Airborne dynamic detection method for unmanned aerial vehicle to intercept low-slow small target, electronic equipment and medium

Abstract

The invention discloses an airborne dynamic detection method for an unmanned aerial vehicle to intercept a low-slow small target, electronic equipment and a medium. The method comprises the following steps: acquiring an original image, and performing feature extraction through a C2SWA module; carrying out feature fusion through an MDEN framework based on the extracted data; decoding based on the fused data to obtain a prediction result; and calculating a loss function of a prediction result and an actual result to obtain a final detection result. According to the invention, by reconstructing a network structure and an optimization strategy, the LSS-YOLO can effectively improve the real-time detection capability of a weak target, and meets the high-precision and high-efficiency requirements of urban low-altitude security and protection on the detection of low-altitude, slow and small targets such as an unmanned aerial vehicle.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle (UAV) airborne dynamic detection technology, and more specifically, to an airborne dynamic detection method, electronic equipment, and medium for a UAV to intercept a low, slow, and small target. Background Art

[0002] With the rapid development of the low-altitude economy, drones (UAVs), owing to their low cost, maneuverability, and wide range of applications, have been widely used in logistics and distribution, urban inspections, aerial photography, and emergency rescue. However, the rapid growth in the number of UAVs and the complex and diverse application scenarios have also brought with them a series of potential problems that cannot be ignored. When executing multi-UAV coordinated pursuit missions in urban low-altitude airspace, target perceptibility becomes a primary challenge. Low, slow, small targets (LSS)—low-altitude, slow-flying, small targets—are characterized by their compact size, slow flight speed, low radar cross-section, and strong camouflage. Their visual features exhibit problems such as low resolution, weak textures, and blurred edges. Especially in urban environments teeming with tall buildings, where background textures are complex and lighting fluctuates dramatically, LSS targets easily blend into the background, further complicating target recognition.

[0003] Traditional object detection algorithms generally face difficulties when detecting small, slow, and low-lying objects. Due to their limited feature extraction capabilities, they struggle to capture the faint and fuzzy features of these objects. Furthermore, their insensitivity to scale prevents them from accurately identifying and locating objects of varying sizes, leading to numerous missed detections and false detections, creating a technical bottleneck that hinders the ability to clearly detect these small, slow, and low-lying objects. Furthermore, due to the high computational demands of current detection technologies, there is a persistent conflict between algorithm accuracy and real-time performance, and between environmental adaptability and computational efficiency.

[0004] To overcome this bottleneck, recent technological developments have gradually focused on the field of deep learning. Researchers have attempted to leverage deep learning's powerful feature extraction and adaptive learning capabilities to improve the accuracy of low-speed, small-target detection. However, this process also faces many key technical difficulties. On the one hand, the problem of complex urban background interference is difficult to effectively address. Accurately extracting the features of low-speed, small-targets from the complex background information is a major challenge for technological development. On the other hand, the high sensitivity requirements for small targets are difficult to meet simultaneously. It is necessary to ensure the detection accuracy of tiny targets while taking into account the real-time and efficiency of the algorithm. This requires in-depth research and innovation in model architecture design, training strategy optimization, and data preprocessing.

[0005] Therefore, it is necessary to develop an airborne dynamic detection method, electronic equipment and medium for UAVs to intercept low, slow and small targets.

[0006] The information disclosed in the background technology section of the present invention is only intended to deepen the understanding of the general background technology of the present invention, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the Invention

[0007] The present invention proposes an airborne dynamic detection method, electronic equipment and medium for drones to intercept low-speed, slow and small targets. It constructs a low-speed, slow and small target detector LSS-YOLO based on the currently popular target detection framework. By performing targeted optimization on the head module, neck network and loss function, the performance of the low-speed, slow and small target detector is significantly improved.

[0008] In a first aspect, the embodiments of the present disclosure provide an airborne dynamic detection method for a drone to intercept a low, slow, and small target, including:

[0009] Get the original image and perform feature extraction through the C2SWA module;

[0010] Based on the extracted data, feature fusion is performed through the MDEN framework;

[0011] Decode based on the fused data to obtain the prediction result;

[0012] Calculate the loss function between the predicted results and the actual results to obtain the final detection results.

[0013] Preferably, the C2SWA module uses Haar wavelet transform to perform multi-resolution decomposition on the original image.

