An RT-DETR-based unmanned aerial vehicle small target detection method and system

Abstract

The application belongs to the technical field of target detection, and discloses a UAV small target detection method and system based on RT-DETR. The method comprises the following steps: performing feature extraction on image data based on different scale feature layers to obtain original features corresponding to each scale feature layer; inputting the original features of different scales into a local convolution branch to correspondingly obtain original local features of each scale; inputting the original features of different scales into an ELGCA branch to correspondingly obtain recalibrated local features of each scale; respectively performing corresponding splicing on each original local feature and each recalibrated local feature according to the same scale to obtain splicing features of each scale, and respectively inputting the splicing features into a CGLU gating unit to obtain output features of each scale; and inputting the output features into an FPN-PAN feature fusion module to obtain fusion features, and then obtaining a target detection result. The application can effectively detect small targets.

Description

Technical Field

[0001] This invention relates to the field of target detection technology, specifically to a method and system for detecting small targets on unmanned aerial vehicles (UAVs) based on RT-DETR. Background Technology

[0002] With the widespread application of drones, and their advantages of high deployment flexibility, strong mobility and wide field of view, they have important application prospects in the field of target detection.

[0003] In the target detection process based on unmanned aerial vehicles (UAVs), image data is first acquired using sensors such as cameras and radar mounted on the UAV. Then, a detection algorithm module, either mounted on the UAV or deployed in the cloud, processes the image data. The detection algorithm module is crucial for ensuring the accuracy and precision of target detection. RT-DETR is an end-to-end target detection architecture that offers flexible algorithm deployment, strong real-time detection capabilities, and effective detection of targets at multiple scales, making it one of the mainstream algorithms used in UAV target detection modules. Specifically, in the RT-DETR architecture, multi-scale feature extraction is first performed. After feature interaction and fusion based on a hybrid encoder, the data sequentially passes through an IoU (Inter-Original Value) perceptual query module and a Transformer decoder before entering the prediction head for classification and regression processing, ultimately yielding the prediction result.

[0004] However, in the RT-DETR architecture, although the FPN-PAN feature fusion module is introduced into the hybrid encoder to address the problem of small targets being easily occluded and ignored due to the large target scale span, the FPN-PAN feature fusion module only uses multi-scale features obtained by simple extraction by the C2f module as input, and it only uses simple convolution stacking when performing feature fusion. This results in limited ability to jointly model cross-scale context and fine-grained texture. Ultimately, the RT-DETR architecture still suffers from significant false negatives and false negatives in small target detection. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for detecting small targets from unmanned aerial vehicles (UAVs) based on RT-DETR, so as to solve the technical problems of high false negative and false positive rates in the detection of small targets in remote sensing data acquired by UAVs in the prior art.

[0006] To achieve the above objectives, the present invention proposes the following technical solution: Firstly, a method for small target detection on UAVs based on RT-DETR is provided, including: Acquire image data captured by the sensor, and extract features from it sequentially based on feature layers of different scales to obtain the original features corresponding to each feature layer of each scale; Original features at different scales are input into the local convolution branch to obtain the corresponding original local features for each scale; at the same time, original features at different scales are input into the ELGCA branch to obtain the corresponding recalibrated local features for each scale; wherein, the local convolution branch includes a 3×3 depthwise separable convolutional layer, and the ELGCA branch includes a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers and a nonlinear activation layer in sequence; Each original local feature and each recalibrated local feature are concatenated at the same scale to obtain concatenated features corresponding to each scale, and then input into the CGLU gate unit to obtain output features corresponding to each scale. Output features at different scales are input into the FPN-PAN feature fusion module to perform bidirectional cross-layer feature fusion, thereby obtaining fused features; wherein, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module; The fused features are sequentially input into the IoU perception query module and the prediction head module to obtain the target detection results corresponding to the image data.

