Delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention

Abstract

The application discloses a delivery platform detection system and method based on sub-band space cooperative frequency domain attention, relates to the technical field of computer vision and unmanned aerial vehicle intelligent sensing, and comprises the following steps: S1, preprocessing a delivery scene image collected by an unmanned aerial vehicle to obtain a standard delivery scene image; S2, inputting the standard delivery scene image into a pre-trained delivery platform detection network to obtain a preliminary delivery platform detection result; and S3, screening the preliminary delivery platform detection result according to a confidence threshold and an intersection over union threshold to obtain a final delivery platform detection result. The application jointly utilizes sub-band information, space structure information and frequency domain information, so that the network can clearly see the details of the delivery platform and accurately see the overall saliency of the delivery platform in a complex scene, thereby improving the detection precision and positioning reliability of the platform target in the unmanned aerial vehicle delivery process, and solving the problems of small delivery platform target, complex background, easy-to-submerge edge and unstable positioning in the unmanned aerial vehicle delivery scene.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and UAV intelligent perception technology, and in particular to a delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention. Background Technology

[0002] With the rapid development of e-commerce and the increasing demands of consumers for timely delivery, traditional logistics methods are gradually revealing problems such as low efficiency and high costs when dealing with the ever-increasing order volume and complex and ever-changing delivery scenarios. Against this backdrop, drone logistics delivery has emerged as a new logistics model and is demonstrating enormous development potential due to its unique advantages.

[0003] Drone logistics delivery boasts significant advantages such as speed, flexibility, and the ability to overcome geographical limitations. It can quickly traverse congested urban areas, delivering goods directly to designated locations, significantly reducing delivery time. This makes it particularly suitable for remote areas, emergency supplies delivery, and last-mile delivery scenarios. For example, in mountainous regions and islands where transportation is difficult, drones can easily overcome terrain obstacles to achieve efficient delivery. During natural disasters, drones can rapidly deliver relief supplies to affected areas, buying valuable time for disaster relief efforts. Furthermore, drone logistics delivery can reduce labor costs, minimize delivery delays caused by traffic congestion, and improve the operational efficiency and competitiveness of logistics companies.

[0004] However, to achieve large-scale and routine application of drone logistics delivery, a series of key technical issues need to be addressed, among which accurate detection and stable positioning of the delivery platform are crucial. As the site for drone cargo loading, unloading, takeoff, and landing, the accurate detection and positioning of the delivery platform directly affects whether the drone can safely and accurately complete its delivery mission. If the delivery platform's detection is inaccurate or its positioning is unstable, the drone may be unable to correctly identify the platform's location, leading to landing failures or even safety accidents, severely impacting the efficiency and reliability of logistics delivery.

[0005] In real-world drone logistics delivery scenarios, the detection and positioning of delivery platforms face interference from numerous complex factors. These factors make it difficult for traditional detection methods to meet practical application requirements, specifically in the following aspects:

[0006] 1) The issue of viewing angle and target size

[0007] Drones typically observe target areas from an aerial perspective, either looking down or at an angle. This perspective differs significantly from the traditional ground-based view. From an aerial viewpoint, the delivery platform appears differently in the image; its shape, proportions, and texture differ from those seen from a ground-based perspective, increasing the difficulty of detection. Furthermore, delivery platforms are often small in size and occupy fewer pixels in images, making it difficult to clearly present their details. Traditional detection methods are prone to missed or false detections when dealing with small targets, making it difficult to accurately identify the boundaries and location of delivery platforms.

[0008] 2) Complex background interference

[0009] Delivery platforms are typically located in complex environments with numerous interfering factors. Building structures, balcony facilities, air conditioner units, billboards, and other building attachments can overlap or obscure the delivery platform in the image, making its features indistinct and difficult to distinguish from the background. Furthermore, road markings, shadows, and reflections can also affect image quality, further interfering with delivery platform detection. For example, shadows and reflections may cause false edges or bright spots in the image, leading traditional detection methods to mistakenly identify these interfering elements as features of the delivery platform, resulting in false detections.

[0010] 3) Flight status impact

[0011] During flight, drones are affected by factors such as airflow and mechanical vibration, resulting in flight jitter and slight blurring. Flight jitter causes blurring and distortion in images captured by the drone, making the edges of the delivery platform unclear and losing details. Slight blurring reduces image resolution and contrast, making it more difficult to extract and identify the features of the delivery platform. Traditional detection methods typically rely on clear images for feature extraction and matching; however, their detection performance drops significantly for blurry and jittery images, making it impossible to accurately detect and locate the delivery platform.

[0012] 4) Limitations of traditional detection methods in delivery platform inspection:

[0013] Faced with various interference factors in the above complex scenarios, traditional detection methods have many limitations in the detection and positioning of delivery platforms in drone logistics delivery, making it difficult to meet the needs of practical applications.

[0014] 5) Unclear edge detection

[0015] Traditional edge detection algorithms, such as the Sobel and Canny operators, primarily rely on the grayscale gradient information of an image to detect object edges. However, in images captured by drones, the edges of delivery platforms are often blurred due to complex background interference and flight vibrations, resulting in indistinct grayscale gradient changes. Traditional edge detection algorithms struggle to accurately capture these blurred edges, leading to incomplete or inaccurate detection of platform edges and an inability to accurately define the delivery platform's boundaries.