[0014] Preferably, for the original image Generate four subbands through two-dimensional discrete Haar wavelet transform:

[0015]

[0016] Among them, h low 、h high is the basis function of Haar wavelet, The low-frequency LL subband preserves the global contour and large-scale structural information, while the high-frequency LH, HL, and HH subbands capture fine details in the horizontal, vertical, and diagonal directions, respectively.

[0017] Preferably, a top-down and bottom-up bidirectional feature transfer mechanism and weighted fusion are introduced to construct a lightweight bidirectional feature pyramid network to simplify the structural parameters of the MDEN framework.

[0018] Preferably, the top-down and bottom-up bidirectional feature transfer mechanism includes:

[0019] In the Top-Down path, sub-pixel convolution is used to rearrange the deep features to improve the resolution:

[0020]

[0021] Subpixel(P l+1 )=PixelShuffle(Conv 3×3 (P l+1 ))

[0022] Among them, the PixelShuffle operation compresses the number of channels by 4 times and directly splices them with the shallow features;

[0023] On the bottom-up path, dilated convolutions are embedded in shallow features to expand the receptive field and enhance the contextual perception capabilities of high-resolution layers:

[0024]

[0025] When k=3, d=3, the equivalent receptive field of the dilated convolution is:

[0026] RF eff =(k-1)·d+1=(3-1)·3+1=7.

[0027] Preferably, weighted fusion includes:

[0028] The extracted data is weighted fused using a fast normalization fusion method:

[0029]

[0030] in, is a learnable weight coefficient, constrained by ReLU activation is the input feature map, and the stable numerical coefficient is ε = 0.0001;

[0031] Introducing target motion state perception factors:

[0032] γ=0.5+0.5·sigmoid(v t / 10)

[0033] Among them, v t is the target instantaneous velocity, when v t When v > 20, it is considered a high-speed target, γ→0.8, and motion blur compensation is enhanced; when v t When <5, it is considered to be a hovering target, γ→0.5, maintaining static details;

[0034] For the fusion process of the intermediate layer features, the features of each layer are combined through a weighted mechanism, and the two fused features are:

[0035]

[0036]

[0037] Among them, P i Td is the intermediate feature of the i-th layer, P i Out is the output feature of the jth layer, GhostConv is a lightweight convolution operation, and Resize is an upsampling or downsampling operation.

[0038] Preferably, the loss function is an SSD-IoU loss function that integrates the constraints of scale, shape, and distance.

[0039] Preferably, constructing the loss function includes:

[0040] Perform scale constraint modeling, define the bounding box scale difference factor S, and use the normalized area ratio to measure the size match between the target and the predicted box:

[0041]

[0042] Among them, A gt 、A pred are the true value and the predicted box area, w gt ,h gt The denominator is the actual width and height of the target, and the diagonal length is normalized;

[0043] The aspect ratio similarity measure R is introduced in shape constraint modeling as:

[0044]

[0045] The Wasserstein distance is used to measure the similarity of distributions:

[0046]

[0047] Where μ is the Gaussian mean, σ is the standard deviation, and C = 12.8 is the normalization constant;

[0048] The loss function is:

[0049]

[0050] Among them, the weight coefficient is generated by the two-stream attention network:

[0051]

[0052] in, is the IoU gradient, F confis the classification confidence feature, and σ is the Sigmoid function.

[0053] In a second aspect, an embodiment of the present disclosure further provides an electronic device, the electronic device comprising:

[0054] a memory storing executable instructions;

[0055] A processor runs the executable instructions in the memory to implement the airborne dynamic detection method for a UAV to intercept a low, slow, and small target.

[0056] In a third aspect, an embodiment of the present disclosure further provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the airborne dynamic detection method for a UAV to intercept low, slow and small targets.

[0057] Its beneficial effects are:

[0058] (1) We construct a cross-subband wavelet attention module (C2SWA) that combines the frequency domain decomposition mechanism with the attention fusion strategy, and propose a backbone network optimization method based on the C2SWA module, which significantly improves the model's ability to perceive small targets and its anti-interference performance;

[0059] (2) A multi-scale dynamic enhancement network (MDEN) is proposed, which includes a lightweight bidirectional feature pyramid, a lightweight attention module, and an adaptive weight fusion mechanism. Through cross-resolution interactive enhancement, multi-scale attention guidance, and lightweight feature fusion design, it achieves highly robust detection in complex scenarios.