[0007] Furthermore, the step of inputting the original features at different scales into the ELGCA branch to obtain the corresponding recalibrated local features for each scale includes: The original features at different scales are input into a 3×3 convolutional layer to obtain the local response corresponding to each scale; Local responses at different scales are input into a global average pooling layer to obtain a global description vector corresponding to each scale; The global description vectors at different scales are processed sequentially using two 1×1 convolutional layers and a nonlinear activation layer to generate channel attention weights corresponding to each scale. The original features at different scales are recalibrated to obtain the recalibrated local features corresponding to each scale: ; in, F loc For local response, s For channel attention weights, This indicates channel-by-channel multiplication.

[0008] Furthermore, the step of inputting output features of different scales into the FPN-PAN feature fusion module for bidirectional cross-layer feature fusion includes: Calculate the equivalent convolution weights and equivalent biases for the 1×1 convolution branch, 3×3 convolution branch, and identity branch during the training phase, respectively. Based on the equivalent convolution weights and equivalent biases, the 1×1 convolution branch, 3×3 convolution branch and identity branch are structurally reparameterized and fused to obtain a target 3×3 convolutional layer. Specifically, the 1×1 convolution branch is equivalent to the 3×3 convolution by filling the center of the kernel with weights and padding the rest with zeros; the identity branch is equivalent to the 3×3 convolution by constructing a unit convolution kernel at the diagonal position and performing BN fusion.

[0009] Furthermore, the step of sequentially inputting the fused features into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data includes: The fused features are input into the IoU-aware query module to obtain enhanced query features based on the EIoU loss function; wherein, the query features include overlap information, center position information, and aspect ratio information; The query enhancement features are input into the prediction head module for regression and classification processing, and the category and bounding box of each target in the image data are output.

[0010] Furthermore, including: Before the local convolutional branches and ELGCA branches of the original features input at different scales, they are processed by channel compression using 1×1 convolutional layers; at the same time, the outputs of each CGLU gated unit are processed by channel restoration using 1×1 convolutional layers to obtain output features corresponding to different scales.

[0011] Furthermore, acquiring image data based on sensor capture includes: The original image is resized according to the preset dimensions and the principle of minimizing black borders. The resized original image is subjected to image normalization processing to obtain a standard image; Standard images are converted to tensor format image data.

[0012] Secondly, a small target detection system for UAVs based on RT-DETR is provided, including: The raw extraction module is used to acquire image data captured by the sensor and extract features from it sequentially based on feature layers of different scales to obtain the original features corresponding to each feature layer of each scale. The dual-branch extraction module is used to input the original features at different scales into the local convolutional branch to obtain the original local features corresponding to each scale; at the same time, it inputs the original features at different scales into the ELGCA branch to obtain the recalibrated local features corresponding to each scale; wherein, the local convolutional branch includes a 3×3 depthwise separable convolutional layer, and the ELGCA branch includes a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers, and a non-linear activation layer in sequence; The feature stitching module is used to stitch each original local feature and each recalibrated local feature at the same scale to obtain stitched features corresponding to each scale, and input them into the CGLU gate unit to obtain output features corresponding to each scale. The feature fusion module is used to input output features at different scales into the FPN-PAN feature fusion module for bidirectional cross-layer feature fusion, thereby obtaining fused features; wherein, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module; The target detection module is used to sequentially input the fused features into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data.

[0013] Furthermore, the feature fusion module includes: The computation unit is used to calculate the equivalent convolution weights and equivalent biases of the 1×1 convolution branch, 3×3 convolution branch, and identity branch during the training phase. The reparameterization unit is used to perform structural reparameterization fusion of 1×1 convolutional branches, 3×3 convolutional branches and identity branches to obtain a target 3×3 convolutional layer. Specifically, the 1×1 convolution branch is equivalent to the 3×3 convolution by filling the center of the kernel with weights and padding the rest with zeros; the identity branch is equivalent to the 3×3 convolution by constructing a unit convolution kernel at the diagonal position and performing BN fusion.

