A method and system for detecting a target of a complex low-altitude scene by using a UAV

Abstract

The application belongs to the technical field of target detection, and provides a UAV target detection method and system for complex low-altitude scenes, which comprises the following steps: acquiring UAV aerial images of complex low-altitude scenes; extracting multi-layer basic features of the images; wherein the multi-layer basic features comprise shallow high-resolution basic features, middle-low-layer basic features, middle-high-layer basic features and deep-layer basic features; inputting the shallow high-resolution basic features and the middle-low-layer basic features into a low-altitude small target detail injection unit to obtain small target detail enhancement features; inputting the small target detail enhancement features, the middle-low-layer basic features, the middle-high-layer basic features and the deep-layer basic features into a low-altitude scene interference gate unit to obtain multi-layer detection features; inputting the multi-layer detection features into a cross-scale target consistency reconstruction module to obtain multi-scale reconstruction detection features; inputting the multi-scale reconstruction detection features into a detection decoder to obtain a detection result, thereby completing the UAV target detection for complex low-altitude scenes.

Description

Technical Field

[0001] This invention belongs to the field of target detection technology, specifically relating to a method and system for UAV target detection in complex low-altitude scenarios. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the development of applications such as low-altitude economy, intelligent transportation, maritime inspection, urban security and emergency rescue, low-altitude aircraft such as drones need to perform visual perception of targets such as vehicles, pedestrians and ships in complex low-altitude scenarios, detect and identify target categories, bounding box positions and confidence levels from aerial images, and realize autonomous inspection, low-altitude supervision and intelligent analysis of drones.

[0004] Compared to ordinary ground images, complex low-altitude aerial images exhibit greater variations in perspective and background complexity. During drone photography, altitude, pitch angle, lighting, motion, and shooting distance can all change, resulting in significant size differences for the same type of target within the image. Distant vehicles, pedestrians, or ships typically occupy only a few pixels in the image, with weaker target edges, contours, and local textures. Simultaneously, water reflections, road textures, building edges, shoreline structures, shadows, and areas where land and sea blend easily produce local responses similar to the target, increasing the risk of false detections.

[0005] Existing target detection methods are widely applicable in general image target detection, but they still face problems such as small target scale, strong background interference, large scale variation and unstable target response in complex low-altitude aerial images. Summary of the Invention

[0006] To address the aforementioned issues, this invention proposes a UAV target detection method and system for complex low-altitude scenarios. It employs a Low-altitude Target Enhancement and Reconstruction Detection Network (LTERD-Net) to preserve small target details while suppressing interference from complex backgrounds. This method performs consistent reconstruction of multi-scale detection features, solving the problems of insufficient small target detail representation, strong non-target response interference from complex backgrounds, and inconsistent multi-scale detection feature representation in existing UAV target detection methods for complex low-altitude scenarios. This improves the stability of feature representation during the detection process.

[0007] According to some embodiments, the first solution of the present invention provides a UAV target detection method for complex low-altitude scenarios, which adopts the following technical solution: A method for UAV target detection in complex low-altitude scenarios includes: Acquire drone aerial images of complex low-altitude scenes (I); Based on the acquired images and the UAV target detection network, complete the UAV target detection for complex low-altitude scenarios; The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network. During UAV target detection, a basic feature extraction network extracts multi-layer basic features from the acquired image. These extracted multi-layer basic features include shallow high-resolution basic features B1, mid-to-low-level basic features B2, mid-to-high-level basic features B3, and deep basic features B4. The shallow high-resolution basic features B1 and mid-to-low-level basic features B2 are input into the Low-Altitude Small Target Detail Injection (LSDI) unit to obtain small target detail enhancement feature D1. The small target detail enhancement feature D1, mid-to-low-level basic features B2, mid-to-high-level basic features B3, and deep basic features B4 are input into the Low-Altitude Scene Interference Gating (LSIG) unit to obtain interference-gated multi-layer detection features G1, G2, G3, and G4. The multi-layer detection features G1, G2, G3, and G4 are input into the Cross-Scale Target Consistency Reconstruction (CTCR) module to obtain multi-scale reconstruction detection features R1, R2, and R3. The multi-scale reconstruction detection features R1 and R2 are then... The R3 input detector head or detector decoder is used to obtain the detection result Y, which includes the target category, bounding box position and confidence level, thus completing the UAV target detection for complex low-altitude scenarios.

[0008] As a further technical limitation, the low-altitude small target detail injection unit includes a shallow detail adaptation branch, a semantic adaptation branch, a detail injection gating branch, and an output integration branch. The shallow detail adaptation branch is used to extract local details and scale-adapt shallow high-resolution basic features to obtain a shallow detail response. The semantic adaptation branch is used to perform channel adaptation and semantic constraint expression on mid-to-low-level basic features to obtain semantic constraint features. The detail injection gating branch is used to generate detail injection gating coefficients based on the semantic constraint features to obtain a global semantic response. The output integration branch is used to inject the gated shallow detail response into the semantic constraint features to obtain small target detail enhancement features.

[0009] As a further technical limitation, the low-altitude scene interference gating unit includes a reference feature adaptation branch, a context projection branch, a context fusion branch, a target response gating branch, a background interference gating branch, and a residual modulation output branch. Specifically, the reference feature adaptation branch selects a corresponding reference feature based on the current output level and adapts it to the target output channel to obtain a reference response. The context projection branch projects small target detail enhancement features, low-to-mid-level basic features, mid-to-high-level basic features, and deep basic features onto channels to obtain multi-layer context features with a unified context channel. The context fusion branch integrates the multi-layer context features after aligning them to the current output scale to obtain context fusion features. The target response gating branch estimates the target region response based on the reference response and the context fusion features, generating target response gating coefficients. The background interference gating branch estimates the background interference response based on the reference response, generating background interference gating coefficients. The residual modulation output branch modulates the reference response based on the target response gating coefficients and the background interference gating coefficients, and outputs the interference gating features of the current level.

[0010] As a further technical limitation, the cross-scale target consistency reconstruction module includes a multi-layer feature projection branch, a scale alignment branch, a cross-scale response description branch, a scale weight generation branch, a weighted reconstruction branch, and an output integration branch; wherein... The multi-layer feature projection branch is used to project multi-layer detection features onto a unified channel dimension; the scale alignment branch is used to upsample or downsample multi-layer projection features to the same spatial scale according to the target output scale; the cross-scale response description branch is used to form a cross-scale target response description based on the aligned multi-layer features; the scale weight generation branch is used to generate scale weights for features of different levels based on the cross-scale target response description; the weighted reconstruction branch is used to perform weighted aggregation of multi-layer aligned features using scale weights to obtain reconstructed features at the current scale; and the output integration branch is used to perform two-stage 3x3 convolution-normalization-activation processing on the reconstructed features to output multi-scale reconstructed detection features.