[0060] (3) Aiming at the problems of bounding box positioning sensitivity, shape deviation and distance coupling in small target detection from the perspective of drones, a scale, shape and distance intersection of union loss function (SSD IoU) is constructed, which integrates the constraints of bounding box scale, shape and distance. By constructing a multi-dimensional geometric constraint field, the bounding box positioning error is compressed to meet the precise control requirements of the pursuit system and realize the joint optimization of the bounding box parameters.

[0061] The methods and apparatus of the present invention have other features and advantages that will be apparent from or will be described in detail in the accompanying drawings and subsequent detailed descriptions incorporated herein, which together serve to explain the specific principles of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] The above and other objects, features and advantages of the present invention will become more apparent through a more detailed description of exemplary embodiments of the present invention with reference to the accompanying drawings, wherein like reference numerals generally represent like components throughout the exemplary embodiments of the present invention.

[0063] Figure 1 A flowchart showing the steps of an airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to an embodiment of the present invention is shown.

[0064] Figure 2 A schematic diagram of a low-speed, slow, and small target algorithm framework for complex urban background interference according to an embodiment of the present invention is shown.

[0065] Figure 3 A diagram showing a low-speed, slow, and small target detection network architecture according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0066] The preferred embodiments of the present invention will be described in more detail below. Although the preferred embodiments of the present invention are described below, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein.

[0067] Figure 1 A flowchart showing the steps of an airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to an embodiment of the present invention is shown.

[0068] like Figure 1 As shown, the airborne dynamic detection method for UAV to intercept low, slow and small targets includes:

[0069] Step 101: Obtain the original image and perform feature extraction through the C2SWA module;

[0070] Step 102: performing feature fusion based on the extracted data through the MDEN framework;

[0071] Step 103: Decode based on the fused data to obtain a prediction result;

[0072] Step 104: Calculate the loss function between the predicted result and the actual result to obtain the final detection result.

[0073] In one example, the C2SWA module uses Haar wavelet transform to perform multi-resolution decomposition on the original image.

[0074] In one example, for the original image Generate four subbands through two-dimensional discrete Haar wavelet transform:

[0075]

[0076] Among them, h low 、h high is the basis function of Haar wavelet, The low-frequency LL subband preserves the global contour and large-scale structural information, while the high-frequency LH, HL, and HH subbands capture fine details in the horizontal, vertical, and diagonal directions, respectively.

[0077] In one example, a top-down and bottom-up bidirectional feature transfer mechanism and weighted fusion are introduced to construct a lightweight bidirectional feature pyramid network to simplify the structural parameters of the MDEN framework.

[0078] In one example, the top-down and bottom-up bidirectional feature transfer mechanism includes:

[0079] In the Top-Down path, sub-pixel convolution is used to rearrange the deep features to improve the resolution:

[0080]

[0081] Subpixel(P l+1 )=PixelShuffle(Conv 3×3 (P l+1 ))

[0082] Among them, the PixelShuffle operation compresses the number of channels by 4 times and directly splices them with the shallow features;

[0083] On the bottom-up path, dilated convolutions are embedded in shallow features to expand the receptive field and enhance the contextual perception capabilities of high-resolution layers:

[0084]

[0085] When k=3, d=3, the equivalent receptive field of the dilated convolution is:

[0086] RF eff =(k-1)·d+1=(3-1)·3+1=7.

[0087] In one example, weighted fusion includes:

[0088] The extracted data is weighted fused using a fast normalization fusion method:

[0089]

[0090] in, is a learnable weight coefficient, constrained by ReLU activation is the input feature map, and the stable numerical coefficient is ε = 0.0001;

[0091] Introducing target motion state perception factors:

[0092] γ=0.5+0.5·sigmoid(v t / 10)

[0093] Among them, v t is the target instantaneous velocity, when v t When v > 20, it is considered a high-speed target, γ→0.8, and motion blur compensation is enhanced; when v t When <5, it is considered to be a hovering target, γ→0.5, maintaining static details;

[0094] For the fusion process of the intermediate layer features, the features of each layer are combined through a weighted mechanism, and the two fused features are:

[0095]

[0096]

[0097] Among them, P i Td is the intermediate feature of the i-th layer, P i Out is the output feature of the jth layer, GhostConv is a lightweight convolution operation, and Resize is an upsampling or downsampling operation.

[0098] In one example, the loss function is an SSD-IoU loss function that integrates three attribute constraints: scale, shape, and distance.