[0016] 6) Significant loss of detail

[0017] Traditional detection methods typically downsample or filter images to reduce computation and noise. However, this process leads to the loss of detail, especially for smaller delivery platforms where subtle features are more easily overlooked. This loss of detail prevents the full extraction and utilization of the delivery platform's unique characteristics, thus impacting the accuracy and reliability of the detection.

[0018] 7) Background confusion is a prominent issue.

[0019] Traditional detection methods often lack the ability to effectively distinguish complex backgrounds. In drone images, delivery platforms are highly similar to their surroundings, making it difficult for traditional methods to accurately identify their features. They may misidentify similar objects in the background as delivery platforms, or confuse the delivery platform with the background, failing to separate them accurately. For example, a billboard's shape and color may be similar to a delivery platform, leading to incorrect detection results due to traditional methods.

[0020] 8) Inaccurate center positioning

[0021] Accurate center positioning is crucial for the safe landing of drones. Traditional detection methods typically calculate the center position of a delivery platform based on detected edges or feature points. However, due to issues such as unclear edges, loss of detail, and background obfuscation mentioned earlier, the calculated center position deviates significantly from the actual center position. This inaccurate positioning prevents the drone from accurately aligning with the delivery platform during landing, increasing the risk of landing failure and potentially damaging both the drone and the cargo.

[0022] In summary, due to the complexity and unique characteristics of drone logistics delivery scenarios, traditional detection methods suffer from numerous problems in delivery platform detection and localization, making it difficult to meet the needs of practical applications. Therefore, there is an urgent need to develop a delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention to address the practical problems of "difficult detection, prone to false detection, and unstable localization" of delivery platforms in drone logistics delivery, and to promote the further development and application of drone logistics delivery technology. Summary of the Invention

[0023] To address the aforementioned technical problems in related technologies, this invention proposes a delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention.

[0024] In a first aspect, the present invention provides a delivery platform detection system based on sub-band spatial cooperative frequency domain attention, comprising:

[0025] The image acquisition and preprocessing module is used to preprocess the delivery scene image I0 acquired by the drone's onboard camera to obtain a standard delivery scene image;

[0026] The target detection module is used to input the standard delivery scene image into a pre-trained delivery platform detection network to obtain preliminary delivery platform detection results;

[0027] The delivery platform detection network includes a backbone network, an attention enhancement module, a feature fusion network, and a detection head. The backbone network includes a STEM convolutional layer, a first residual stage, a second residual stage, a third residual stage, and a fourth residual stage. The standard delivery scene image is input into the backbone network, passing through the STEM convolutional layer, the first residual stage, and the second residual stage sequentially to obtain a second residual feature. This second residual feature is then input into the attention enhancement module to obtain a second enhanced feature. The second enhanced feature is then input into the third residual stage to obtain a third residual feature, which is then input into the attention enhancement module again to obtain a third enhanced feature. The third enhanced feature is then input into the fourth residual stage to obtain high-level features. The second enhanced feature, the third enhanced feature, and the high-level features are then fused at multiple scales through the feature fusion network to obtain a second fused feature, a third fused feature, and a fourth fused feature. These fused features are then input into the detection head to perform confidence classification, bounding box regression, and center point regression for the drone delivery platform, resulting in preliminary delivery platform detection results.

[0028] The post-processing module is used to filter the preliminary delivery platform detection results based on the confidence threshold and the crossover ratio threshold to obtain the final delivery platform detection results.

[0029] Specifically, the attention enhancement module includes a Haar wavelet subband decomposition unit for processing the l-th residual feature. Haar discrete wavelet transform decomposition yields four wavelet subbands, including:

[0030] ,

[0031] in, It is a low-frequency sub-band. For low- and high-frequency detail subbands, For high and low frequency detail subbands, For high-frequency detail subbands; the dimensions of the four wavelet subbands are all... where l is the stage index; The number of channels for the l-th residual feature; The height of the l-th residual feature; Indicates the width of the second residual feature; () indicates the operation of extracting LL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting LH subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting HL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting the HH subband after performing Haar discrete wavelet transform on the input features;

[0032] Specifically, the attention enhancement module also includes a subband space tensor construction unit, used to stack four wavelet subbands to construct a subband space tensor. ;in, The first dimension represents the channel dimension; the second dimension represents the sub-band dimension, where 4 is the length of the second dimension, corresponding to four fixed wavelet sub-bands. The third dimension represents the row dimension; The fourth dimension represents Levy.

[0033] Specifically, the attention enhancement module further includes a global frequency domain channel gating unit, used to adjust the subband space tensor along the subband dimension, row dimension, and column dimension. Perform a three-dimensional fast Fourier transform to obtain the complex spectrum. and for complex spectrum Modulus is taken to obtain amplitude spectrum Then analyze the amplitude spectrum separately. After performing global average pooling and global max pooling, the inputs are fed into the same multilayer perceptron to obtain the initial global frequency domain channel gating. Finally, the initial global frequency domain channel gating is performed. Extending along the sub-band dimension, row dimension, and column dimension yields the global frequency domain channel gating. .

[0034] Specifically, the attention enhancement module also includes a local sub-band direction gating unit for gating the sub-band spatial tensor. Execute one layer Convolution yields the first intermediate feature Then the first intermediate feature Input basic blocks Obtain the second intermediate feature Finally, regarding the second intermediate feature Execute one more layer Convolution followed by Sigmoid activation to obtain local subband orientation gating .

