used to input the pre-processed aerial image to be identified into the trained vehicle re-identification model to obtain the vehicle re-identification result;

[0031] Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

[0032] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the aerial vehicle re-identification method based on arbitrary angle detection described in the first aspect.

[0033] In a fourth aspect, the present invention provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps in the method for aerial vehicle re-identification based on arbitrary angle detection described in the first aspect are implemented.

[0034] Compared with the prior art, the present invention has the following beneficial effects:

[0035] (1) In view of the difficulties of the diversity of postures, orientations, and sizes of vehicles in aerial traffic scenes, the present invention designs a vehicle detection algorithm at any angle to cope with the diverse postures and orientations of vehicles to improve the detection accuracy of vehicles in different orientations.

[0036] (2) In the arbitrary-angle vehicle detection algorithm, the present invention designs a channel attention module that integrates position coding to enhance the attention weight of the detection network to vehicle information; at the same time, it designs a multi-head attention module that integrates axial compression to enhance the global semantic features of the network and reduce the amount of computation.

[0037] (3) Based on the detection results of the vehicle at any angle, the present invention designs a vehicle re-identification method based on posture correction and cosine similarity, performs posture calibration to reduce the impact of the diversity of detected vehicle postures, and uses cosine similarity to determine the vehicle re-identification result based on the final feature vector.

[0038] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings constituting a part of the present invention are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their description are used to explain the present invention but do not constitute a limitation of the present invention.

[0040] Figure 1 A main flow chart of an aerial vehicle re-identification method based on arbitrary angle detection provided by an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of the structure of a vehicle re-identification model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0043] Embodiment 1

[0044] like Figure 1 As shown, this embodiment discloses a method for re-identifying aerial vehicles based on arbitrary angle detection, comprising the following steps:

[0045] S1: Obtain the aerial image to be identified and pre-process it;

[0046] S2: Input the preprocessed aerial image to be identified into the trained vehicle re-identification model to obtain the vehicle re-identification result;

[0047] Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

[0048] Next, combine Figure 2, a method for aerial vehicle re-identification based on arbitrary angle detection disclosed in this embodiment is described in detail.

[0049] 1) Position encoding fusion module

[0050] The position coding fusion module is specifically a channel attention module that integrates position coding. The position coding fusion module includes a feature extraction and encoding unit, a feature fusion and enhancement unit, and a channel attention fusion unit.

[0051] In the feature extraction and encoding unit, the aerial images of the vehicle are scaled to different scales and then input into the ResNet34 network to extract multi-scale features to obtain a multi-scale pyramid feature map {P 2 ,P 3 ,P 4 ,P 5}.

[0052] In order to cope with the detection network's ability to detect small target vehicles, artificial position encoding information L is added to the multi-scale pyramid feature map. p , in order to improve the sensitivity of the position information in the feature map, and at the same time increase the weight of the spatial position information of the vehicle target at different scales, so as to increase the attention level of the target vehicle to be identified. Then, the transformer-based attention mechanism is enhanced, and the scales are normalized and spliced ​​together to obtain the multi-scale feature F with position enhancement. p .

[0053] Among them, the artificial position coding is generated based on the sine-cosine coding method with pixels at each position in the feature map. This method calculates sine and cosine based on the coding dimension and pixel position information to obtain the position coding.

[0054] The scale of the multi-scale pyramid feature map is uniformly scaled to P 4 Size, in the feature fusion and enhancement unit, will be scaled to P 4 The feature map and position-enhanced multi-scale feature F p Multiply and fuse them to obtain a multi-scale feature map with enhanced position encoding {F 2 ,F 3 ,F 4 ,F 5}, and {F 2 ,F 3 ,F 4 ,F 5 The feature maps of the same size in these four layers are channel-merged to obtain a channel-merged multi-scale feature map.

[0055] In the channel attention fusion unit, the multi-scale feature map merged by the channel is sequentially subjected to global average pooling, two 1*1 convolution operations, and S function channel association to obtain the channel attention weight feature F c At the same time, the feature map after global average pooling is combined with the channel attention weight feature F c Perform multiplication and fusion to obtain the position encoding fusion channel attention feature map F a , where the fusion method is to layer four blocks of F c Multiply it with the pooling feature map at the channel level.