[0099] In one example, constructing the loss function includes:

[0100] Perform scale constraint modeling, define the bounding box scale difference factor S, and use the normalized area ratio to measure the size match between the target and the predicted box:

[0101]

[0102] Among them, A gt 、A pred are the true value and the predicted box area, w gt ,h gt The denominator is the actual width and height of the target, and the diagonal length is normalized;

[0103] The aspect ratio similarity measure R is introduced in shape constraint modeling as:

[0104]

[0105] The Wasserstein distance is used to measure the similarity of distributions:

[0106]

[0107] Where μ is the Gaussian mean, σ is the standard deviation, and C = 12.8 is the normalization constant;

[0108] The loss function is:

[0109]

[0110] Among them, the weight coefficient is generated by the two-stream attention network:

[0111]

[0112] in, is the IoU gradient, F conf is the classification confidence feature, and σ is the Sigmoid function.

[0113] Figure 2 A schematic diagram of a low-speed, slow, and small target algorithm framework for complex urban background interference according to an embodiment of the present invention is shown.

[0114] Specifically, the present invention is a dynamic recognition and detection method for drones intercepting low, slow, and small targets in complex urban environments. By performing targeted optimization on the head module, neck network, and loss function, the performance of the low, slow, and small target detector is significantly improved, and the problems of weak significance and scale insensitivity in traditional methods are solved. Under the premise of ensuring real-time performance, higher precision and recall rate are achieved than traditional algorithms, effectively improving the real-time and accuracy of dynamic recognition and detection tasks of low, slow, and small targets. Research framework such as Figure 2 shown.

[0115] Figure 3 A diagram showing a low-speed, slow, and small target detection network architecture according to an embodiment of the present invention is shown.

[0116] The low-speed, slow, and small target detection algorithm LSS-YOLO is based on the YOLOv11 network architecture. It has made a series of optimization improvements to the difficulty of detecting low-speed, slow, and small targets under complex urban background interference. The overall network architecture is as follows: Figure 3 First, in the backbone network, a C2SWA module is designed to reduce the false alarm rate in complex backgrounds by using a joint fusion method of frequency-domain and spatial-domain features. Secondly, in the neck network, a lightweight convolution module is introduced to optimize the BiFPN network, and the MDEN architecture is designed to reduce the computational burden while solving the problem of weakened small target features. In addition, the SSD IoU loss function is introduced to reduce the influence of the appearance of low-speed and small targets, achieving more accurate regression of the detection box.

[0117] To address the challenges of feature ambiguity and background noise in detecting small, slow, and low-lying drone targets, a backbone network optimization method based on the C2SWA module is proposed. This method significantly improves the model's small target perception and interference resistance through the collaborative design of wavelet-domain multi-band decomposition and cross-subband attention fusion. The C2SWA module consists of a wavelet decomposition layer and cross-subband attention gating.

[0118] Wavelet convolution WTConv uses wavelet transform and inverse wavelet transform to process the multi-scale information of the input image, thereby enhancing the model's ability to extract detailed features, especially for detecting low, slow, and small targets in urban backgrounds. To achieve frequency-domain and spatial domain joint feature modeling, C2SWA uses Haar wavelet transform to perform multi-resolution decomposition of input features. Given an input feature map Generate four subbands through two-dimensional discrete Haar wavelet transform:

[0119]

[0120] in, and is the basis function of the Haar wavelet. The low-frequency LL subband retains global contours and large-scale structural information, while the high-frequency LH / HL / HH subbands capture fine details in the horizontal, vertical, and diagonal directions, respectively. Unlike fixed-size convolution, WTConv divides the input features into the frequency domain, enabling the network to more effectively distinguish high-frequency edges and low-frequency backgrounds. It performs convolution operations on the high-frequency subbands, and then hierarchically reversely combines the stored LL and high-frequency subbands and reconstructs the features to the original resolution. Finally, the inverse wavelet transform is used to add the reconstructed wavelet features to the original features through residual connections, and resynthesize them into a complete output, ensuring seamless integration of multi-frequency information. This enhances the detection of low, slow, and small targets and the extraction of complex texture details.

[0121] Although wavelet decomposition and reconstruction introduce additional computational steps, WTConv is still lightweight compared to traditional large kernel convolution or repeated stacking of complex layers. This design maintains computational efficiency while significantly improving the detection accuracy of small objects and high-detail features.