[0035] Specifically, the attention enhancement module further includes a collaborative gating enhancement unit for gating global frequency domain channels. Local sub-band direction gating To achieve synergistic integration, the sub-band space tensor Adaptive enhancement yields the co-enhanced subband space tensor. :

[0036] ,

[0037] in, This represents element-wise multiplication. Represents the sub-band space tensor A single sheet of uniform size; The first learnable weight coefficient; This is the second learnable weight coefficient; This is the third learnable weight coefficient.

[0038] Specifically, the attention enhancement module further includes an inverse wavelet reconstruction output unit, used to first convert the collaboratively enhanced subband spatial tensor into an inverse wavelet reconstruction output unit. It is split along the sub-band dimension into four enhancement sub-bands, including For low-frequency structure enhancement subband, To enhance low and high frequency details, sub-bands To enhance high and low frequency details, sub-bands A sub-band for enhancing high-frequency details was created; subsequently, inverse wavelet reconstruction was performed on the four enhanced sub-bands to obtain the reconstructed incremental features. Finally, the incremental features will be reconstructed. Compared with the original l-th residual feature By summing the residuals, we obtain the enhanced output features. .

[0039] Specifically, the feature fusion network adopts a pyramid-style multi-scale fusion method, first aligning the channels of features at different layers, then performing upsampling, downsampling, and element-wise addition or splicing fusion, thereby forming three-scale fusion features: the second fusion feature, the third fusion feature, and the fourth fusion feature.

[0040] Specifically, the post-processing module includes:

[0041] First, based on the confidence threshold For the candidate box set and its corresponding confidence set Set of center points The process involves filtering out low-confidence candidate targets to obtain the remaining candidate targets, and then determining the target based on the intersection-union threshold. Non-maximum suppression is applied to the remaining candidate targets to obtain the final detection result of the drone delivery platform; the candidate targets are the set of candidate boxes. The corresponding confidence set and the set of center points A set of results at the same index position; the final drone delivery platform detection result includes the final candidate box set, the final confidence set, and the final center point set.

[0042] Secondly, the present invention also provides a delivery platform detection method based on sub-band spatial cooperative frequency domain attention, which, based on the delivery platform detection system based on sub-band spatial cooperative frequency domain attention described in the first aspect above, specifically includes the following steps:

[0043] S1. Preprocess the delivery scene images captured by the UAV's onboard camera to obtain standard delivery scene images;

[0044] S2. Input the standard delivery scene image into a pre-trained delivery platform detection network to obtain preliminary delivery platform detection results;

[0045] S3. The preliminary delivery platform detection results are filtered according to the confidence threshold and the crossover ratio threshold to obtain the final delivery platform detection results.

[0046] The delivery platform detection system and method based on sub-band spatial collaborative frequency domain attention provided by this invention achieves more targeted feature enhancement by introducing a sub-band spatial collaborative frequency domain attention mechanism into a detection network with ResNet as the backbone. First, the backbone network is used to extract multi-layer visual features in the delivery scene. Then, wavelet decomposition is performed on the features in the intermediate key layer to separate the originally mixed main structure information and detailed information such as edges, textures, and corners into different sub-bands, thereby explicitly preserving the outline and local identification features of the delivery platform. On this basis, each sub-band is constructed into a sub-band-spatial joint representation, and further combined with global frequency domain modeling, the effective channel response related to the delivery platform is enhanced, while suppressing irrelevant background information such as roof debris, building structure, advertising facilities, and ground interference. Subsequently, through the synergistic effect of local sub-band enhancement and global frequency domain enhancement, the enhanced features are reconstructed back to the original spatial feature domain and fed into the subsequent feature fusion and detection output part to achieve stable prediction of the delivery platform's location, boundary, and center point. This invention does not rely solely on conventional spatial domain convolution for "image recognition." Instead, it utilizes subband information, spatial structure information, and frequency domain information to enable the network to not only see the details of the delivery platform but also accurately assess its overall saliency in complex scenarios. This improves the detection accuracy and positioning reliability of platform targets during drone delivery and solves the problems of small delivery platform targets, complex backgrounds, easily obscured edges, and unstable positioning in drone delivery scenarios.

[0047] Furthermore, the delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention provided by this invention has the following advantages:

[0048] 1) Introduce the "subband spatial collaborative frequency domain attention" mechanism in the ResNet backbone network, which combines wavelet subband decomposition with global frequency domain channel enhancement, and no longer relies solely on conventional spatial domain convolution for target representation;

[0049] 2) By separating and modeling low-frequency structural information with high-frequency edge, texture, and corner details, and combining the synergistic effect of local sub-band enhancement and global frequency domain enhancement, the discriminability of the delivery platform under complex background, occlusion, jitter, and slight blur conditions is improved.