[0014] Thirdly, an electronic device is provided, including at least one processor coupled to a memory storing a computer program configured to be executed by the processor when run.

[0015] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement the method.

[0016] Beneficial effects: As can be seen from the above technical solutions, the technical solution of the present invention provides a small target detection method for UAVs based on RT-DETR to solve the defects of various current target detection algorithms in small target detection.

[0017] First, in the feature extraction stage, considering the smaller spatial receptive field of high-resolution feature maps, which can preserve the precise boundary coordinates and local structural details of the target, this approach has a significant advantage in detecting small targets. Therefore, to improve the overall accuracy of small target detection by UAVs while suppressing the attenuation of local details and the loss of small target information caused by downsampling, this technical solution introduces both local convolutional branches and ELGCA branches in the high-resolution feature extraction module of the backbone network. At this point, local edge and texture information can be extracted through the local convolutional branch, while the ELGCA branch explicitly encodes local details and global context simultaneously, thus providing a more discriminative representation for subsequent small target discrimination. Furthermore, a CGLU gating unit is introduced to suppress redundant channels and highlight key information relevant to the target during the fusion of the two branches.

[0018] Secondly, in the feature fusion stage, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module to enhance multi-scale feature representation by using a multi-branch convolutional structure and a reparameterization strategy, thereby reducing the false detection rate of small targets.

[0019] Ultimately, based on the above scheme, a more accurate detection of various small targets was achieved.

[0020] It should be understood that all combinations of the foregoing concepts and the additional concepts described in more detail below can be considered part of the inventive subject matter of this disclosure, provided that such concepts do not contradict each other.

[0021] The foregoing and other aspects, embodiments, and features of the teachings of the present invention will be more fully understood from the following description in conjunction with the accompanying drawings. Other additional aspects of the invention, such as features and / or beneficial effects of exemplary embodiments, will become apparent from the following description or may be learned through practice of specific embodiments according to the teachings of the present invention. Attached Figure Description

[0023] The accompanying drawings are not intended to be drawn to scale. In the drawings, each identical or nearly identical component shown in the various figures may be denoted by the same reference numeral. For clarity, not every component is labeled in each figure. Embodiments of various aspects of the invention will now be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the UAV small target detection method based on RT-DETR described in this embodiment; Figure 2 Here is a flowchart of the process for acquiring the image data; Figure 3 A flowchart for recalibrating local features; Figure 4This is a flowchart illustrating the folding process for the RepC3 module; Figure 5 This is a flowchart of the IoU-aware query module. Figure 6 This is a structural block diagram of the UAV small target detection system based on RT-DETR described in this embodiment; Figure 7 This is a structural block diagram of the electronic device described in this embodiment. Detailed Implementation

[0024] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used herein should have the ordinary meaning understood by those skilled in the art to which this invention pertains.

[0025] The terms "first," "second," and similar words used in this application specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, unless the context clearly indicates otherwise, the singular forms of "an," "a," or "the," etc., do not indicate a quantity limitation, but rather indicate the presence of at least one. Terms such as "comprising" or "including" mean that the element or object preceding "comprising" encompasses the features, wholes, steps, operations, elements, and / or components listed following "comprising" or "including," and do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components, and / or sets thereof. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0026] In UAV target detection, the detection algorithm is crucial for ensuring accuracy and precision. Detection algorithms based on the RT-DETR architecture are widely used in UAV target detection due to their ability to effectively detect targets at multiple scales. However, they suffer from limitations in high-resolution feature extraction and the core hybrid encoder's limited ability to jointly model cross-scale context and fine-grained texture. Therefore, they still exhibit significant false negatives and missed detections in small target detection. This embodiment aims to provide an RT-DETR-based UAV small target detection method to address these technical problems.

[0027] The following section, with reference to the accompanying drawings, provides a detailed description of the UAV small target detection method based on RT-DETR described in this embodiment.