[0011] As a further technical limitation, the UAV target detection network adopts an end-to-end deep learning monitoring framework. In the process of UAV target detection, the original feature map of the acquired image is extracted by a convolutional neural network based on the residual network; the extracted original feature map is fused and enhanced to obtain an enhanced feature map; the enhanced feature map is analyzed to obtain the target category and bounding box position of the image, thus completing the UAV target detection for complex low-altitude scenarios.

[0012] Furthermore, in the process of fusing and enhancing the extracted original feature maps, a feature fusion path combining top-down and bottom-up approaches is adopted in the feature encoder; the feature encoder includes a spatial intelligent transformer, a streamlined fusion model, and a collaborative spatial context fusion. The enhanced feature map obtained by the target decoder is used to parse it. The interaction between the learnable object query and the enhanced feature map is carried out through a multi-level attention mechanism. The category and bounding box position of all targets in the image are output in parallel through the classification head and the regression head.

[0013] According to some embodiments, the second aspect of the present invention provides a UAV target detection system for complex low-altitude scenarios, employing the following technical solution: A drone target detection system for complex low-altitude scenarios includes: The acquisition module is configured to acquire drone aerial images of complex low-altitude scenes; The detection module is configured to perform UAV target detection in complex low-altitude scenarios based on the acquired images and the UAV target detection network. The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network, and the detection module includes: The feature extraction submodule is configured to extract multi-layer basic features of the acquired image based on the basic feature extraction network; wherein the extracted multi-layer basic features include shallow high-resolution basic features, low-to-medium layer basic features, high-to-medium layer basic features, and deep basic features. The low-altitude small target detail injection submodule is configured to input shallow high-resolution basic features and mid-to-low-level basic features into the low-altitude small target detail injection unit to obtain small target detail enhancement features. The low-altitude scene interference gating submodule is configured to input small target detail enhancement features, low-to-mid-level basic features, mid-to-high-level basic features and deep basic features into the low-altitude scene interference gating unit to obtain multi-layer detection features after interference gating. The cross-scale target consistency reconstruction submodule is configured to input multi-layer detection features into the cross-scale target consistency reconstruction module to obtain multi-scale reconstructed detection features. The detection output submodule is configured to input multi-scale reconstructed detection features into the detection head or detection decoder to obtain detection results including target category, bounding box position and confidence level, thus completing UAV target detection for complex low-altitude scenarios.

[0014] According to some embodiments, a third aspect of the present invention provides a computer-readable storage medium, employing the following technical solution: A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a UAV target detection method for complex low-altitude scenarios as described in the first aspect of the present invention.

[0015] According to some embodiments, the fourth aspect of the present invention provides an electronic device, which adopts the following technical solution: An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the UAV target detection method for complex low-altitude scenarios as described in the first aspect of the present invention.

[0016] According to some embodiments, the fifth aspect of the present invention provides a computer program product, which adopts the following technical solution: A computer program product includes software code, wherein the program in the software code performs the steps of a UAV target detection method for complex low-altitude scenarios as described in the first aspect of the present invention.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention combines shallow high-resolution detail response with mid-to-low-level semantic constraint information through a Low-altitude Small-object Detail Injection Unit (LSDI). By gating target-related details, it enhances the edge, contour, and local texture expression of distant small targets, improves the detail response of low-altitude small targets, and reduces the risk of small targets being weakened by downsampling. This invention uses a Low-altitude Scene Interference Gating Unit (LSIG) to simultaneously estimate the target response and background interference response, and employs residual gating to modulate multi-layer detection features. This helps to reduce the interference of non-target areas such as water surface reflection, road texture, building edges, shoreline structures, and shadows on the detection features, thereby effectively suppressing non-target responses in complex backgrounds.

[0018] This invention employs a Cross-scale Target Consistency Reconstruction Module (CTCR) to perform scale alignment, cross-scale response aggregation, and consistency reconstruction on gated multi-layer features, thereby improving the consistency of target representation in the input features of the detection head at different scales and making multi-scale detection features more suitable for category prediction and bounding box localization. Attached Figure Description

[0019] The accompanying drawings, which form part of this embodiment, are used to provide a further understanding of this embodiment. The illustrative embodiments and their descriptions are used to explain this embodiment and do not constitute an improper limitation of this embodiment.

[0020] Figure 1 This is a flowchart of a UAV target detection method for complex low-altitude scenarios according to Embodiment 1 of the present invention; Figure 2 This is a schematic diagram of the overall structure of the low-altitude target enhancement and reconstruction detection network in Embodiment 1 of the present invention; Figure 3 This is a schematic diagram of the low-altitude small target detail injection unit structure in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram of the low-altitude scene interference gating unit structure in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram of the cross-scale target consistency reconstruction module structure in Embodiment 1 of the present invention; Figure 6 This is a structural block diagram of a UAV target detection system for complex low-altitude scenarios according to Embodiment 2 of the present invention; Figure 7 This is a flowchart of a UAV target detection method for complex low-altitude scenarios according to Embodiment 3 of the present invention; Figure 8 This is a diagram illustrating the overall architecture of the UAV target detection method for complex low-altitude scenarios in Embodiment 3 of the present invention. Figure 9 This is a schematic diagram of the streamlined hybrid structure in Embodiment 3 of the present invention; Figure 10 This is a schematic diagram of the collaborative spatial context fusion structure in Embodiment 3 of the present invention; Figure 11 This is a schematic diagram of the spatial intelligent converter in Embodiment 3 of the present invention. Detailed Implementation

[0021] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0023] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0024] In this invention, terms such as "upper," "lower," "left," "right," "front," "back," "vertical," "horizontal," "side," and "bottom" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used only to facilitate the description of the structural relationships of the various components or elements of this invention and do not specifically refer to any component or element in this invention. They should not be construed as limiting the invention.

[0025] In this invention, terms such as "fixed connection," "connected," and "linked" should be interpreted broadly, indicating a fixed connection, an integral connection, or a detachable connection; a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can determine the specific meaning of these terms in this invention based on the specific circumstances, and they should not be construed as limitations on the invention.

[0026] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0027] Example 1 Embodiment 1 of this invention introduces a method for UAV target detection in complex low-altitude scenarios.

[0028] Existing drone target detection methods in complex low-altitude scenarios have at least the following shortcomings: (1) Insufficient representation of small target details. In UAV aerial images, distant targets usually have a small pixel ratio. After multiple downsampling, the target edges, contours and local texture information are easily weakened. If the detection network mainly relies on deep low-resolution semantic features, it is difficult to fully preserve the spatial details of small targets, which can easily lead to missed detection of distant small targets or inaccurate bounding box localization.