[0050] 3) For drone logistics delivery scenarios, achieve more stable detection and center positioning of delivery platform targets, providing more reliable visual perception results for subsequent hovering alignment, path fine-tuning and accurate delivery. Attached Figure Description

[0051] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a schematic diagram of a delivery platform detection system based on sub-band spatial cooperative frequency domain attention provided in an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of the attention enhancement module structure provided in an embodiment of the present invention;

[0054] Figure 3 This is a schematic diagram of a delivery platform detection method based on sub-band spatial cooperative frequency domain attention provided in an embodiment of the present invention. Detailed Implementation

[0055] The invention will be explained in detail through the following embodiments. The purpose of this invention is to protect all technical improvements within its scope. In the description of this invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0056] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] refer to Figure 1 This embodiment discloses a delivery platform detection system based on sub-band spatial cooperative frequency domain attention, including the following modules:

[0059] The image acquisition and preprocessing module is used to preprocess the delivery scene image I0 acquired by the UAV's onboard camera to obtain the standard delivery scene image I;

[0060] Original delivery scene images were captured using drone-borne cameras. Then, the original delivery scene image I0 is preprocessed to ensure that the input image meets the fixed input requirements of the subsequent ResNet backbone network. After preprocessing, a standard delivery scene image is obtained. Where H0 and W0 represent the image height and width of the delivery scene image, respectively; H and W represent the normalized image height and width, respectively; and 3 represents the RGB three-dimensional color channel.

[0061] The preprocessing includes, but is not limited to, image resizing, normalization, and format standardization.

[0062] This step is only used to unify the scale and range of input data and does not change the target semantic information of the delivery drone platform.

[0063] The target detection module will analyze the standard delivery scene image. The data is input into a pre-trained delivery platform detection network to obtain preliminary detection results; the preliminary detection results include a set of candidate boxes. Confidence set and the set of center points ;

[0064] The delivery platform detection network includes a backbone network, an attention enhancement module, a feature fusion network, and a detection head;

[0065] The backbone network is a ResNet backbone network, which includes a stem convolutional layer, a first residual stage, a second residual stage, a third residual stage, and a fourth residual stage. The stem convolutional layer is the initial convolutional extraction part after input, used to extract initial features; the four residual stages are used to extract feature information at different levels step by step.

[0066] Attention enhancement modules are embedded at the end of the second and third residual stages of the backbone network to enhance the residual features output at the end of the stages.

[0067] The standard delivery scene image I is input into the backbone network, and sequentially passes through a STEM convolutional layer, a first residual stage, and a second residual stage to obtain the second residual feature. Among them, the second residual characteristic It's still essentially a diagram. This represents the second residual feature of the output of the second residual stage. Size, The number of channels for the second residual feature. The height of the second residual feature. The width of the second residual feature is given. It is worth noting that the superscript (2) in the formula indicates that the feature comes from the second residual stage, not the squared meaning.

[0068] Then, the second residual feature is input into the attention enhancement module to obtain the second enhancement feature; the second enhancement feature is input into the third residual stage to obtain the third residual feature, and then the third residual feature is input into the attention enhancement module to obtain the third enhancement feature;

[0069] refer to Figure 2 The attention enhancement module, officially named the subband space collaborative frequency domain attention module, is a core enhancement module embedded in the ResNet backbone network. It enhances the detailed representation and global discrimination capabilities of delivery drone platforms through wavelet subband decomposition, three-dimensional frequency domain channel gating, local subband / direction gating, and reconstruction enhancement. The attention enhancement module includes a Haar wavelet subband decomposition unit, a subband space tensor construction unit, a global frequency domain channel gating unit, a local subband direction gating unit, a collaborative gating enhancement unit, and an inverse wavelet reconstruction output unit, as detailed below:

[0070] Haar wavelet subband decomposition unit: for the l-th residual feature Applying the Haar discrete wavelet transform, it is decomposed into one low-frequency subband and three high-frequency subbands, for a total of four wavelet subbands, as shown below:

[0071] ,

[0072] in, This indicates a low-frequency sub-band. , and These represent high-frequency detail subbands in different directions. It is a low-frequency sub-band. For low- and high-frequency detail subbands, For high and low frequency detail subbands, For high-frequency detail subbands; the dimensions of the four wavelet subbands are all... where l is the stage index; The number of channels for the l-th residual feature; The height of the l-th residual feature; Indicates the width of the second residual feature; () indicates the operation of extracting LL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting LH subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting HL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting the HH subband after performing Haar discrete wavelet transform on the input features;

[0073] This step is used to explicitly separate the main structural information of the delivery drone platform from detailed information such as edges, corners, and marking textures.

[0074] Subband space tensor construction unit: according to fixed subband order Stack the four wavelet subbands to construct the subband space tensor. ;

[0075] in, The first dimension represents the channel dimension; the second dimension represents the sub-band dimension, where 4 is the length of the second dimension, corresponding to four fixed wavelet sub-bands. The third dimension represents the row dimension; The fourth dimension represents Levy.

[0076] This step reorganizes the original frequency band representation into a tensor with a "pseudo-depth" structure, enabling subsequent 3D frequency domain channel modeling to simultaneously characterize inter-subband correlations and spatial domain correlations.

[0077] Understandable, This is merely a preferred sub-band order provided in this embodiment and is not intended to limit the sub-band stacking order. In other possible implementations, the sub-band order can be adjusted accordingly. Arbitrary combinations yield a fixed sub-band order, which will not be elaborated here.

[0078] Global frequency domain channel gating unit: along the subband dimension, row dimension, and column dimension of the subband space tensor Perform a three-dimensional fast Fourier transform to obtain the complex spectrum. :

[0079] ,

[0080] Preserving subband space tensor First dimension (channel dimension) The first dimension remains unchanged, while the second dimension (sub-band dimension), the third dimension (row dimension), and the fourth dimension (column dimension) are subjected to a three-dimensional fast Fourier transform.