[0056] 2) Multi-head attention module

[0057] The multi-head attention module is specifically an axial compression module based on multi-head attention. In the multi-head attention module, the feature map F after position enhancement and channel attention fusion is a After 3*3 convolution operations in sequence, three feature maps are obtained: Q feature, K feature and V feature. These three feature maps are used for compression in the horizontal axis direction, vertical axis direction and channel direction respectively. The compression process is to gather global information on one axis.

[0058] The specific process is to aggregate the Q, K, and V feature maps of size H×W×C on the horizontal axis according to the mean values ​​of Q, K, and V along the vertical dimension in the horizontal axis direction to obtain {Q h ,K h ,V h}; On the vertical axis, the Q, K, and V feature maps of size H×W×C are aggregated according to the mean values ​​of Q, K, and V in the horizontal dimension in the vertical axis direction to obtain {Q v ,K v ,V v}; In the channel direction, Q, K, and V with a size of H×W×C are aggregated according to the mean values ​​of Q, K, and V in the horizontal and vertical dimensions in the channel direction to obtain {Q c ,K c ,V c}.

[0059] All nine feature maps {Q h ,K h ,V h ,Q v ,K v ,V v ,Q c ,K c ,V c} Perform calculations based on the multi-head attention mechanism to obtain three feature maps of size (C, H, 1), (C, 1, W), and (C, 1, 1). Then, add and fuse the three feature maps to obtain the feature map F based on channel compression attention. z .

[0060] The position encoding is fused with the channel attention feature map F a Compared with the channel compression and multi-head attention mechanism to extract global features, the channel compression attention feature map F z After multiplication and fusion, the multi-head attention axial compression feature map F is finally obtained e , used for arbitrary angle detection.

[0061] In this embodiment, a multi-head attention mechanism is used. This mechanism can focus on different feature subspaces when processing feature maps. By aggregating information in different directions according to the means of Q, K, and V in turn, it can effectively extract the key features of the vehicle, thereby better capturing the characteristics of the vehicle in different postures and orientations; by performing feature compression and information aggregation in three directions: the horizontal axis, the channel direction, and the vertical axis, it can effectively process the diverse postures and orientations of the vehicle, enhance the global semantic feature extraction of the network, and improve the detection accuracy.

[0062] 3) Re-identification module based on arbitrary angle detection

[0063] Compress the multi-head attention axis feature map F e Further operations are performed through two branches. The first branch obtains the heat map of the input image after passing through deconvolution, convolution and S-function units in sequence. The axis and long and short side information of the arbitrary angle detection frame are obtained based on the heat map results, and the general outline and position relationship of the vehicle are perceived as a whole; the second branch is used to generate predictions of five information (x, y, w, h, θ) in the arbitrary angle detection results after passing through two convolution units, including the center point (x, y), the width and height (w, h) of the horizontal detection frame, and the angle θ, so as to accurately capture the detailed features of the vehicle, especially with high sensitivity to angle changes. The final arbitrary angle detection frame is determined based on (x, Y, w, h, θ) and the axis and long and short side information obtained by the first branch.

[0064] This embodiment comprehensively covers the target's posture and orientation information from macro to micro. The overall outline provided by the heat map provides a basis for positioning detailed information such as angles, while precise parameters such as angles can correct possible deviations in the heat map, greatly improving the sensitivity of the detection frame to information such as the target's posture and orientation, making detection more accurate and reliable.

[0065] Based on the final arbitrary angle detection results, all detected vehicles are corrected based on their angle information. After extracting features through multi-layer convolution operations, a 128-dimensional feature vector V is obtained through the fully connected layer. d The posture correction process is as follows: the image of the detected target detection frame at any angle is rotated based on the angle information obtained, and the target image after posture correction is obtained based on the new vertex positions mapped based on the vertex position information of the rotated detection frame and the rotated image.

[0066] The target vehicle to be re-identified also undergoes the same posture calibration and multiple convolutional layers to obtain the feature vector V r ; By calculating the cosine similarity of the feature vectors of the detected vehicle and the target vehicle to be re-identified, the detected vehicle with the smallest similarity is used as the re-identification result of the vehicle to be re-identified; the cosine similarity calculation process is to calculate the cosine value of the angle between the two vectors, and use this cosine value as the similarity value.