[0029] (2) Complex backgrounds cause strong interference from non-target responses. Water surface reflections, road textures, building edges, shoreline structures, shadows, and mixed land and sea areas may produce responses similar to the target area in the feature space. If existing methods lack a mechanism to distinguish between target responses and background interference responses in the context of low-altitude backgrounds, they are prone to misidentifying non-target areas as targets, leading to an increase in false detections in the detection results.

[0030] (3) Inconsistent feature representation across multiple scales. Shallow features have high spatial resolution but weak semantic representation, while deep features have strong semantic representation but lack spatial detail. Directly concatenating, adding, or conventionally weighting multi-scale features may lead to inconsistent response representations of targets at different scales, affecting the stable prediction of categories and bounding boxes by the detection head.

[0031] Therefore, this embodiment proposes a method such as Figure 1 The UAV target detection method shown is designed for complex low-altitude scenarios and includes: Acquire drone aerial images of complex low-altitude scenes; Based on the acquired images and the UAV target detection network, complete the UAV target detection for complex low-altitude scenarios; The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network. During UAV target detection, a basic feature extraction network extracts multi-layer basic features from the acquired image. These extracted multi-layer basic features include shallow high-resolution basic features, mid-to-low-level basic features, mid-to-high-level basic features, and deep basic features. The shallow high-resolution and mid-to-low-level basic features are input into a low-altitude small target detail injection unit to obtain small target detail enhancement features. The small target detail enhancement features, mid-to-low-level basic features, mid-to-high-level basic features, and deep basic features are input into a low-altitude scene interference gating unit to obtain interference-gated multi-layer detection features. The multi-layer detection features are input into a cross-scale target consistency reconstruction module to obtain multi-scale reconstructed detection features. The multi-scale reconstructed detection features are input into a detection head or detection decoder to obtain detection results including target category, bounding box position, and confidence level, thus completing UAV target detection for complex low-altitude scenes.

[0032] like Figure 2 As shown, this embodiment receives an aerial image I captured by a UAV, and outputs B1, B2, B3, and B4 through a basic feature extraction network. Among them, B1 and B2 are input into LSDI to obtain D1, D1 and B2, B3, and B4 are input into LSIG to obtain G1, G2, G3, and G4, G1, G2, G3, and G4 are input into CTCR to obtain R1, R2, and R3, and finally the detection head or detection decoder outputs the detection result Y.

[0033] In one embodiment, the size of the input image I can be set to 640x640. The spatial scales of B1, B2, B3, and B4 output by the basic feature extraction network are 160x160, 80x80, 40x40, and 20x20, respectively, with 64, 128, 256, and 512 channels, respectively. LSDI outputs D1, with a spatial scale of 80x80 and 128 channels; LSIG outputs G1, G2, G3, and G4; CTCR outputs R1, R2, and R3, where the spatial scales of R1, R2, and R3 are 80x80, 40x40, and 20x20, respectively, and each has 256 channels.

[0034] In one implementation based on PyTorch or other deep learning frameworks, the core processing unit of LTERD-Net can be implemented using the operator-level structure shown in Table 1.

[0035] Table 1 Operator-level structure

[0036] As one or more implementations, the basic feature extraction network is used to convert the input image I into multi-layer basic features. The input image I is a drone aerial image of a complex low-altitude scene, which may contain detection targets such as vehicles, pedestrians, and ships, as well as complex background areas such as water surfaces, roads, buildings, shorelines, and shadows. The basic feature extraction network may consist of convolution, normalization, activation, downsampling, and feature blocks, or it may employ other backbone feature extraction networks suitable for target detection tasks.

[0037] The basic feature extraction network outputs shallow high-resolution basic features B1, mid-to-low-level basic features B2, mid-to-high-level basic features B3, and deep basic features B4. Among them, B1 mainly preserves shallow details such as edges, contours, and textures; B2 has high spatial resolution and contains mid-to-low-level semantic information; B3 contains mid-to-high-level target semantics and contextual information; and B4 contains deep scene semantics and global contextual information.

[0038] In one embodiment, when the input image size is 640x640, the spatial scales of B1, B2, B3, and B4 can be 160x160, 80x80, 40x40, and 20x20, respectively, and the number of channels can be 64, 128, 256, and 512, respectively. The above dimensions and number of channels are merely one implementation method, and those skilled in the art can make equivalent adjustments based on the backbone network type, input image size, and computing resources.

[0039] As one or more implementations, the Low-altitude Small-object Detail Injection Unit (LSDI) is located after the basic feature extraction network and before the low-altitude scene interference gating unit, and is used to generate small object detail enhancement features D1 based on shallow high-resolution basic features and mid-to-low-level semantic features.

[0040] like Figure 3 As shown, the inputs to LSDI include B1 and B2, where B1 is the shallow high-resolution basic feature and B2 is the mid-to-low-level basic feature. The output of LSDI is the small target detail enhancement feature D1. In one embodiment, B1 has a spatial scale of 160x160 and 64 channels, B2 has a spatial scale of 80x80 and 128 channels, and the LSDI output D1 is aligned with B2, with a spatial scale of 80x80 and 128 channels.

[0041] LSDI includes shallow detail adaptation branch, semantic adaptation branch, detail injection gating branch, and output integration branch.

[0042] The shallow detail adaptation branch is used to extract local details and adapt the scale of B1. In one embodiment, the shallow detail adaptation branch may include a two-stage convolution-normalization-activation combination layer. The first stage may use a convolution with a kernel size of 3 and a stride of 2 to perform initial alignment from the B1 scale to the B2 scale. The second stage may use a convolution with a kernel size of 3 and a stride of 1 to further extract local detail responses. The shallow detail response Fd is obtained through this branch, and Fd contains information such as small object edges, contours, and local textures.

[0043] The semantic adaptation branch is used to perform channel adaptation and semantic constraint representation on B2. In one embodiment, the semantic adaptation branch can use a 1x1 convolution-normalization-activation combination layer or other channel mapping operations to convert B2 into semantic constraint features Fs. Fs is used to provide stronger semantic constraints on the target region than B1, avoiding the subsequent detail injection process from indiscriminately introducing irrelevant background textures into the detection features.

[0044] The injection-of-details (IoD) gating branch is used to generate IoD gating coefficients Wd based on the semantic constraint features Fs. In one embodiment, the IoD gating branch may include operations such as adaptive average pooling, 1x1 convolution, and sigmoid activation to extract the global semantic response from Fs and convert it into gating weights acting on the shallow detail response Fd. When the spatial dimensions of Fd and Fs are inconsistent due to variations in the input size or backbone network, bilinear interpolation can be used to align Fd to the spatial dimensions of Fs.