[0081] And for the complex spectrum Modulus is taken to obtain amplitude spectrum :

[0082] ,

[0083] Then analyze the amplitude spectrum separately. Perform global average pooling and global max pooling, and input the two statistical results into the same multilayer perceptron to obtain the initial global frequency domain channel gating. :

[0084] ,

[0085] Where AvgPool() represents global average pooling; MaxPool() represents global max pooling; MLP() represents multilayer perceptron; ( ) indicates Sigmoid activation; ;

[0086] Subsequently, the initial global frequency domain channel gating is performed. Extending along the sub-band dimension and spatial dimension (row and column dimensions) yields the final extended global frequency domain channel gating. ;

[0087] The phrase "extending along the sub-zone dimension and spatial dimension" in the text specifically refers to: extending along the sub-zone dimension and spatial dimension. Extend along the sub-band dimension to a length of 4, and extend along the two spatial dimensions respectively. and Thus obtain .

[0088] This step utilizes three-dimensional frequency domain statistics to establish a global frequency domain channel relationship between the delivery drone platform target and the surrounding background, thereby suppressing complex background interference and enhancing the platform's relevant channel response.

[0089] Local subband direction gating unit: for subband space tensor Execute one layer Convolution yields the first intermediate feature. :

[0090] ,

[0091] in, express convolution;

[0092] Subsequently, the first intermediate feature Input basic blocks The second intermediate feature is obtained. :

[0093] ,

[0094] in, This represents a basic block consisting of depthwise convolution, layer normalization, GELU activation, and pointwise convolution.

[0095] Finally, regarding the second intermediate feature Execute one more layer Convolution is performed and activated by a sigmoid function to obtain a local subband orientation gate. : ,in, .

[0096] This step performs local adaptive weighting in the wavelet domain for different sub-bands, directions, and spatial locations to highlight the edge contours, platform markings, and structural details of the delivery drone platform.

[0097] Collaborative Gating Enhancement Unit: Gating global frequency domain channels Local sub-band direction gating To achieve synergistic integration, the sub-band space tensor Adaptive enhancement yields the co-enhanced subband space tensor. :

[0098] ,

[0099] in, This represents element-wise multiplication. Represents the sub-band space tensor A single sheet of uniform size; The first learnable weight coefficient; This is the second learnable weight coefficient; This is the third learnable weight coefficient.

[0100] This step enables joint modeling of global frequency domain channel selection and local wavelet subband enhancement, thereby simultaneously improving the effective frequency components and structural details of the delivery drone platform.

[0101] Inverse wavelet reconstruction output unit: First, according to the fixed sub-band order Coordinate the enhancement of subband space tensor The subband is split along the subband dimension into four enhancement subbands, denoted as follows: , , and , For low-frequency structure enhancement subband, To enhance low and high frequency details, sub-bands To enhance high and low frequency details, sub-bands Subbands for enhancing high- and high-frequency details;

[0102] Subsequently, inverse wavelet reconstruction was performed on the four enhanced subbands to obtain the reconstruction increment features. :

[0103] ,

[0104] Finally, the incremental features will be reconstructed. Compared with the original l-th residual feature By summing the residuals, we obtain the enhanced output features. :

[0105] ,

[0106] This step restores the enhanced frequency domain information back to the spatial feature domain while maintaining the integrity of the main structural and detailed information of the delivery drone platform.

[0107] That is, the second residual feature The input is fed into the attention enhancement module to obtain the second enhanced feature. ; the second enhancement feature The third residual feature is obtained by inputting the data into the third residual stage. Then the third residual feature The input to the attention enhancement module yields the third enhanced feature. ;

[0108] Through the above two attention enhancement modules (sub-band spatial collaborative frequency domain attention module), the joint enhancement of local high-frequency details, directional textures, and global frequency domain correlations of the delivery drone platform is achieved.

[0109] The third enhancement feature is input into the fourth residual stage to obtain high-level features. Then, the second enhancement feature, the third enhancement feature, and the high-level features are fused at multiple scales through the feature fusion network to obtain the second fused feature, the third fused feature, and the fourth fused feature. The second fused feature, the third fused feature, and the fourth fused feature are input into the detection head to perform confidence classification, bounding box regression, and center point regression on the drone delivery platform to obtain preliminary delivery platform detection results. The preliminary delivery platform detection results include a candidate box set, a confidence set, and a center point set.

[0110] The third enhancement feature The input is fed into the fourth residual stage of ResNet to obtain high-level features. ;

[0111] The second enhanced feature is obtained through a feature fusion network. Third Enhancement Feature and high-level characteristics Multi-scale fusion is performed to obtain the second fusion feature. Third fusion feature and the fourth fusion feature Then, the second fusion feature is fused. Third fusion feature and the fourth fusion feature Input the detection head, perform confidence classification, bounding box regression, and center point regression for the drone delivery platform, and obtain preliminary detection results for the delivery platform, including a set of candidate boxes. Confidence set and the set of center points ;

[0112] in, The candidate bounding box for the nth candidate drone delivery platform; confidence level Confidence level of the nth candidate drone delivery platform; This represents the coordinates of the center point of the nth candidate drone delivery platform; Let x and y represent the center point coordinates of the nth candidate drone delivery platform, respectively; N is the number of candidate drone delivery platforms initially detected.

[0113] Candidate box set Confidence set and the set of center points It's a one-to-one correspondence. That is, the nth candidate box... The nth confidence level and the nth center point They correspond to the same candidate delivery drone platform.

[0114] This step only completes the routine multi-scale feature fusion and candidate target prediction.

