Unmanned aerial vehicle target detection method based on space-time attention mechanism and track matching
By deploying detection methods based on space-time attention mechanism and track matching on drones, the drone target detection challenges of traditional systems in dynamic environments and variable lighting conditions are solved, and efficient and accurate drone identification is achieved. and tracking, improving the flexibility and coverage of the monitoring system.
Patent Information
- Application Number
- CN202411177319.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-26
- Publication Date
- 2024-12-03
AI Technical Summary
Traditional ground or static platform-based drone monitoring systems have limitations in dynamic environments and rapid responses in vast areas, especially in terms of dynamic background adaptation, stable performance under variable lighting conditions and standard tracking of high-speed motion objects. Drone target detection.
Adopting a drone target detection method based on spatiotemporal attention mechanism and track matching, by obtaining visible light video sequences, processing single-frame images, determining image features, and using algorithm models to detect targets, combining track correlation technology, intercepting three consecutive times The frame image is generated and correlation judgment is performed. If the correlation coefficient exceeds the preset value, it is determined to be the same drone.
It significantly improves the recognition efficiency and reliability of drone target detection, reduces misjudgment and repeated tracking, optimizes system performance, and improves task efficiency and accuracy.
Smart Images

Figure CN119068367A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of target detection technology, and in particular to a method for detecting unmanned aerial vehicle targets based on spatiotemporal attention mechanism and track matching. Background Art
[0002] In the current security and surveillance field, drones have become a common and versatile tool for performing a variety of tasks including border patrols, traffic monitoring, environmental monitoring, etc. However, with the widespread use of drones, the security threats they bring are also increasing, especially for illegal, hostile or drones that mistakenly enter no-fly zones. These threats need to be addressed in an efficient and reliable way.
[0003] Therefore, it is urgent to develop an efficient UAV-on-UAV target detection system. Traditional ground or static platform-based UAV monitoring systems can effectively monitor UAVs to a certain extent, but they have limitations in rapid response in dynamic environments and wide areas. To overcome these limitations, deploying target detection systems on UAVs to enable mobile monitoring in the air has become a promising solution. This approach not only increases the flexibility and coverage of monitoring, but also provides a more multi-angle surveillance field of view. In UAV-on-UAV systems, visible light imaging technology has attracted much attention due to its advantages in environmental adaptability and detail capture. Compared with infrared or other sensing technologies, visible light imaging can provide high-resolution images, which is crucial for identifying small or long-distance targets. However, using visible light technology for UAV target detection faces several challenges, including adaptation to dynamic backgrounds, stable performance under changing lighting conditions, and accurate tracking of high-speed moving targets. Summary of the invention
[0004] The technical problem to be solved by the present invention is how to effectively extract image features and enhance the accuracy and stability of target detection; in view of this, the present invention provides a method and device for drone target detection based on spatiotemporal attention mechanism and track matching.
[0005] The technical solution adopted by the present invention is a method for detecting unmanned aerial vehicle targets based on spatiotemporal attention mechanism and track matching, comprising:
[0006] Step S1, obtaining a known UAV visible light video sequence, and processing the video sequence into a single frame image;
[0007] Step S2, determining corresponding image features according to the image samples formed by the single frame image;
[0008] Step S3, based on the image features, using the configured algorithm model, determining the drone target detection result corresponding to the image sample;
[0009] Step S4, capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
[0010] In one embodiment, the step S2 includes: processing the image sample using a spatiotemporal attention mechanism, specifically including:
[0011] Acquire a plurality of track points represented by the single-frame images, which include input feature vectors of corresponding multiple features, and obtain an attention weight vector by mapping the input feature vectors through a single-layer neural network;
[0012] Normalize each attention weight coefficient to obtain the attention weight of each feature;
[0013] Using the attention weights, the input feature vector is re-determined as a weighted vector, that is, the output of the spatiotemporal attention mechanism at the current moment.
[0014] In one embodiment, the step S3 further includes an optimization process for the algorithm model, and the optimization process is specifically implemented by the following formula:
[0015]
[0016] Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively; N pos is the number of positive samples, loss VFL is the VFL cross entropy loss, 1oss GIOU =1-GIOU,loss DFL is the DFL cross entropy loss function, 1oss GIOU is the loss function of GIoU, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
[0017] In one embodiment, step S4 includes:
[0018] Three frames of visible light images are captured from the beginning, middle and end of the target's track respectively;
[0019] Perform convolution processing on the visible light image, merge the feature maps of the three frames of image data after processing in channel order, enhance the features through the downsampling convolution layer and channel attention mechanism, and generate a feature map;
[0020] The feature map is dimensionally adjusted and feature mapped through a fully connected layer, and a correlation coefficient is determined by comparing the feature difference between the feature map and the drone target detection result;
[0021] If the correlation coefficient exceeds the preset threshold, it is determined that the target and the known track belong to the same UAV.
[0022] Another aspect of the present invention further provides a UAV target detection device with spatiotemporal attention mechanism and track matching, comprising:
[0023] An acquisition unit is configured to acquire a known UAV visible light video sequence and process the video sequence into a single frame image;
[0024] A feature unit, configured to determine corresponding image features according to the image sample formed by the single frame image;
[0025] A detection unit configured to determine a drone target detection result corresponding to the image sample based on the image feature and using a configured algorithm model;
[0026] The matching unit is configured to capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
[0027] In one embodiment, the feature unit is further configured to: process the image sample using a spatiotemporal attention mechanism, specifically including:
[0028] Acquire a plurality of track points represented by the single-frame images, which include input feature vectors of corresponding multiple features, and obtain an attention weight vector by mapping the input feature vectors through a single-layer neural network;
[0029] Normalize each attention weight coefficient to obtain the attention weight of each feature;
[0030] Using the attention weights, the input feature vector is re-determined as a weighted vector, that is, the output of the spatiotemporal attention mechanism at the current moment.