[0045] The output integration branch is used to inject the gated, shallow detail responses into semantically constrained features. In one embodiment, the weighted result of Fd and Wd can be calculated first, then the result can be residually integrated with Fs, and finally D1 can be obtained through a 3x3 convolution-normalization-activation combination layer. This processing logic can be expressed as: D1 = Conv3x3(Fs + Fd × Wd), where Conv3x3 represents the output integration operation consisting of convolution, normalization, and activation with a kernel size of 3. This expression is used to illustrate the processing logic and is not limited to a single mathematical form.

[0046] The steps for processing LSDI are as follows: (1) Input the shallow basic feature B1 into the shallow detail adaptation branch, and obtain the shallow detail response Fd after local convolution and scale adaptation; (2) Input the low-to-mid-level basic features B2 into the semantic adaptation branch to obtain the semantic constraint features Fs; (3) Generate detail injection gating coefficients Wd based on semantic constraint features Fs; (4) Use Wd to perform weighted filtering on Fd to obtain the target-related detailed response; (5) Inject the target-related detail response into Fs, and obtain the small target detail enhancement feature D1 through the output integration branch.

[0047] The output D1 of LSDI is connected to the low-altitude scene interference gating unit LSIG. Through the above structure, LSDI uses the high-resolution detail information of B1 to supplement the semantic features of B2, and at the same time uses the semantic response of B2 to constrain the details in B1 that may belong to pure background texture, thereby providing a more suitable detail enhancement foundation for subsequent background interference suppression and multi-scale reconstruction for small target detection.

[0048] As one or more implementations, the Low-altitude Scene Interference Gating Unit (LSIG) is located after the Low-altitude Small Target Details Injection Unit and before the Cross-Scale Target Consistency Reconstruction Module. It is used to estimate the target response and background interference response based on multi-layered basic features and small target detail enhancement features, and to perform gated modulation on the multi-layered detection features. This module corresponds to the attached... Figure 4 .

[0049] like Figure 4 As shown, the inputs to LSIG include D1, B2, B3, and B4; where D1 is the small target detail enhancement feature, and B2, B3, and B4 are different levels of basic features output by the basic feature extraction network. The outputs of LSIG are multi-layer detection features G1, G2, G3, and G4 after interference gating. In one embodiment, the spatial scales of G1 and G2 are 80x80, the spatial scale of G3 is 40x40, and the spatial scale of G4 is 20x20; G1, G2, G3, and G4 are used to preserve target response and background suppression results at different scales, respectively.

[0050] LSIG includes a baseline feature adaptation branch, a context projection branch, a context fusion branch, a target response gating branch, a background interference gating branch, and a residual modulation output branch.

[0051] The baseline feature adaptation branch is used to select the corresponding baseline feature based on the current output level and adapt the baseline feature to the target output channel. In one embodiment, LSIG can adopt a single output level wrapper structure, where G1, G2, G3, or G4 is generated each time it is executed, determined by the output level parameter out_level. Correspondingly, when out_level is 1, D1 is used as the baseline feature; when out_level is 2, B2 is used as the baseline feature; when out_level is 3, B3 is used as the baseline feature; and when out_level is 4, B4 is used as the baseline feature. The selected baseline feature can be adapted to the target output channel through a 1x1 convolution-normalization-activation combination layer to obtain the baseline response Fb.

[0052] The context projection branch is used to project channels D1, B2, B3, and B4 separately to obtain multi-layer context features with a unified context channel. Since D1, B2, B3, and B4 may differ in spatial scale and number of channels, the four inputs can be mapped to the unified context channel through a 1x1 convolution-normalization-activation combination layer to form a comparable and fusionable context representation. In one embodiment, the number of unified context channels can be set to 128.

[0053] The context fusion branch is used to integrate multi-layer contextual features after aligning them to the current output scale. In one embodiment, bilinear interpolation or other scale alignment operations can be used to adjust the projected features of D1, B2, B3, and B4 to the spatial scale of the current output level, then concatenate them along the channel dimension, and perform context fusion through a 1x1 convolution-normalization-activation combination layer to obtain the context fusion feature Fc. Fc simultaneously contains small object details, local semantics, and deep contextual information.

[0054] The target response gating branch is used to estimate the target region response based on Fb and Fc, generating the target response gating coefficients Wt. In one embodiment, Fb and Fc can be added together and then sequentially passed through a 3x3 convolution-normalization-activation combination layer, a 1x1 convolution, and a sigmoid activation to obtain Wt.

[0055] The background interference gating branch is used to estimate the background interference response based on Fb, generating the background interference gating coefficients Wb. In one embodiment, Fb can be sequentially passed through a 3x3 convolution-normalization-activation combination layer, a 1x1 convolution, and a sigmoid activation to obtain Wb. Wt is used to enhance regions that may contain targets, while Wb is used to characterize non-target responses that may cause interference, such as water surface reflections, road textures, building edges, shoreline structures, and shadows.

[0056] The residual modulation output branch is used to modulate Fb according to Wt and Wb, and outputs the interference gating feature Gk of the current level. In one embodiment, Gk can be obtained by residual gating of Fb, Wt, and Wb, for example, Gk = Conv3x3(Fb × Wt × (1 - Wb) + Fb), where k is the current output level, and Conv3x3 represents the output integration operation consisting of convolution, normalization, and activation with a kernel size of 3. This expression is used to illustrate the processing logic of target response enhancement, background interference suppression, and residual preservation, and is not limited to a single mathematical form.

[0057] The steps for processing LSIG are as follows: (1) Receive D1, B2, B3 and B4; (2) Select the reference feature according to the current output level, and perform channel adaptation on the reference feature to obtain the reference response Fb; (3) Channel projection is performed on D1, B2, B3 and B4 respectively to obtain multi-layer context features; (4) Align the multi-layer context features to the current output scale and then concatenate and fuse them to obtain the context fusion feature Fc; (5) Generate the target response gating coefficient Wt based on Fb and Fc; (6) Generate the background interference gating coefficient Wb based on Fb; (7) Modulate Fb according to Wt and Wb, and output the interference gating feature Gk through residual method.

[0058] In one embodiment, LSIG repeatedly performs the gating calculation described above for different output levels. Each execution uses D1, B2, B3, and B4 as common context inputs and outputs G1, G2, G3, or G4 respectively based on the current output level. Through this structure, LSIG can reduce non-target texture responses in complex low-altitude backgrounds while preserving target features. The residual modulation method also preserves the original baseline features, reducing the risk of target information loss due to over-suppression.

[0059] As one or more implementations, the Cross-scale Target Consistency Reconstruction Module (CTCR) is located after the low-altitude scene interference gating unit and before the detection head or detection decoder. It is used to perform scale alignment, cross-scale response aggregation and consistency reconstruction on the multi-layer detection features after interference gating to obtain multi-scale reconstructed detection features that are adapted to the detection output.