[0028] The method described in this embodiment is run in the following configuration environment: an Intel® Xeon(R) Gold 5218R CPU@ 2.10GHz processor with 64GB of RAM, an NVIDIA A10 graphics card, and PyTorch 2.8.0 deep learning framework. Based on this, combined with Figure 1 As shown, the method includes the following steps: Step S202: Acquire image data based on sensor capture, and extract features from it sequentially based on feature layers of different scales to obtain the original features corresponding to each scale feature layer.

[0029] In this embodiment, the feature layers of different scales are specifically: the C3 feature layer, the C4 feature layer, and the C5 feature layer. To facilitate feature extraction from each feature layer, combined with... Figure 2 As shown, the image data is obtained through processing in the following manner: Step S20202: Adjust the size of the original image according to the preset size and the principle of minimizing black borders.

[0030] Step S20204: Perform image standardization processing on the original image after size adjustment to obtain a standard image.

[0031] Step S20206: Convert the standard image to obtain tensor format image data.

[0032] At this point, based on steps S20202 to S20206, dimensionality mismatch caused by inconsistent tensor sizes in subsequent network layers can be avoided, numerical distribution differences can be eliminated, and the convergence speed of the corresponding model can be improved. Ultimately, high-quality image data that meets practical needs can be obtained.

[0033] Step S204: Input the original features at different scales into the local convolution branch to obtain the original local features corresponding to each scale; at the same time, input the original features at different scales into the ELGCA branch to obtain the recalibrated local features corresponding to each scale.

[0034] The local convolutional branch described in this step includes a 3×3 depthwise separable convolutional layer for extracting local edge and texture information. The ELGCA branch is used to explicitly encode local details and global context simultaneously on the high-resolution feature map, providing a more discriminative representation for subsequent small object discrimination. Based on this, it is configured to sequentially include a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers, and a non-linear activation layer. Correspondingly, combined with... Figure 3 As shown, the ELGCA branch processes the original features at different scales as follows: Step S20402: Input the original features at different scales into a 3×3 convolutional layer to obtain the local response corresponding to each scale.

[0035] Step S20404: Input the local responses at different scales into the global average pooling layer to obtain the global description vector corresponding to each scale.

[0036] Step S20406: The global description vectors at different scales are processed sequentially using two 1×1 convolutional layers and a nonlinear activation layer to generate channel attention weights corresponding to each scale.

[0037] Step S20408: Recalibrate the original features at different scales to obtain recalibrated local features corresponding to each scale.

[0038] At this point, the recalibrated local features obtained in steps S20402 to S20408 are: ; in, F loc For local response, s For channel attention weights, This indicates channel-by-channel multiplication.

[0039] At this point, steps S20402 to S20408 can alleviate the interference of complex backgrounds on small target detection.

[0040] Continuing, to further suppress redundant channels and highlight key information relevant to the target, the following channel splicing operation is performed: Step S206: Perform corresponding splicing of each original local feature and each recalibrated local feature at the same scale to obtain spliced ​​features corresponding to each scale, and input them into the CGLU gate unit to obtain output features corresponding to each scale.

[0041] At this point, the CGLU gating unit in step S206 can adaptively suppress channels that are irrelevant to or redundant with small targets, and enhance the significant response related to small targets of the UAV. Steps S204 to S206 enhance the contour cues and spatial position signals of small targets in the high-resolution layer, which helps to improve the recall rate and positioning accuracy of small targets; it also solves the problem that small targets occupy few pixels, so it is difficult to maintain their discriminativeness after multiple downsampling in traditional local convolution.

[0042] As a preferred implementation, to reduce overhead, the original features at different scales are processed by channel compression using 1×1 convolutional layers before entering the local convolutional branches and ELGCA branches; at the same time, the outputs of the CGLU gated units are processed by channel restoration using 1×1 convolutional layers to obtain output features corresponding to different scales.