[0122] Small, slow, and low-lying targets exhibit significant scale variations during dynamic flight. The pixel ratios during hovering and high-speed dives differ significantly. Traditional feature fusion methods suffer from cross-resolution semantic fragmentation, resulting in insufficient interaction between shallow detail features and deep semantic features. BiFPN, a core feature fusion mechanism, effectively integrates multi-scale features input from the backbone network. However, the cross-layer connections of traditional BiFPN increase the number of parameters and lead to computational redundancy.

[0123] To address the problems of weakened features and scale sensitivity of small targets from the perspective of drones, a multi-scale dynamic enhancement network (MDEN) is proposed, which includes a lightweight bidirectional feature pyramid, a cross-scale token attention module and an adaptive weight fusion mechanism. Through cross-resolution interactive enhancement, multi-scale attention guidance and lightweight feature fusion design, highly robust detection in complex scenarios is achieved.

[0124] Although traditional feature pyramid networks (FPNs) can enhance semantic information in high-level feature maps, they often lose positional information in low-level feature maps during the fusion process, affecting object detection accuracy. To overcome this shortcoming, BiFPN builds on FPN by introducing a bidirectional feature transfer mechanism, top-down and bottom-up, enabling simultaneous transfer of positional and semantic information, significantly improving object detection accuracy. A lightweight bidirectional feature pyramid network (BiFPN-Lite) is constructed to achieve more efficient fusion of multi-level feature information.

[0125] In the Top-Down path, sub-pixel convolution is used to rearrange the deep features (P5) to improve the resolution:

[0126]

[0127] Subpixel(P l+1 )=PixelShuffle(Conv 3×3 (P l+1 )) (3)

[0128] The PixelShuffle operation compresses the number of channels by a factor of 4 (8x upsampling) and directly concatenates them with the shallow features (P4). In the bottom-up path, a dilated convolution (dilation = 3) is embedded in the shallow features to expand the receptive field and enhance the contextual awareness of the high-resolution layer.

[0129]

[0130] The equivalent receptive field of the dilated convolution (dilation=3) is calculated as:

[0131] RF eff =(k-1)·d+1=(3-1)·3+1=7 (5)

[0132] By cascading 3 layers of dilated convolution, the final receptive field is expanded to 7 3 =343, effectively capturing long-range context associations.

[0133] In the feature fusion process, a dynamic weighting mechanism is used to adjust the contribution of different feature maps. Considering the training efficiency and performance, in order to balance the contribution of multi-scale features, a fast normalization fusion method is selected to perform weighted fusion on the input feature group {P3, P4, P5}:

[0134]

[0135] in, is a learnable weight coefficient, constrained by ReLU activation is the input feature map, and the stable numerical coefficient is ε = 0.0001.

[0136] Introducing target motion state perception factors:

[0137] γ=0.5+0.5·sigmoid(v t / 10) (7)

[0138] Among them, v t is the target instantaneous speed (pixels / frame). t When v > 20, it is considered a high-speed target, γ→0.8, and motion blur compensation is enhanced; when v t When γ is less than 5, it is considered to be a hovering target, and γ→0.5 maintains static details.

[0139] For the fusion process of the intermediate layer features, the features of each layer are combined through a weighted mechanism. The specific formula of the two fused features is:

[0140]

[0141]

[0142] Among them, P i Td is the intermediate feature of the i-th layer, P i Out is the output feature of the jth layer. GhostConv is a lightweight convolution operation that expands the receptive field by using fewer feature channels. Resize is an upsampling or downsampling operation. Through these improvements, we constructed the lightweight BiFPN-Lite, which effectively simplifies the parameters of the MDEN neck network structure. This not only effectively captures positional and semantic information in complex scenes, but also ensures more accurate fusion of features at different scales.

[0143] Specifically, the MDEN neck network offers the following important improvements for our task. First, the network structure is simplified, preserving sufficient information in low-speed, small-object features. This reduces computational complexity without sacrificing performance. Second, by adding feature branches between the original input and output nodes of the same layer, the richness of feature fusion is improved, thereby enhancing the network's expressiveness. Finally, due to the varying resolutions of feature maps at different scales, BiFPN-Lite utilizes a trainable weighting mechanism to balance the influence of each feature layer, achieving deep feature fusion.

[0144] To address the issues of bounding box positioning sensitivity, shape deviation, and distance coupling in small target detection from drones, we propose an SSD-IoU loss function that integrates three attribute constraints: scale, shape, and distance. Its core idea is to achieve joint optimization of bounding box parameters by constructing a multi-dimensional geometric constraint field.