[0060] like Figure 5 As shown, the inputs of CTCR include G1, G2, G3, and G4, which are the multi-layer detection features after interference gating. The outputs of CTCR include R1, R2, and R3, which are the multi-scale reconstructed detection features. In one embodiment, the spatial scales of R1, R2, and R3 are 80x80, 40x40, and 20x20, respectively, and the number of channels can all be set to 256 to adapt to the multi-scale input of the detection head or detection decoder.

[0061] CTCR includes a multi-layer feature projection branch, a scale alignment branch, a cross-scale response description branch, a scale weight generation branch, a weighted reconstruction branch, and an output integration branch.

[0062] The multi-layer feature projection branch is used to project G1, G2, G3, and G4 onto a unified channel dimension, respectively. Since the number of channels and spatial scale of the four gated features may differ, unified channel projection ensures that features from different levels have a consistent intermediate representation dimension during subsequent cross-scale aggregation. In one embodiment, G1, G2, G3, and G4 can be projected onto a unified intermediate channel via a 1x1 convolution-normalization-activation combination layer, and the number of unified intermediate channels can be set to 256.

[0063] The scale alignment branch is used to upsample or downsample multi-layer projection features to the same spatial scale based on the target output scale. In one embodiment, CTCR can employ a single-output hierarchical packaging structure, performing reconstruction calculations for each of the three detection output scales. When out_level is 1, the spatial scale of G1 is used as the target scale to generate R1; when out_level is 2, the spatial scale of G3 is used as the target scale to generate R2; when out_level is 3, the spatial scale of G4 is used as the target scale to generate R3. Each execution receives G1, G2, G3, and G4, and bilinear interpolation can be used to align the four projection features to the current target scale.

[0064] The cross-scale response description branch is used to form a cross-scale target response description based on the aligned multi-layer features. Since targets at different scales respond differently to shallow localization information and deep semantic information, the cross-scale response description is used to express the contribution of each layer of features to target detection at the current output scale.

[0065] The scale weight generation branch is used to generate scale weights for different levels of features based on the cross-scale target response description. In one embodiment, four aligned features can be concatenated along the channel dimension to form descriptive features, and then four scale weights can be generated through adaptive average pooling, 1x1 convolution, and sigmoid activation to adjust the contribution of each layer of features at the current output scale.

[0066] The weighted reconstruction branch is used to perform weighted aggregation of multi-layer alignment features using scale weights to obtain the reconstructed features at the current scale. In one embodiment, the four-way alignment features can be stacked first, and then weighted summation can be performed according to the four-way scale weights. The output integration branch is used to perform two-stage 3x3 convolution-normalization-activation processing on the reconstructed features, outputting R1, R2, or R3.

[0067] The processing steps for CTCR are as follows: (1) Multi-layer detection features G1, G2, G3 and G4 after receiving interference gating; (2) Perform channel projection on G1, G2, G3 and G4 respectively to make them have a uniform number of intermediate channels; (3) Align the projected multi-layer features to the same spatial size according to the current output scale; (4) Generate cross-scale target response descriptions based on the aligned multi-layer features; (5) Generate scale weights for each layer based on the cross-scale target response description; (6) Use scale weights to perform weighted aggregation of multi-layer features to obtain the reconstructed features at the current scale; (7) Generate R1, R2 or R3 by integrating the output branches.

[0068] In one embodiment, CTCR outputs R1, R2, and R3 through three reconstruction calculations at different target scales. R1 focuses on high-resolution detection features, suitable for localization and response preservation of smaller targets; R2 focuses on medium-scale target representation; and R3 focuses on deep semantics and large-scale contextual representation. Through this structure, CTCR enables shallow localization information and deep semantic information to be reconstructed at the same target scale, reducing the inconsistency in representation caused by directly inputting features of different scales into the detection head.

[0069] As one or more implementations, the detection head or detection decoder receives R1, R2, and R3, and outputs the detection result Y. The detection head performs target category prediction, bounding box position prediction, and confidence prediction based on the multi-scale reconstructed detection features to obtain the final detection result. In one embodiment, R1, R2, and R3 can all be set to 256 channels, where R1 retains higher resolution detection features, R2 retains medium-scale detection features, and R3 retains low-resolution deep detection features.

[0070] In one embodiment, the detection decoder can receive reconstructed detection features with the same number of channels across three layers and output a fixed number of candidate detection results. Each candidate detection result includes the target category, bounding box coordinates, and confidence score. The detection head can be an end-to-end detection decoder or other detection head structures suitable for multi-scale target detection. The detection decoder is used to complete the final prediction output and is not considered a core innovation of this embodiment that distinguishes it from existing technologies.

[0071] To address the issue of small target pixel ratio and easy loss of details during downsampling in complex low-altitude scenes, this embodiment combines shallow high-resolution detail responses with mid-to-low-level semantic constraint information through LSDI, enabling target-related details to be enhanced before entering subsequent detection feature construction. B1 provides high-resolution information such as small target edges, contours, and textures, while B2 provides mid-to-low-level semantic constraints. The two are gated and injected to obtain D1, thereby avoiding reliance solely on deep low-resolution semantic features.

[0072] To address non-target responses generated by complex backgrounds such as water surfaces, roads, buildings, shorelines, and shadows, this embodiment uses LSIG to simultaneously estimate target responses and background interference responses, and employs gating to modulate multi-layer detection features. Target response gating enhances regions that may contain targets, while background interference gating reduces the impact of non-target textures, reflections, edges, and shadows. Residual output is used to avoid over-suppression that could lead to target information loss.

[0073] To address the issue of inconsistent semantic and spatial representations among features at different scales, this embodiment uses CTCR to perform scale alignment, cross-scale response aggregation, and consistency reconstruction on the gated multi-layer features, resulting in a more consistent target response representation for the multi-scale features output to the detection head. Small targets rely more on high-resolution features, while larger targets or contexts rely more on deep semantics. CTCR reconstructs features at different scales around the target response through unified channels, scale alignment, and cross-scale weighted aggregation.

[0074] This embodiment does not enhance only a single level of features, but processes the detection features step by step in the order of "shallow detail injection, low-altitude background interference gating, and cross-scale consistency reconstruction", so that small target details, background suppression information and multi-scale semantic information form a unified target detection feature expression before the detection head.

[0075] This embodiment combines shallow high-resolution detail responses with mid-to-low-level semantic constraint information through a Low-altitude Small-object Detail Injection Unit (LSDI). It uses gating to filter target-related details, enhancing the edge, contour, and local texture representation of distant small targets, thus improving the low-altitude small target detail response and reducing the risk of small targets being weakened during downsampling. A Low-altitude Scene Interference Gating Unit (LSIG) simultaneously estimates target and background interference responses, using residual gating to modulate multi-layer detection features. This helps to reduce interference from non-target areas such as water reflections, road textures, building edges, shoreline structures, and shadows, effectively suppressing non-target responses in complex backgrounds. A Cross-scale Target Consistency Reconstruction Module (CTCR) performs scale alignment, cross-scale response aggregation, and consistency reconstruction on the gated multi-layer features, improving the consistency of target representation across different scales in the detection head input features, making multi-scale detection features more suitable for category prediction and bounding box localization.