[0043] Step S208: Input the output features at different scales into the FPN-PAN feature fusion module to perform bidirectional cross-layer feature fusion, thereby obtaining fused features.

[0044] In this embodiment, a RepC3 module is also introduced at any fusion node of the FPN-PAN feature fusion module, thereby enhancing multi-scale feature representation through the multi-branch convolutional structure and reparameterization strategy of the RepC3 module. Specifically, combined with Figure 4 As shown, during the inference process, the RepC3 module trained through multiple branches is folded as follows: Step S20802: Calculate the equivalent convolution weights and equivalent biases of the 1×1 convolution branch, 3×3 convolution branch and identity branch during the training phase, respectively.

[0045] Specifically, the equivalent convolution weights are: ; in, γ i , σ i All of these are BN parameters for any branch; K i represents the original convolution weights for any branch.

[0046] The equivalent bias is: ; in, b i Let be the original bias for any branch. µ i , β i These are all BN parameters for any branch.

[0047] Step S20804: Based on each equivalent convolution weight and equivalent bias, perform structural reparameterization fusion on the 1×1 convolution branch, 3×3 convolution branch and identity branch to obtain a target 3×3 convolutional layer.

[0048] The target 3×3 convolutional layer is: ; in, K rep For reparameterized weights, pad() means padding a 1×1 convolution kernel or identity kernel with zeros to expand it to a 3×3 size; The equivalent convolution weights for the 3×3 convolution branches, The equivalent convolution weights for a 1×1 convolution branch, The equivalent convolution weights for the identity branches; b rep For reparameterized bias, ; This is the equivalent bias for the 3×3 convolution branch. The equivalent bias for a 1×1 convolution branch. This is the equivalent bias for the identity branch.

[0049] At this point, based on steps S20802 to S20804, training with multiple branches and inference with a single branch can be achieved, which reduces inference overhead and improves detection efficiency while ensuring inference accuracy.

[0050] Step S210: Input the fused features sequentially into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data.

[0051] As a preferred implementation method, combined with Figure 5 As shown, step S210 further includes: Step S21002: Input the fused features into the IoU-aware query module to obtain enhanced query features based on the EIoU loss function.

[0052] The query features include overlap information, center position information, and aspect ratio information.

[0053] Step S21004: Input the query enhancement features into the prediction head module for regression and classification processing, and output the category and bounding box of each target in the image data.

[0054] At this point, based on steps S21002 to S21004, since the EIoU loss function is introduced in the IoU perception query module, the bounding box of small targets can be accurately measured, thereby effectively optimizing the detection accuracy of small targets.

[0055] In summary, this embodiment discloses a UAV target detection scheme based on RT-DETR. It extracts features through a dual-branch approach using local convolutional and ELGCA branches in parallel, and introduces a CGLU gating unit for branch fusion to enhance cross-scale semantic aggregation, thereby achieving saliency modeling for small targets. A RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module to significantly improve the quality of fused features without increasing inference overhead. Furthermore, an EIoU loss function is introduced to guide the training process to focus more on the localization error of extremely small targets, thereby improving the detection rate of small targets.

[0056] The aforementioned program can run in a processor or be stored in memory (or a computer-readable storage medium). Computer-readable media includes both permanent and non-permanent, removable and non-removable media, and information storage can be achieved by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include temporary computer-readable media, such as modulated data signals and carrier waves.

[0057] These computer programs may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes can be implemented using different modules, and different steps can be implemented using different modules.

[0058] This embodiment also provides a small target detection system for UAVs based on RT-DETR. Combined with... Figure 6 As shown, it includes: The raw extraction module is used to acquire image data captured by the sensor and extract features from it sequentially based on feature layers of different scales to obtain the original features corresponding to each feature layer of each scale.

[0059] The dual-branch extraction module is used to input the original features at different scales into the local convolutional branch to obtain the original local features corresponding to each scale; at the same time, it inputs the original features at different scales into the ELGCA branch to obtain the recalibrated local features corresponding to each scale; wherein, the local convolutional branch includes a 3×3 depthwise separable convolutional layer, and the ELGCA branch includes a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers and a nonlinear activation layer in sequence.