[0115] The feature fusion network adopts a conventional pyramid-style multi-scale fusion approach. First, it performs channel alignment on features at different levels, then performs upsampling, downsampling, and element-wise addition (or concatenation and fusion) to form a second fusion feature with three scales. Third fusion feature and the fourth fusion feature Using feature fusion networks for multi-scale fusion is an existing technology; therefore, this embodiment only provides a general overview and will not elaborate further.

[0116] The detection head is a key component in the object detection model. Its main function is to process the features extracted by the backbone network and the neck network to complete the object classification and bounding box regression tasks.

[0117] The detection head sets up prediction branches at each fusion scale, automatically calculating the confidence score, bounding box parameters, and center point coordinates of candidate targets at that scale based on the fusion features of that scale; then, the prediction results at each scale are aggregated to form a candidate box set. Confidence set and the set of center points The role of the detection head is to complete these three types of predictions based on fused features, using a conventional detection output structure.

[0118] It is understood that the ResNet of the delivery platform detection network in this embodiment is used as the backbone feature extraction network; the feature fusion network and the detection head are subsequent modules built on the backbone features. The role of the detection head in this application is to perform confidence classification, bounding box regression, and center point regression on the fused features.

[0119] The post-processing module is used to filter the preliminary delivery platform detection results based on the confidence threshold and the intersection-union ratio threshold to obtain the final drone delivery platform detection results;

[0120] Preset reliability threshold and the crossover ratio threshold First, based on the confidence threshold For the candidate box set and its corresponding confidence set Set of center points The process involves filtering and eliminating low-confidence candidate targets to obtain the remaining candidate targets.

[0121] Understandably, the candidate box set Confidence set and the set of center points It is a one-to-one correspondence, with the confidence threshold... With confidence set Each confidence level in Compare them one by one; if the confidence level of a certain item is... Below the confidence threshold Then from the candidate box set Confidence set and the set of center points The corresponding nth item is deleted synchronously.

[0122] The candidate target is a set of candidate boxes. The corresponding confidence set and the set of center points A set of results at the same index position, i.e., the nth candidate target. .

[0123] Subsequently, based on the crossover-union threshold Non-maximum suppression is performed on the remaining candidate targets to eliminate overlapping and redundant candidate boxes, retaining the final drone delivery platform detection result. The final drone delivery platform detection result includes: the final candidate box set. Final confidence set and the final set of center points .

[0124] in, For the m-th final drone delivery platform, the candidate bounding box is defined as follows: Confidence level Confidence level of the m-th final drone delivery platform; This represents the coordinates of the center point of the m-th final drone delivery platform; Let x and y represent the center point coordinates of the m-th final drone delivery platform, respectively; M is the number of final candidate drone delivery platforms detected.

[0125] Based on the crossover ratio threshold The non-maximum suppression (NMS) applied to the remaining candidate targets specifically includes: first, sorting the candidate boxes from highest to lowest confidence level among the remaining candidate targets after confidence level filtering, and retaining the candidate box with the highest confidence level; then calculating its intersection-union ratio (IU / I) with other candidate boxes, and selecting boxes with an IU / I greater than the IU / I threshold. Overlapping and redundant candidate boxes are deleted. At the same time as deleting the candidate boxes, their corresponding confidence items and center point items are also deleted.

[0126] Therefore, the final , , The final result retained after confidence screening and non-maximum suppression is the final detection result of the drone delivery platform.

[0127] The results obtained from this step can be directly used for visual positioning, hovering alignment, and subsequent delivery control of the delivery drone platform.

[0128] The core theme of the delivery platform detection system based on sub-band spatial cooperative frequency domain attention provided in this embodiment is to enable drones to more accurately "see and recognize" the actual delivery platform during flight delivery. In simpler terms, it equips drones with a smarter visual recognition method, allowing them to quickly locate the target platform for delivery or docking from a large number of complex backgrounds such as rooftops, balconies, ground markings, railings, air conditioner units, and billboards in aerial footage, and further determine the approximate boundaries and center position of the platform. For example, in scenarios like drone food delivery in Shenzhen, when the drone flies near buildings, the camera's view often presents problems such as small targets, cluttered backgrounds, large changes in lighting, flight vibrations, and slight blurring. The function of this core theme is to reliably identify the delivery platform under such complex conditions, providing reliable visual basis for subsequent hovering alignment, path fine-tuning, accurate delivery, and safe landing.

[0129] To address the challenges of small target size, complex background, easily obscured edges, and unstable positioning in drone delivery scenarios, this embodiment provides a delivery platform detection system based on sub-band spatial collaborative frequency domain attention. This system introduces a sub-band spatial collaborative frequency domain attention mechanism into a detection network with a ResNet backbone to achieve more targeted feature enhancement. First, the backbone network extracts multi-layered visual features from the delivery scene. Then, wavelet decomposition is performed on the features in the intermediate key layer, separating the originally mixed main structural information from details such as edges, textures, and corners into different sub-bands, thus explicitly preserving the outline and local identifier features of the delivery platform. Based on this, each sub-band is constructed into a sub-band-spatial joint representation, and further combined with global frequency domain modeling, the effective channel responses related to the delivery platform are enhanced while suppressing irrelevant background information such as roof clutter, building structures, advertising facilities, and ground interference. Subsequently, through the synergistic effect of local sub-band enhancement and global frequency domain enhancement, the enhanced features are reconstructed back into the original spatial feature domain and fed into the subsequent feature fusion and detection output, achieving stable prediction of the delivery platform's location, boundaries, and center point. In other words, this embodiment does not rely solely on conventional spatial domain convolution for "image recognition," but rather utilizes subband information, spatial structure information, and frequency domain information to enable the network to not only see the details of the delivery platform but also accurately assess its overall saliency in complex scenarios, thereby improving the detection accuracy and positioning reliability of platform targets during drone delivery.