[0076] Example 2 Embodiment 2 of the present invention introduces a UAV target detection system for complex low-altitude scenarios.

[0077] like Figure 6 The UAV target detection system shown includes: The acquisition module is configured to acquire drone aerial images of complex low-altitude scenes; The detection module is configured to perform UAV target detection in complex low-altitude scenarios based on the acquired images and the UAV target detection network. The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network, and the detection module includes: The feature extraction submodule is configured to extract multi-layer basic features of the acquired image based on the basic feature extraction network; wherein the extracted multi-layer basic features include shallow high-resolution basic features, low-to-medium layer basic features, high-to-medium layer basic features, and deep basic features. The low-altitude small target detail injection submodule is configured to input shallow high-resolution basic features and mid-to-low-level basic features into the low-altitude small target detail injection unit to obtain small target detail enhancement features. The low-altitude scene interference gating submodule is configured to input small target detail enhancement features, low-to-mid-level basic features, mid-to-high-level basic features and deep basic features into the low-altitude scene interference gating unit to obtain multi-layer detection features after interference gating. The cross-scale target consistency reconstruction submodule is configured to input multi-layer detection features into the cross-scale target consistency reconstruction module to obtain multi-scale reconstructed detection features. The detection output submodule is configured to input multi-scale reconstructed detection features into the detection head or detection decoder to obtain detection results including target category, bounding box position and confidence level, thus completing UAV target detection for complex low-altitude scenarios.

[0078] The detailed steps are the same as those provided in Example 1 for UAV target detection in complex low-altitude scenarios, and will not be repeated here.

[0079] Example 3 Based on the UAV target detection method for complex low-altitude scenarios introduced in Embodiment 1, Embodiment 3 of this invention introduces a UAV target detection method using an end-to-end deep learning monitoring framework for the UAV target detection network.

[0080] like Figure 7 The method for UAV target detection in complex low-altitude scenarios, as shown, includes: Acquire drone aerial images of complex low-altitude scenes; A convolutional neural network based on residual networks is used to extract the original feature map of the acquired image. The extracted original feature maps are fused and enhanced to obtain the enhanced feature map; The enhanced feature map is analyzed to obtain the target category and bounding box location of the image, thus completing the drone target detection for complex low-altitude scenes.

[0081] To address the problem that existing target detection models struggle to identify small targets due to the loss of spatial details in high-level feature maps, this embodiment effectively integrates the rich spatial details of low-level feature maps with the strong semantic information of high-level features, thereby enhancing the feature representation of small, dense targets from a UAV aerial perspective. Regarding the issue of existing feature pyramids easily generating conflicts and inconsistent information representations when fusing multi-scale features, this embodiment addresses how to adaptively integrate features from all scales to resolve spatial and scale conflicts and generate an information-rich and highly unified feature representation for subsequent detection heads. Finally, to address the difficulty of existing target detection models in efficiently establishing long-range dependencies in complex scenes, this embodiment addresses how to efficiently model global context and long-range dependencies, thereby enhancing the model's ability to distinguish heterogeneous targets in the scene (such as ships and vehicles in a port).

[0082] The architecture diagram of the UAV target detection method for complex low-altitude scenarios in this embodiment is as follows: Figure 8 As shown, an end-to-end deep learning monitoring framework is adopted, which includes the following three core functional modules connected in sequence: The feature extraction module (Backbone) is used to extract multi-level raw feature maps from the input drone aerial images; The feature encoder module (AUP-Encoder) is used to perform deep and targeted fusion and enhancement on the original feature map; The target decoder module (Decoder) is used to parse and generate the final target category and bounding box location information based on the enhanced features.

[0083] This embodiment receives a frame of drone aerial image, and the feature extraction module extracts multi-level original feature maps. The feature encoder module performs complex fusion and enhancement on these original feature maps to solve problems such as small target representation, multi-scale conflict and global context modeling. The target decoder module generates the category and location information of the target in the image based on the enhanced features output by the feature encoder module.

[0084] (1) Feature extraction module The feature extraction module is a convolutional neural network (CNN) based on a residual network (Res Net). Its specific structure and data flow strictly follow the definition of the backbone section in the AUP-DETR.yaml configuration file. This module receives image input and passes it sequentially through a Conv Norm Layer with a stride of 2, two Conv Norm Layers with a stride of 1, and an nn.MaxPool2d layer. The feature map goes through four Blocks stages, each containing multiple basic residual blocks. Finally, it outputs four feature maps of different scales, denoted as S2, S3, S4, and S5, which are then transmitted to the subsequent feature encoder module.

[0085] (2) Feature encoder module The feature encoder module is the core of this embodiment. Internally, it employs a combination of top-down and bottom-up feature fusion paths and contains the following three key sub-modules. Their detailed structure and connections strictly adhere to the definition in the head section of the AUP-DETR.yaml configuration file: Submodule 1: Spatial Intelligent Transformer (SAT) To address the challenge of associating drones with heterogeneous targets such as distant ships and nearby vehicles in large-scale scenarios like ports, this embodiment employs the following approach: Figure 11 The SAT module shown here. Compared to the traditional Transformer, its Agent Attention mechanism uses a small set of agent tokens to "aggregate" and "distribute" global information, significantly reducing the computational complexity of attention while maintaining the global receptive field. Its SA-FFN ensures that spatial information crucial for target localization is not destroyed during feature transformation.

[0086] The detailed structure of the Spatial Intelligent Transformer (SAT) is as follows: Figure 11 As shown, the input of this module is connected to the highest-level feature map S5 output by the feature extraction module for efficient modeling of the global context; internally, the SAT module consists of two sequentially stacked "adaptive residual blocks" (num_layers=2) and uses eight attention heads (num_heads=8). Each "adaptive residual block" (such as...) Figure 11 (A) in the image contains an "AgentAttention component" (as shown in the image). Figure 11 (as shown in (B)) and a "Spatial Aware Feedforward Network (SA-FFN) component" (as shown in (B)). Figure 11 (as shown in (C)).

[0087] In this embodiment, the Agent Attention component employs a two-stage "collect-broadcast" mechanism: In the first stage, a small set of Agent Tokens is generated by adaptive average pooling of the query (Q) feature map. These Agent Tokens "collect" global contextual information from the complete key (K) and value (V) feature maps. In the second stage, the original query (Q) feature map "broadcasts" information from the Agent Tokens that already contain global information to update itself, thereby efficiently establishing long-distance dependencies.