[0060] The feature stitching module is used to stitch each original local feature and each recalibrated local feature at the same scale to obtain stitched features corresponding to each scale, and input them into the CGLU gated unit to obtain output features corresponding to each scale.

[0061] The feature fusion module is used to input output features of different scales into the FPN-PAN feature fusion module for bidirectional cross-layer feature fusion, thereby obtaining fused features; wherein, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module.

[0062] The target detection module is used to sequentially input the fused features into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data.

[0063] Since the system is built based on the method described above, the points already explained will not be repeated here.

[0064] For example, the feature fusion module includes: The computation unit is used to calculate the equivalent convolution weights and equivalent biases of the 1×1 convolutional branches, 3×3 convolutional branches, and identity branches during the training phase.

[0065] The reparameterization unit is used to perform structural reparameterization fusion of 1×1 convolutional branches, 3×3 convolutional branches and identity branches to obtain a target 3×3 convolutional layer.

[0066] Specifically, the 1×1 convolution branch is equivalent to the 3×3 convolution by filling the center of the kernel with weights and padding the rest with zeros; the identity branch is equivalent to the 3×3 convolution by constructing a unit convolution kernel at the diagonal position and performing BN fusion.

[0067] Furthermore, in combination Figure 7 As shown, an electronic device is also provided, including at least one processor coupled to a memory storing a computer program configured to be executed by the processor when run.

[0068] Additionally, a computer-readable storage medium is provided having a computer program stored thereon, the computer program being executed by a processor to implement the method.

[0069] Since the system, electronic devices, and storage media are all used to implement the method or are built based on the method, they have significant technical advantages over other existing detection algorithms in terms of detection rate and accuracy when performing small target detection on UAVs. They are more suitable for detecting small and extremely small targets in complex real-world scenarios.

[0070] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for detecting small targets on a UAV based on RT-DETR, characterized in that, include: Acquire image data captured by the sensor, and extract features from it sequentially based on feature layers of different scales to obtain the original features corresponding to each feature layer of scale; Original features at different scales are input into the local convolution branch to obtain the corresponding original local features for each scale; at the same time, original features at different scales are input into the ELGCA branch to obtain the corresponding recalibrated local features for each scale; wherein, the local convolution branch includes a 3×3 depthwise separable convolutional layer, and the ELGCA branch includes a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers and a nonlinear activation layer in sequence; Each original local feature and each recalibrated local feature are concatenated at the same scale to obtain concatenated features corresponding to each scale, and then input into the CGLU gate unit to obtain output features corresponding to each scale. Output features at different scales are input into the FPN-PAN feature fusion module to perform bidirectional cross-layer feature fusion, thereby obtaining fused features; wherein, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module; The fused features are sequentially input into the IoU perception query module and the prediction head module to obtain the target detection results corresponding to the image data.

2. The UAV small target detection method based on RT-DETR according to claim 1, characterized in that, The step of inputting the original features at different scales into the ELGCA branch to obtain the corresponding recalibrated local features for each scale includes: The original features at different scales are input into a 3×3 convolutional layer to obtain the local response corresponding to each scale; Local responses at different scales are input into a global average pooling layer to obtain a global description vector corresponding to each scale; The global description vectors at different scales are processed sequentially using two 1×1 convolutional layers and a nonlinear activation layer to generate channel attention weights corresponding to each scale. The original features at different scales are recalibrated to obtain the recalibrated local features corresponding to each scale: ; in, F loc For local response, s For channel attention weights, This indicates channel-by-channel multiplication.