[0145] The first step is scale constraint modeling. The bounding box scale difference factor S is defined, and the normalized area ratio is used to measure the size match between the target and the predicted box:

[0146]

[0147] Among them, A gt 、A pred are the true value and the predicted box area, w gt ,h gt The denominator is normalized by the diagonal length to represent the true width and height of the target, eliminating the impact of the target's absolute size on the loss. This design adaptively balances the scale penalty weights for large and small targets.

[0148] In shape constraint modeling, the aspect ratio similarity metric R is introduced, which draws on the aspect ratio penalty term of CIoU and is improved into a hyperbolic tangent function:

[0149]

[0150] The first term captures aspect ratio differences, while the second term suppresses extreme shape deviations (such as extremely flat or long boxes) through the tanh function, addressing the square-like problem caused by extreme maneuvering of small, slow, and low-lying objects. In object detection, when the predicted box and the ground-truth box have no overlap, the IoU = 0 and there is no gradient update signal, resulting in the zero-overlap vanishing gradient problem. Furthermore, the traditional IoU loss does not adequately penalize width and height deviations of small objects, resulting in bounding box offsets at the edges of drone blades and high sensitivity to drone shape.

[0151] The bounding box is modeled as a two-dimensional Gaussian distribution, and the distribution similarity is measured by Wasserstein distance:

[0152]

[0153] Where μ is the Gaussian mean, σ is the standard deviation, and C = 12.8 is the normalization constant. This metric is less sensitive to small position deviations (<5 pixels) than IoU and can still provide effective gradients when there is no overlap. To avoid the optimization direction deviation caused by fixed weights, a dynamic fusion strategy based on the target state is proposed:

[0154]

[0155] The weight coefficients are generated through a two-stream attention network:

[0156]

[0157] in, is the IoU gradient, F conf is the classification confidence feature, and σ is the Sigmoid function. This mechanism enables the model to focus on distance optimization in the early stages of training and strengthen shape constraints in the later stages.

[0158] SSD-IoU, for the first time, jointly models the three elements of scale, shape, and distance, overcoming the geometric constraints of traditional IoU loss through a dynamic coupling mechanism. This design demonstrates significant advantages in small target detection tasks for drones, laying a theoretical foundation for the construction of low-speed, slow, and small target detectors.

[0159] The key points and points to be protected of the present invention include:

[0160] 1. Cross-Subband Wavelet Attention Module (C2SWA): This module proposes an algorithm based on the fusion of wavelet frequency domain decomposition and spatial channel attention. It uses multi-level subband decomposition and injects attention weighting strategy into each subband to achieve accurate extraction of weak, low-contrast small target texture and edge features and suppress background interference.

[0161] 2. Multi-scale Dynamic Enhancement Network (MDEN): This design utilizes a feature fusion network based on bidirectional information flow and learnable fusion weights. It combines cross-resolution interaction enhancement, multi-scale attention guidance, and a lightweight fusion unit to achieve highly robust detection and real-time performance for small objects at multiple scales in complex urban scenarios.

[0162] 3. Multi-attribute constrained intersection-over-union (SSD-IoU) loss function: A joint constraint loss is constructed that integrates the three-dimensional geometric attributes of bounding box scale, shape, and detection distance. By jointly penalizing the deviation between the predicted box and the true box in the multi-dimensional attribute space, it achieves accurate compression of the bounding box positioning error, meeting the requirements of high-precision interception control of drones.

[0163] The present invention also provides an electronic device, which includes: a memory storing executable instructions; and a processor running the executable instructions in the memory to implement the above-mentioned airborne dynamic detection method for a UAV to intercept low, slow and small targets.

[0164] The present invention also provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned airborne dynamic detection method for unmanned aerial vehicles to intercept low, slow and small targets.

[0165] To facilitate understanding of the solutions and effects of the embodiments of the present invention, three specific application examples are given below. Those skilled in the art should understand that these examples are only for facilitating understanding of the present invention, and any specific details thereof are not intended to limit the present invention in any way.