[0088] In this embodiment, the Spatial Aware Feedforward Network (SA-FFN) component uses a fully convolutional structure instead of the traditional fully connected layer. Specifically, it consists of a 1x1 convolution (for channel expansion), a 3x3 depthwise separable convolution (for feature transformation), and a 1x1 convolution (for channel compression) connected sequentially to completely preserve the two-dimensional spatial topological information during feature transformation.

[0089] Submodule 2: Fusion-SHC (Fusion Streamlined Hybrid Core) Conventional feature fusion can lead to the fine spatial features of small targets (such as pedestrians and small boats) from the drone's perspective being overwhelmed by high-level semantic features. In this embodiment, for example... Figure 9 The Fusion-SHC module shown here, by specifically fusing low-level high-resolution features from S2 and S3 and employing designs such as Spatial Multi-Scale Aggregation (SCA) and Large Kernel Convolution, can maximize the preservation and enhancement of the detailed information of these tiny targets, thereby improving the detection recall rate.

[0090] The detailed structure of the integrated streamlined hybrid core is as follows: Figure 9 As shown, the Fusion-SHC module is used to fuse low-level features to enhance the feature representation of small targets. Its core is an "SHC-Kernel." It employs a channel segmentation design, processing only a subset of channels (e.g., 25%) to improve efficiency. The processing part includes multiple parallel branches: a "Spatial Multi-Scale Aggregation (SCA) branch" composed of parallel 1x1, 3x3, and 5x5 (or 7x7) convolutions to capture diverse local features; a separate 7x7 (or other large-size) "Large Kernel Convolution Enhancement Branch" to expand the effective receptive field; and a "Channel Attention Branch" that performs global average pooling on the input features, generates channel weights through two convolutional layers, and applies them to the output of the SCA branch. The outputs of these branches are concatenated with the residuals of the input to form the output of the processing part. This output is then concatenated with another segmented through-channel and fused using a 1x1 convolution.

[0091] Submodule 3: Collaborative Spatial Context Fusion (SSCF) To address the semantic conflicts and inconsistencies arising from the direct fusion of features at different scales, this embodiment employs the following approach: Figure 10 The SSCF module shown intelligently solves the problem of which scale feature to "trust more" at each spatial location by adaptively learning weights; the subsequent MSFF stage further enriches the contextual information of the fused features through multi-kernel convolution, ensuring that the decoder obtains a fully optimized and information-unified optimal feature representation.

[0092] Collaborative Spatial Context Fusion (SSCF), its detailed structure is attached. Figure 10 As shown, located at the end of the encoder, it is used to receive and unify the enhancement features from all levels, generating a single high-quality input for the decoder.

[0093] The SSCF module in this embodiment includes two stages (such as...). Figure 10 (A) in the middle): The first stage is "Adaptive Spatial Feature Fusion (ASFF)" (e.g., ... Figure 10 As shown in (B) in the diagram), this stage first passes through the "channel alignment block" (as shown in the diagram). Figure 10 (C) shows the channel dimensions of the unified input multi-scale feature maps (P3, P4, P5); P4 and P5 are upsampled to the same resolution as P3; then, pixel-wise spatial weights are learned for the three aligned feature maps and weighted sums are performed accordingly to adaptively eliminate scale conflicts.

[0094] The second stage is "multi-scale context enhancement". This stage feeds the output of ASFF into a parallel lightweight MSFF structure containing 1x1, 3x3 and 5x5 convolutions, and concatenates the outputs of the three branches and then performs a final fusion through a 1x1 hybrid convolution to capture rich contextual information.

[0095] In this embodiment, the internal data flow of the feature encoder module strictly follows the definition in the head section of the AUP-DETR.yaml file, specifically as follows: Figure 8 As shown: The S5 from the feature extraction module first performs channel mapping through a 1x1 convolutional layer, and then inputs it into the SAT module; The output of the SAT module is then passed through a 1x1 convolutional layer to generate a high-level enhanced feature map P5. P5 increases the resolution through a DySample upsampling module. Meanwhile, S4 from the feature extraction module performs channel mapping through a 1x1 convolutional layer. The upsampled P5 and the channel-mapped S4 are concatenated along the channel dimension (Concat), and then deep fused through a RepC3 module to generate the mid-layer enhanced feature map P4. P4 increases the resolution through a DySample upsampling module to obtain intermediate feature Y4. Meanwhile, S2 from the feature extraction module is processed through a Focus module. The processed S2, S3 from the feature extraction module, and intermediate feature Y4 are concatenated (Concat). The concatenated features are input into the Fusion-SHC module, and then deep fusion is performed through a RepC3 module to generate a low-level enhanced feature map P3. Starting from the bottom-up path, P3 is downsampled through a Frequency Focused Down Sampling module and concatenated with P4. Then, an updated P4 is obtained through a RepC3 module. The updated P4 is downsampled through a Frequency Focused Down Sampling module and concatenated with P5. Then, the updated P5 is obtained through a RepC3 module. The final P3, the updated P4, and the updated P5 are then input into the SSCF module for final collaborative fusion.

[0096] (3) Target decoder module The object decoder module is an RTDETR decoder, whose input is connected to the output of the SSCF module. Internally, it uses a multi-level attention mechanism to enable the interaction between a set of learnable object queries and the enhanced feature maps output by the SSCF module. Finally, it outputs the category and bounding box location of all objects in the image in parallel through a classification head and a regression head.

[0097] To address the issue of losing fine spatial features of small targets due to downsampling during feature extraction, this embodiment utilizes a streamlined hybrid core (Fusion-SHC) module to specifically process the low-level high-resolution feature maps (S2, S3) from the backbone network and combine them with semantic information from higher levels, thereby maximizing the preservation and enhancement of detailed representations of small targets.

[0098] Example 4 Embodiment 4 of the present invention provides a computer-readable storage medium.

[0099] A computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of a UAV target detection method for complex low-altitude scenarios as described in Embodiment 1 of the present invention.

[0100] The detailed steps are the same as those provided in Example 1 for UAV target detection in complex low-altitude scenarios, and will not be repeated here.

[0101] Example 5 Embodiment 5 of the present invention provides an electronic device.

[0102] An electronic device includes a memory, a processor, and a program stored in the memory and running on the processor. When the processor executes the program, it implements the steps in the UAV target detection method for complex low-altitude scenarios as described in Embodiment 1 of the present invention.

[0103] The detailed steps are the same as those provided in Example 1 for UAV target detection in complex low-altitude scenarios, and will not be repeated here.

[0104] Example 6 Embodiment 6 of the present invention provides a computer program product.

[0105] A computer program product includes software code, wherein the program in the software code performs the steps of a UAV target detection method for complex low-altitude scenarios as described in Embodiment 1 of the present invention.

[0106] The detailed steps are the same as those provided in Example 1 for UAV target detection in complex low-altitude scenarios, and will not be repeated here.

[0107] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of the present invention can be implemented using various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0108] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions 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 of the function specified in one or more boxes.