[0067] For the loss function of the arbitrary angle detection part, based on the feature map F e Two loss functions are designed based on the heat maps obtained from the two detection branches and the five parameters of OBB (x, y, w, H, θ): the first is the cross entropy loss function L based on the heat map h ; The second is the OBB loss function L o , L o It consists of a regression loss on (x,y,w,h) and a slope loss obtained from θ.

[0068] The final loss function L of the arbitrary angle detection network is:

[0069] L=r 1 ·L h +r 2 ·L o

[0070] Among them, r 1 and r 2 is the weight coefficient of the two-part loss function.

[0071] The re-identification module based on arbitrary angle detection in this embodiment effectively solves the difficulties of vehicle posture diversity, orientation diversity, and size diversity in aerial traffic scenes through its complex detection and re-identification mechanism, thereby improving the accuracy of vehicle detection and re-identification.

[0072] In order to verify the effect of this embodiment, the following implementation method is proposed:

[0073] This embodiment is based on the self-collected UAV aerial traffic scene re-identification dataset UARD (UAV Aerial Re-Identification Dataset). The UARD dataset includes 65,365 images, including about 300 targets to be identified, and the training set and validation set are divided into 8:2.

[0074] The re-identification model is trained and tested based on the constructed UARD dataset, and the mAP and Top-1 indicators are used to measure the test results: mAP represents the average precision of the number of correctly detected targets, and Top-1 represents the accuracy of re-identification sample ID judgment in the re-identification task.

[0075] The experimental results show that the mAP value of the ADR-Net model designed in this embodiment is 94.4%, and the Top-1 value is 93.5%. According to these two evaluation index data, it can be seen that the ADR-Net model has achieved good results in the test of vehicle re-identification in aerial traffic scenes.

[0076] Embodiment 2

[0077] This embodiment provides an aerial vehicle re-identification system based on arbitrary angle detection, including:

[0078] A data acquisition module is used to acquire the aerial images to be identified for preprocessing;

[0079] The re-identification module is used to input the pre-processed aerial image to be identified into the trained vehicle re-identification model to obtain the vehicle re-identification result;

[0080] Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

[0081] Embodiment 3

[0082] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps in the aerial vehicle re-identification method based on arbitrary angle detection as described in the first embodiment above are implemented.

[0083] Embodiment 4

[0084] This embodiment provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the steps in the aerial vehicle re-identification method based on arbitrary angle detection as described in the first embodiment above are implemented.

[0085] The steps or modules involved in the above embodiments 2 to 4 correspond to those in embodiment 1. For the specific implementation, please refer to the relevant description of embodiment 1. The term "computer-readable storage medium" should be understood as a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0086] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for aerial vehicle re-identification based on arbitrary angle detection, characterized in that: include: Obtain the aerial image to be identified for preprocessing; The pre-processed aerial image to be identified is input into the trained vehicle re-identification model to obtain the vehicle re-identification result; Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

2. The method for aerial vehicle re-identification based on arbitrary angle detection according to claim 1, characterized in that: The preprocessing includes: using data enhancement methods of randomly flipping images horizontally, randomly flipping images vertically, and randomly rotating images to unify the size of the aerial images to be identified.

3. The method for aerial vehicle re-identification based on arbitrary angle detection according to claim 1, characterized in that: The position encoding fusion module is used to detect vehicles on the input image and extract preliminary vehicle features, specifically including: The position encoding fusion module includes a feature extraction and encoding unit, a feature fusion and enhancement unit, and a channel attention fusion unit; The aerial image to be identified is input into the feature extraction and encoding unit, a multi-scale pyramid feature map is extracted, and a multi-scale feature with position enhancement is obtained based on the added target vehicle position encoding information; The multi-scale pyramid feature map is adjusted to a uniform size and input into the feature fusion and enhancement unit together with the position-enhanced multi-scale feature to obtain a multi-scale feature map with position encoding enhancement; The multi-scale feature map enhanced by position coding is input into the channel attention fusion unit. Channel association is performed based on global average pooling, two 1*1 convolution operations, and S function to obtain the channel attention weight features. Then, it is merged with the multi-scale feature map enhanced by position coding to obtain the position coding fused channel attention feature map, i.e., the preliminary vehicle features.