[0166] Example 1

[0167] To validate the detection performance of the proposed dynamic recognition and detection method for low-lying, slow, and small objects in complex environments, two sets of experiments were designed: a module ablation experiment and a comparative experiment with a mainstream algorithm. These experiments compared the impact of each proposed module on the overall algorithm and the performance of the proposed algorithm with that of mainstream algorithms. The experiments used the self-constructed CounterUAV dataset, which contains six categories of low-lying, slow, and small objects, 12 complex urban scenes, and various lighting conditions, making it highly representative and challenging. All experiments were conducted on Ubuntu 22.04, Python 3.10, and PyTorch 2.1, using an NVIDIA RTX 4090 GPU for training and inference. YOLOv11 was used as the base detection framework for training, with an input image resolution of 640×640. SGD was used as the optimizer, with an initial learning rate of 0.01, a total of 400 training epochs, and a batch size of 16. All models were trained using the same data partitioning and random seed to ensure comparability of the experimental results.

[0168] The evaluation metrics used are four common indicators in the field of object detection: Precision and Recall are used to measure the accuracy and completeness of detection results. mAP50 and mAP50-95 represent the average precision at an IoU threshold of 0.5 and the average of multiple IoUs, respectively. Higher values ​​indicate better detector performance. Params / GFLOPs represent the number of model parameters and computational complexity, respectively, reflecting the model's lightweightness. FPS measures the model's inference speed, measured in frames per second, reflecting its real-time performance.

[0169] To clarify the contribution of each submodule to detection performance, ablation experiments were conducted on various model configurations to verify the effectiveness and stability of the low-speed, slow, and small object detection algorithm. The LSS-YOLO algorithm was disassembled into its individual modules. Using YOLOv11 as the baseline network, experiments were conducted with the C2SWA, MDEN, and SSD-Loss modules. Table 1 shows the results on the CounterUAV dataset. The experimental results show that improvements in each component lead to optimal results. LSS-YOLO achieves the best performance in accuracy, recall, and mAP50, and also performs well in the other two metrics. While maintaining a lower parameter count than the baseline model, the entire detection algorithm achieves mAP50 and mAP50-95 scores of 0.8691 and 0.57216, respectively, demonstrating an excellent precision-complexity balance. The recall rate also reaches 79.747%, exceeding that of other configurations and demonstrating its comprehensive detection capabilities.

[0170] Table 1 LSS-YOLO ablation experiment

[0171]

[0172] To further validate the advanced nature of the detection algorithm, the constructed LSS-YOLO was compared with current mainstream object detection models on the public Drone vs Bird dataset. The results are shown in Table 2. The comparison results show that our detector achieves the best mAP50-95 performance, demonstrating stronger generalization capabilities on more stringent target metrics. Although its mAP50 is slightly lower than YOLOv11, its lightweight design makes it more suitable for embedded deployment scenarios such as drones. Overall, the proposed low-speed, slow, and small object detector achieves a good balance between precision, recall, speed, and model complexity, demonstrating significant practical deployment value.

[0173] Table 2 Comparative experiment

[0174]

[0175]

[0176] Example 2

[0177] The present disclosure provides an electronic device, which includes: a memory storing executable instructions; and a processor running the executable instructions in the memory to implement the above-mentioned airborne dynamic detection method for a drone to intercept a low, slow, and small target.

[0178] An electronic device according to an embodiment of the present disclosure includes a memory and a processor.

[0179] The memory is used to store non-transitory computer-readable instructions. Specifically, the memory may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, flash memory, etc.

[0180] The processor may be a central processing unit (CPU) or other form of processing unit having data processing capability and / or instruction execution capability, and may control other components in the electronic device to perform desired functions. In one embodiment of the present disclosure, the processor is used to execute the computer-readable instructions stored in the memory.

[0181] Those skilled in the art should understand that in order to solve the technical problem of how to obtain a good user experience, this embodiment may also include well-known structures such as a communication bus and an interface, and these well-known structures should also be included in the scope of protection of this disclosure.

[0182] For detailed description of this embodiment, please refer to the corresponding description in the aforementioned embodiments, which will not be repeated here.

[0183] Example 3

[0184] An embodiment of the present disclosure provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the computer program implements the airborne dynamic detection method for a drone to intercept a low, slow, and small target.

[0185] According to an embodiment of the present disclosure, a computer-readable storage medium stores non-transitory computer-readable instructions, which, when executed by a processor, execute all or part of the steps of the aforementioned methods of the embodiments of the present disclosure.