[0130] Example 2

[0131] refer to Figure 3 This embodiment provides a delivery platform detection method based on sub-band spatial cooperative frequency domain attention, which is based on the delivery platform detection system based on sub-band spatial cooperative frequency domain attention described in Embodiment 1, and includes the following steps:

[0132] S1. Preprocess the delivery scene images captured by the UAV's onboard camera to obtain standard delivery scene images;

[0133] S2. Input the standard delivery scene image into a pre-trained delivery platform detection network to obtain preliminary delivery platform detection results;

[0134] The delivery platform detection network includes a backbone network, an attention enhancement module, a feature fusion network, and a detection head. The backbone network includes a STEM convolutional layer, a first residual stage, a second residual stage, a third residual stage, and a fourth residual stage. The standard delivery scene image is input into the backbone network, passing through the STEM convolutional layer, the first residual stage, and the second residual stage sequentially to obtain a second residual feature. This second residual feature is then input into the attention enhancement module to obtain a second enhanced feature. The second enhanced feature is then input into the third residual stage to obtain a third residual feature, which is then input into the attention enhancement module again to obtain a third enhanced feature. The third enhanced feature is then input into the fourth residual stage to obtain high-level features. The second enhanced feature, the third enhanced feature, and the high-level features are then fused at multiple scales through the feature fusion network to obtain a second fused feature, a third fused feature, and a fourth fused feature. These fused features are then input into the detection head to perform confidence classification, bounding box regression, and center point regression for the drone delivery platform, resulting in preliminary delivery platform detection results.

[0135] The attention enhancement module includes a Haar wavelet subband decomposition unit, a subband spatial tensor construction unit, a global frequency domain channel gating unit, a local subband direction gating unit, a cooperative gating enhancement unit, and an inverse wavelet reconstruction output unit.

[0136] S3. The preliminary delivery platform detection results are filtered according to the confidence threshold and the crossover ratio threshold to obtain the final delivery platform detection results.

[0137] The delivery platform detection system and method based on sub-band spatial cooperative frequency domain attention provided in this embodiment have the following highlights:

[0138] 1) Introduce the "subband spatial collaborative frequency domain attention" mechanism in the ResNet backbone network, which combines wavelet subband decomposition with global frequency domain channel enhancement, and no longer relies solely on conventional spatial domain convolution for target representation;

[0139] 2) By separating and modeling low-frequency structural information with high-frequency edge, texture, and corner details, and combining the synergistic effect of local sub-band enhancement and global frequency domain enhancement, the discriminability of the delivery platform under complex background, occlusion, jitter, and slight blur conditions is improved.

[0140] 3) For drone logistics delivery scenarios, achieve more stable detection and center positioning of delivery platform targets, providing more reliable visual perception results for subsequent hovering alignment, path fine-tuning and accurate delivery.

[0141] Example 3

[0142] The purpose of this embodiment is to provide a computing device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method of Embodiment 2.

[0143] Example 4

[0144] The purpose of this embodiment is to provide a computer-readable storage medium, which is a computer-readable storage medium.

[0145] The device stores a computer program that, when executed by a processor, performs the steps of the method described in Embodiment 2.

[0146] It is worth noting that in the above embodiments of the delivery platform detection system based on sub-band spatial cooperative frequency domain attention, the various units and modules included are only divided according to functional logic, but are not limited to the above division, as long as the corresponding functions can be achieved; in addition, the specific names of each functional unit are only for easy differentiation and are not used to limit the scope of protection of the present invention.

[0147] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. 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, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 A process, multiple processes, and / or boxes Figure 1 Devices that specify the functions in one or more boxes.

[0148] 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 an instruction device, which is implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0149] These computer program instructions can 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 1The steps of the function specified in one or more boxes.

[0150] The parts of this invention not described in detail are prior art. It will be apparent to those skilled in the art that this invention is not limited to the details of the above exemplary embodiments, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and are intended to encompass all changes falling within the meaning and scope of equivalents within this invention.