3. The UAV small target detection method based on RT-DETR according to claim 1, characterized in that, The step of inputting output features of different scales into the FPN-PAN feature fusion module for bidirectional cross-layer feature fusion includes: Calculate the equivalent convolution weights and equivalent biases for the 1×1 convolution branch, 3×3 convolution branch, and identity branch during the training phase, respectively. Based on the equivalent convolution weights and equivalent biases, the 1×1 convolution branch, 3×3 convolution branch and identity branch are structurally reparameterized and fused to obtain a target 3×3 convolutional layer. Specifically, the 1×1 convolution branch is equivalent to the 3×3 convolution by filling the center of the kernel with weights and padding the rest with zeros; the identity branch is equivalent to the 3×3 convolution by constructing a unit convolution kernel at the diagonal position and performing BN fusion.

4. The UAV small target detection method based on RT-DETR according to claim 1, characterized in that, The step of sequentially inputting the fused features into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data includes: The fused features are input into the IoU-aware query module to obtain enhanced query features based on the EIoU loss function; wherein, the query features include overlap information, center position information, and aspect ratio information; The query enhancement features are input into the prediction head module for regression and classification processing, and the category and bounding box of each target in the image data are output.

5. The UAV small target detection method based on RT-DETR according to claim 1, characterized in that, include: Before the local convolutional branches and ELGCA branches of the original features input at different scales, they are processed by channel compression using 1×1 convolutional layers; at the same time, the outputs of each CGLU gated unit are processed by channel restoration using 1×1 convolutional layers to obtain output features corresponding to different scales.

6. The UAV small target detection method based on RT-DETR according to claim 1, characterized in that, The acquisition of image data based on sensor capture includes: The original image is resized according to the preset dimensions and the principle of minimizing black borders. The resized original image is subjected to image normalization processing to obtain a standard image; Standard images are converted to tensor format image data.

7. A small target detection system for unmanned aerial vehicles based on RT-DETR, characterized in that, include: The raw extraction module is used to acquire image data captured by the sensor and extract features from it sequentially based on feature layers of different scales to obtain the raw features corresponding to each feature layer of each scale. The dual-branch extraction module is used to input the original features at different scales into the local convolutional branch to obtain the original local features corresponding to each scale; at the same time, it inputs the original features at different scales into the ELGCA branch to obtain the recalibrated local features corresponding to each scale; wherein, the local convolutional branch includes a 3×3 depthwise separable convolutional layer, and the ELGCA branch includes a 3×3 convolutional layer, a global average pooling layer, two 1×1 convolutional layers, and a non-linear activation layer in sequence; The feature stitching module is used to stitch each original local feature and each recalibrated local feature at the same scale to obtain stitched features corresponding to each scale, and input them into the CGLU gate unit to obtain output features corresponding to each scale. The feature fusion module is used to input output features at different scales into the FPN-PAN feature fusion module for bidirectional cross-layer feature fusion, thereby obtaining fused features; wherein, a RepC3 module is introduced at any fusion node of the FPN-PAN feature fusion module; The target detection module is used to sequentially input the fused features into the IoU perception query module and the prediction head module to obtain the target detection result corresponding to the image data.

8. The UAV small target detection system based on RT-DETR according to claim 7, characterized in that, The feature fusion module includes: The computation unit is used to calculate the equivalent convolution weights and equivalent biases of the 1×1 convolution branch, 3×3 convolution branch and identity branch during the training phase. The reparameterization unit is used to perform structural reparameterization fusion of 1×1 convolutional branches, 3×3 convolutional branches and identity branches to obtain a target 3×3 convolutional layer. Specifically, the 1×1 convolution branch is equivalent to the 3×3 convolution by filling the center of the kernel with weights and padding the rest with zeros; the identity branch is equivalent to the 3×3 convolution by constructing a unit convolution kernel at the diagonal position and performing BN fusion.

9. An electronic device, characterized in that, It includes at least one processor coupled to a memory storing a computer program configured to be executed by the processor to perform the method of any one of claims 1-6.

10. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which is executed by a processor to implement the method of any one of claims 1-6.

Images