[0186] The above-mentioned computer-readable storage media include, but are not limited to, optical storage media (e.g., CD-ROMs and DVDs), magneto-optical storage media (e.g., MOs), magnetic storage media (e.g., magnetic tapes or mobile hard disks), media with built-in rewritable non-volatile memory (e.g., memory cards), and media with built-in ROM (e.g., ROM cartridges).

[0187] Those skilled in the art should understand that the above description of the embodiments of the present invention is only for the purpose of illustrative purposes only to illustrate the beneficial effects of the embodiments of the present invention, and is not intended to limit the embodiments of the present invention to any given examples.

[0188] While various embodiments of the present invention have been described above, the above description is intended to be illustrative, not exhaustive, and not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. An airborne dynamic detection method for a UAV to intercept a low, slow, and small target, characterized in that: include: Get the original image and perform feature extraction through the C2SWA module; Based on the extracted data, feature fusion is performed through the MDEN framework; Decode based on the fused data to obtain the prediction result; Calculate the loss function between the predicted results and the actual results to obtain the final detection results.

2. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 1, wherein: The C2SWA module uses Haar wavelet transform to perform multi-resolution decomposition on the original image.

3. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 2, wherein: For the original image Generate four subbands through two-dimensional discrete Haar wavelet transform: Among them, h low 、h high is the basis function of Haar wavelet, The low-frequency LL subband preserves the global contour and large-scale structural information, while the high-frequency LH, HL, and HH subbands capture fine details in the horizontal, vertical, and diagonal directions, respectively.

4. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 1, wherein: By introducing top-down and bottom-up bidirectional feature transfer mechanisms and weighted fusion to construct a lightweight bidirectional feature pyramid network, the structural parameters of the MDEN framework are simplified.

5. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 4, wherein: The top-down and bottom-up bidirectional feature transfer mechanisms include: In the Top-Down path, sub-pixel convolution is used to rearrange the deep features to improve the resolution: Subpixel(P l+1 )=PixelShuffle(Conv 3×3 (P l+1 )) Among them, the PixelShuffle operation compresses the number of channels by 4 times and directly splices them with the shallow features; On the bottom-up path, dilated convolutions are embedded in shallow features to expand the receptive field and enhance the contextual perception capabilities of high-resolution layers: When k=3, d=3, the equivalent receptive field of the dilated convolution is: RF eff =(k-1)·d+1=(3-1)·3+1=7。 6. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 4, wherein: Weighted fusion includes: The extracted data is weighted fused using a fast normalization fusion method: in, is a learnable weight coefficient, constrained by ReLU activation is the input feature map, and the stable numerical coefficient is ε = 0.0001; Introducing target motion state perception factors: γ=0.5+0.5·sigmoid(v t / 10) Among them, v t is the target instantaneous velocity, when v t When v > 20, it is considered a high-speed target, γ→0.8, and motion blur compensation is enhanced; when v t When <5, it is considered to be a hovering target, γ→0.5, maintaining static details; For the fusion process of the intermediate layer features, the features of each layer are combined through a weighted mechanism, and the two fused features are: Among them, P i Td is the intermediate feature of the i-th layer, P i Out is the output feature of the jth layer, GhostConv is a lightweight convolution operation, and Resize is an upsampling or downsampling operation.

7. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 1, wherein: The loss function is an SSD-IoU loss function that integrates the constraints of scale, shape, and distance.

8. The airborne dynamic detection method for a UAV to intercept a low, slow, and small target according to claim 7, wherein: Constructing the loss function includes: Perform scale constraint modeling, define the bounding box scale difference factor S, and use the normalized area ratio to measure the size match between the target and the predicted box: Among them, A gt 、A pred are the true value and the predicted box area, w gt ,h gt The denominator is the actual width and height of the target, and the diagonal length is normalized; The aspect ratio similarity measure R is introduced in shape constraint modeling as: The Wasserstein distance is used to measure the similarity of distributions: Where μ is the Gaussian mean, σ is the standard deviation, and C = 12.8 is the normalization constant; The loss function is: Among them, the weight coefficient is generated by the two-stream attention network: in, is the IoU gradient, F conf is the classification confidence feature, and σ is the Sigmoid function.

9. An electronic device, characterized in that: The electronic device comprises: a memory storing executable instructions; A processor that runs the executable instructions in the memory to implement the airborne dynamic detection method for a drone to intercept a low, slow, and small target as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the airborne dynamic detection method for a drone to intercept a low, slow, and small target according to any one of claims 1 to 8.

Images