4. The method for aerial vehicle re-identification based on arbitrary angle detection as claimed in claim 3, characterized in that: After the channel attention weight feature is obtained, it is merged with the multi-scale feature map enhanced by position coding. Specifically, the channel attention weight feature is divided into 4 layers, and multiplied with the multi-scale feature map enhanced by pooled position coding at the channel level to obtain the position coding fused channel attention feature map.

5. The method for aerial vehicle re-identification based on arbitrary angle detection according to claim 1, characterized in that: The multi-head attention module is used to capture the multi-scale features of preliminary vehicle features in different postures and orientations, including: The multi-head attention module includes a 3*3 convolutional layer, three posture compression channels and a multi-head attention unit; The position encoding fusion channel attention feature map is input into the 3*3 convolution layer to obtain Q features, K features and V features, which are respectively input into three posture compression channels to extract posture orientation features in the horizontal axis, channel direction and vertical axis, and input into the multi-head attention unit to obtain a fusion feature map converged on one axis; The position encoding fusion channel attention feature map and the fusion feature are multiplied and fused to obtain the multi-scale feature.

6. The method for aerial vehicle re-identification based on arbitrary angle detection as claimed in claim 5, characterized in that: The Q feature, K feature and V feature are input into three posture compression channels respectively to extract posture orientation features in the horizontal axis, channel direction and vertical axis, specifically including: On the horizontal axis, the Q, K, and V feature maps of size H×W×C are aggregated according to the mean values ​​of Q, K, and V in the vertical dimension in the horizontal direction to obtain {Q h ,K h ,V h }; On the vertical axis, the Q, K, and V feature maps of size H×W×C are aggregated according to the mean values ​​of Q, K, and V along the horizontal dimension in the vertical axis direction to obtain {Q v ,K v ,V v }; In the channel direction, Q, K, and V with a size of H×W×C are aggregated according to the mean values ​​of Q, K, and V in the horizontal and vertical dimensions in the channel direction to obtain {Q c ,K c ,V c }.

7. The method for aerial vehicle re-identification based on arbitrary angle detection as claimed in claim 1, characterized in that: The determining of the vehicle arbitrary angle detection frame based on the multi-scale features specifically includes: The re-identification module based on arbitrary angle detection includes two branches. The first branch is used to obtain a heat map of the input image after the input multi-scale features are sequentially subjected to deconvolution, convolution and S-function units, and preliminarily obtain the axis and long and short side information of the arbitrary angle detection frame according to the heat map result; the second branch is used to generate a prediction of five information (x, y, w, h, θ) in the arbitrary angle detection result after the input multi-scale features pass through two convolution units, where (x, y) is the center point, and (w, h, θ) is the width, height and angle of the horizontal detection frame; According to (x, y, w, H, θ) and the axis and long and short side information obtained by the first branch, the final vehicle arbitrary angle detection frame is determined.

8. An aerial vehicle re-identification system based on arbitrary angle detection, characterized in that: include: A data acquisition module is used to acquire the aerial images to be identified for preprocessing; The re-identification module is used to input the pre-processed aerial image to be identified into the trained vehicle re-identification model to obtain the vehicle re-identification result; Among them, the vehicle re-identification model includes a position coding fusion module, a multi-head attention module and a re-identification module based on arbitrary angle detection; the position coding fusion module is used to perform vehicle detection on the input image and extract preliminary vehicle features; the multi-head attention module is used to capture the multi-scale features of the preliminary vehicle features in different postures and orientations; the re-identification module based on arbitrary angle detection is used to determine the vehicle arbitrary angle detection frame based on multi-scale features, and perform cosine similarity calculation between the vehicle of interest and the target vehicle in the vehicle arbitrary angle detection frame to obtain the vehicle re-identification result.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the aerial vehicle re-identification method based on arbitrary angle detection as described in any one of claims 1 to 7 are implemented.

10. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the steps of the aerial vehicle re-identification method based on arbitrary angle detection as described in any one of claims 1-7 are implemented.

Images