Claims

1. A delivery platform detection system based on sub-band spatial cooperative frequency domain attention, characterized in that, Includes the following modules: The image acquisition and preprocessing module is used to preprocess the delivery scene images acquired by the UAV's onboard camera to obtain standard delivery scene images; The target detection module is used to input the standard delivery scene image into a pre-trained delivery platform detection network to obtain preliminary delivery platform detection results; The delivery platform detection network includes a backbone network, an attention enhancement module, a feature fusion network, and a detection head. The backbone network includes a STEM convolutional layer, a first residual stage, a second residual stage, a third residual stage, and a fourth residual stage. The standard delivery scene image is input into the backbone network, passes through the STEM convolutional layer, the first residual stage, and the second residual stage in sequence to obtain the second residual feature. Then, the second residual feature is input into the attention enhancement module to obtain the second enhanced feature. The second enhancement feature is input into the third residual stage to obtain the third residual feature, and then the third residual feature is input into the attention enhancement module to obtain the third enhancement feature; The third enhancement feature is input into the fourth residual stage to obtain the high-level feature; then, the second enhancement feature, the third enhancement feature, and the high-level feature are fused at multiple scales through the feature fusion network to obtain the second fused feature, the third fused feature, and the fourth fused feature; the second fused feature, the third fused feature, and the fourth fused feature are input into the detection head to perform confidence classification, bounding box regression, and center point regression of the drone delivery platform to obtain the preliminary detection results of the delivery platform; The attention enhancement module includes a Haar wavelet subband decomposition unit, a subband spatial tensor construction unit, a global frequency domain channel gating unit, a local subband direction gating unit, a cooperative gating enhancement unit, and an inverse wavelet reconstruction output unit. Haar wavelet subband decomposition unit, used for the l-th residual feature Haar discrete wavelet transform decomposition yields four wavelet subbands: a low-low frequency structure subband, a low-high frequency detail subband, a high-low frequency detail subband, and a high-high frequency detail subband; the dimensions of each of the four wavelet subbands are [missing information]. where l is the stage index; The number of channels for the l-th residual feature; The height of the l-th residual feature; Indicates the width of the second residual feature; Subband space tensor building unit, used to stack four wavelet subbands to construct the subband space tensor. ;in, The first dimension represents the channel dimension; the second dimension represents the sub-band dimension. The third dimension represents the row dimension; The fourth dimension represents Levy; Global frequency domain channel gating unit, used to manipulate the subband space tensor along the subband dimension, row dimension, and column dimension. Perform a three-dimensional fast Fourier transform to obtain the complex spectrum. and for complex spectrum Modulus is taken to obtain amplitude spectrum Then analyze the amplitude spectrum separately. After performing global average pooling and global max pooling, the inputs are fed into the same multilayer perceptron to obtain the initial global frequency domain channel gating. Finally, the initial global frequency domain channel gating is performed. Extending along the sub-band dimension, row dimension, and column dimension yields the global frequency domain channel gating. ; Local subband direction gating unit, used for subband spatial tensor Execute one layer Convolution yields the first intermediate feature Then the first intermediate feature Input basic blocks Obtain the second intermediate feature Finally, regarding the second intermediate feature Execute one more layer Convolution followed by Sigmoid activation to obtain local subband orientation gating ; Cooperative gating enhancement unit, used to gate global frequency domain channels Local sub-band direction gating To achieve synergistic integration, the sub-band space tensor Adaptive enhancement yields the co-enhanced subband space tensor. : ， in, This represents element-wise multiplication. Represents the sub-band space tensor A single sheet of uniform size; The first learnable weight coefficient; This is the second learnable weight coefficient; This is the third learnable weight coefficient; The post-processing module is used to filter the preliminary delivery platform detection results based on the confidence threshold and the crossover ratio threshold to obtain the final delivery platform detection results.

2. The system according to claim 1, characterized in that, The four wavelet sub-bands specifically include: ， in, It is a low-frequency sub-band. For low- and high-frequency detail subbands, For high and low frequency detail subbands, For high-frequency detail subbands; () indicates the operation of extracting LL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting LH subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting HL subbands after performing Haar discrete wavelet transform on the input features; () indicates the operation of extracting the HH subband after performing Haar discrete wavelet transform on the input features.

3. The system according to claim 1, characterized in that, The attention enhancement module also includes an inverse wavelet reconstruction output unit, used to first enhance the co-enhanced subband spatial tensor It is split along the sub-band dimension into four enhancement sub-bands, including For low-frequency structure enhancement subband, To enhance low and high frequency details, sub-bands To enhance high and low frequency details, sub-bands A sub-band for enhancing high-frequency details was created; subsequently, inverse wavelet reconstruction was performed on the four enhanced sub-bands to obtain the reconstructed incremental features. Finally, the incremental features will be reconstructed. Compared with the original l-th residual feature By summing the residuals, we obtain the enhanced output features. .

4. The system according to claim 1, characterized in that, The feature fusion network adopts a pyramid-style multi-scale fusion method. First, it performs channel alignment on features at different levels, and then performs upsampling, downsampling, and element-wise addition or splicing fusion to form three-scale fusion features: the second fusion feature, the third fusion feature, and the fourth fusion feature.

5. The system according to claim 1, characterized in that, The post-processing module specifically includes: First, based on the confidence threshold For the candidate box set and its corresponding confidence set Set of center points The process involves filtering out low-confidence candidate targets to obtain the remaining candidate targets, and then determining the target based on the intersection-union threshold. Non-maximum suppression is applied to the remaining candidate targets to obtain the final detection result of the drone delivery platform; the candidate targets are the set of candidate boxes. The corresponding confidence set and the set of center points A set of results at the same index position; the final drone delivery platform detection result includes the final candidate box set, the final confidence set, and the final center point set.

6. A delivery platform detection method based on sub-band spatial cooperative frequency domain attention, based on the delivery platform detection system based on sub-band spatial cooperative frequency domain attention as described in any one of claims 1-5, characterized in that, Includes the following steps: S1. Preprocess the delivery scene images captured by the UAV's onboard camera to obtain standard delivery scene images; S2. Input the standard delivery scene image into a pre-trained delivery platform detection network to obtain preliminary delivery platform detection results; S3. The preliminary delivery platform detection results are filtered according to the confidence threshold and the crossover ratio threshold to obtain the final delivery platform detection results.

Images