[0111] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention.

[0112] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

[0113] The above description is merely a preferred embodiment of this practice and is not intended to limit the scope of this practice. Various modifications and variations can be made to this practice by those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this practice should be included within the protection scope of this practice.

Claims

1. A method for UAV target detection in complex low-altitude scenarios, characterized in that, include: Acquire drone aerial images of complex low-altitude scenes; Based on the acquired images and the UAV target detection network, complete the UAV target detection for complex low-altitude scenarios; The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network. During UAV target detection, a basic feature extraction network extracts multi-layer basic features from the acquired image. These extracted multi-layer basic features include shallow high-resolution basic features, mid-to-low-level basic features, mid-to-high-level basic features, and deep basic features. The shallow high-resolution and mid-to-low-level basic features are input into a low-altitude small target detail injection unit to obtain small target detail enhancement features. The small target detail enhancement features, mid-to-low-level basic features, mid-to-high-level basic features, and deep basic features are input into a low-altitude scene interference gating unit to obtain interference-gated multi-layer detection features. The multi-layer detection features are input into a cross-scale target consistency reconstruction module to obtain multi-scale reconstructed detection features. The multi-scale reconstructed detection features are input into a detection head or detection decoder to obtain detection results including target category, bounding box position, and confidence level, thus completing UAV target detection for complex low-altitude scenes.

2. The UAV target detection method for complex low-altitude scenarios as described in claim 1, characterized in that, The low-altitude small target detail injection unit includes a shallow detail adaptation branch, a semantic adaptation branch, a detail injection gating branch, and an output integration branch. The shallow detail adaptation branch extracts local details and adapts scale to shallow high-resolution basic features to obtain a shallow detail response. The semantic adaptation branch performs channel adaptation and semantic constraint expression on mid-to-low-level basic features to obtain semantic constraint features. The detail injection gating branch generates detail injection gating coefficients based on the semantic constraint features to obtain a global semantic response. The output integration branch injects the gated shallow detail response into the semantic constraint features to obtain small target detail enhancement features.

3. The UAV target detection method for complex low-altitude scenarios as described in claim 1, characterized in that, The low-altitude scene interference gating unit includes a reference feature adaptation branch, a context projection branch, a context fusion branch, a target response gating branch, a background interference gating branch, and a residual modulation output branch. Specifically, the reference feature adaptation branch selects a corresponding reference feature based on the current output level and adapts it to the target output channel to obtain a reference response. The context projection branch projects small target detail enhancement features, low-to-mid-level basic features, mid-to-high-level basic features, and deep basic features into channels to obtain multi-layer context features with a unified context channel. The context fusion branch integrates the multi-layer context features after aligning them to the current output scale to obtain context fusion features. The target response gating branch estimates the target region response based on the reference response and context fusion features, generating target response gating coefficients. The background interference gating branch estimates the background interference response based on the reference response, generating background interference gating coefficients. The residual modulation output branch modulates the reference response based on the target response gating coefficients and background interference gating coefficients, and outputs the interference gating features of the current level.

4. The UAV target detection method for complex low-altitude scenarios as described in claim 1, characterized in that, The cross-scale target consistency reconstruction module includes a multi-layer feature projection branch, a scale alignment branch, a cross-scale response description branch, a scale weight generation branch, a weighted reconstruction branch, and an output integration branch; wherein... The multi-layer feature projection branch is used to project multi-layer detection features onto a unified channel dimension; the scale alignment branch is used to upsample or downsample multi-layer projection features to the same spatial scale according to the target output scale; the cross-scale response description branch is used to form a cross-scale target response description based on the aligned multi-layer features; the scale weight generation branch is used to generate scale weights for features of different levels based on the cross-scale target response description; the weighted reconstruction branch is used to perform weighted aggregation of multi-layer aligned features using scale weights to obtain reconstructed features at the current scale; and the output integration branch is used to perform two-stage 3x3 convolution-normalization-activation processing on the reconstructed features to output multi-scale reconstructed detection features.

5. The UAV target detection method for complex low-altitude scenarios as described in claim 1, characterized in that, The UAV target detection network adopts an end-to-end deep learning monitoring framework. During the UAV target detection process, the original feature map of the acquired image is extracted by a convolutional neural network based on the residual network. The extracted original feature maps are fused and enhanced to obtain enhanced feature maps; the enhanced feature maps are analyzed to obtain the target category and bounding box location of the image, thus completing the UAV target detection for complex low-altitude scenes.

6. The UAV target detection method for complex low-altitude scenarios as described in claim 5, characterized in that, In the process of fusing and enhancing the extracted original feature maps, a feature fusion path combining top-down and bottom-up approaches is adopted in the feature encoder. The feature encoder includes a spatial intelligent transformer, a streamlined hybrid fusion, and a collaborative spatial context fusion. The enhanced feature map obtained by the target decoder is used to parse it. The interaction between the learnable object query and the enhanced feature map is carried out through a multi-level attention mechanism. The category and bounding box position of all targets in the image are output in parallel through the classification head and the regression head.

7. A UAV target detection system for complex low-altitude scenarios, characterized in that, include: The acquisition module is configured to acquire drone aerial images of complex low-altitude scenes; The detection module is configured to perform UAV target detection in complex low-altitude scenarios based on the acquired images and the UAV target detection network. The UAV target detection network employs a low-altitude target enhancement and reconstruction detection network, and the detection module includes: The feature extraction submodule is configured to extract multi-layer basic features of the acquired image based on the basic feature extraction network; wherein the extracted multi-layer basic features include shallow high-resolution basic features, low-to-medium layer basic features, high-to-medium layer basic features, and deep basic features. The low-altitude small target detail injection submodule is configured to input shallow high-resolution basic features and mid-to-low-level basic features into the low-altitude small target detail injection unit to obtain small target detail enhancement features. The low-altitude scene interference gating submodule is configured to input small target detail enhancement features, low-to-mid-level basic features, mid-to-high-level basic features and deep basic features into the low-altitude scene interference gating unit to obtain multi-layer detection features after interference gating. The cross-scale target consistency reconstruction submodule is configured to input multi-layer detection features into the cross-scale target consistency reconstruction module to obtain multi-scale reconstructed detection features. The detection output submodule is configured to input multi-scale reconstructed detection features into the detection head or detection decoder to obtain detection results including target category, bounding box position and confidence level, thus completing UAV target detection for complex low-altitude scenarios.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of a UAV target detection method for complex low-altitude scenarios as described in any one of claims 1-6.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the program, it implements the steps of a UAV target detection method for complex low-altitude scenarios as described in any one of claims 1-6.

10. A computer program product, comprising software code, characterized in that, The program in the software code executes the steps of a UAV target detection method for complex low-altitude scenarios as described in any one of claims 1-6.

Images