[0031] In one embodiment, the detection unit is further configured to optimize the algorithm model, and the optimization process is specifically implemented by the following formula:
[0032]
[0033] Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively; N pos is the number of positive samples, loss VFLis the VFL cross entropy loss, 1oss GIOU =1-GIOU,loss DFL is the DFL cross entropy loss function, 1oss GIOU is the loss function of GIoU, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
[0034] In one embodiment, the matching unit is further configured as follows:
[0035] Capture three frames of visible light images from the start, middle and end of the target's track respectively;
[0036] Perform convolution processing on the visible light image, merge the feature maps of the three frames of image data after processing in channel order, enhance the features through the downsampling convolution layer and channel attention mechanism, and generate a feature map;
[0037] The feature map is dimensionally adjusted and feature mapped through a fully connected layer, and a correlation coefficient is determined by comparing the feature difference between the feature map and the drone target detection result;
[0038] If the correlation coefficient exceeds the preset threshold, it is determined that the target and the known track belong to the same UAV.
[0039] Another aspect of the present invention provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the steps of the drone target detection method based on spatiotemporal attention mechanism and track matching as described in any of the above items.
[0040] Another aspect of the present invention further provides a computer storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the drone target detection method based on spatiotemporal attention mechanism and track matching as described in any of the above items are implemented.
[0041] Compared with the prior art, the present invention has at least the following advantages:
[0042] Compared with the traditional single image processing method, the present invention can more effectively cope with the challenges of target detection in different scenarios and significantly improve the recognition efficiency and reliability of the results; moreover, by combining the track correlation technology, the present invention can effectively determine whether it is multiple tracking of the same target, thereby reducing misjudgment and repeated tracking, optimizing system performance, and not only reducing the workload, but also improving the efficiency and accuracy of task execution. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] Figure 1A schematic diagram of a flow chart of a method for detecting a drone target based on a spatiotemporal attention mechanism and track matching according to an embodiment of the present invention;
[0044] Figure 2 A logical schematic diagram of a method for detecting a drone target based on a spatiotemporal attention mechanism and track matching according to an embodiment of the present invention;
[0045] Figure 3 A diagram showing a network structure for extracting features from visible light image data according to an embodiment of the present invention;
[0046] Figure 4 A diagram showing a positioning and classification network structure according to an embodiment of the present invention;
[0047] Figure 5 A schematic diagram of a spatiotemporal attention structure according to an embodiment of the present invention;
[0048] Figure 6 The figure is a schematic diagram of the structure of a UAV target detection device based on spatiotemporal attention mechanism and track matching according to an embodiment of the present invention. DETAILED DESCRIPTION
[0049] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined purpose, the present invention is described in detail below in conjunction with the accompanying drawings and preferred embodiments.
[0050] Unless otherwise defined, all terms (including technical terms and scientific terms) used in this article have the same meaning as commonly understood by ordinary technicians in the field to which this application belongs. It should also be understood that terms (such as terms defined in commonly used dictionaries) should be interpreted as having the same meaning as their meaning in the context of the relevant technology, and will not be interpreted in an idealized or overly formal sense unless explicitly defined in this article.
[0051] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.
[0052] The embodiment of the present invention is a method for detecting drone targets based on spatiotemporal attention mechanism and track matching. Figure 1 As shown, including:
[0053] Step S1, obtaining a known UAV visible light video sequence, and processing the video sequence into a single frame image;
[0054] Step S2, determining corresponding image features according to the image samples formed by the single frame image;
[0055] Step S3, based on the image features, using the configured algorithm model, determining the drone target detection result corresponding to the image sample;
[0056] Step S4, capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
[0057] Combination Figures 2 to 5 , the method provided in this embodiment will be described in detail step by step below.
[0058] The detailed steps of S1 are:
[0059] The drone is equipped with a high-resolution visible light camera that can capture clear video under different lighting conditions. During the drone flight mission, the camera continuously captures video sequences on the ground or in the air, which cover most of the monitoring area under the drone and record potential drone intrusion events. The captured video is extracted from the video sequence using a custom script, with 25 frames extracted per second.
[0060] The detailed steps of S2 are:
[0061] S2.1: CSPRepResStage is used as the core network framework, which is a deep learning network component that integrates the CSPNet architecture, repetitive feature extraction, and residual cascading. It aims to achieve efficient processing of feature maps and reduce computational complexity. By adopting a repetitive feature extraction mechanism, this technology significantly improves the accuracy and robustness of target detection. In addition, the addition of residual cascade (ResStage) further enhances the effective propagation and learning of deep features, ensuring the efficiency of information flow in the deep network structure.
[0062] S2.2: An innovative spatiotemporal attention mechanism is used as the main attention framework. This mechanism combines the attention of spatial and temporal dimensions, allowing the model to simultaneously analyze the temporal relationship between frames and the spatial details within the frames when processing video sequences. In this way, the system can dynamically adjust the attention to specific spatial regions and time points, effectively focusing on drone targets in continuous videos.
[0063] Spatial attention mechanism, taking the t-th track point as an example, the input feature vector x containing M features t =[x 1t , x 2t , …, x Mt ]. The attention weight vector e is obtained by mapping through a single-layer neural network t :
[0064] et =W e x t +b e
[0065] Where: e t =[e 1t , e 2t ,…,e mt ,…,e Mt ] is the attention weight coefficient corresponding to each input feature at the tth time step; W e is the trainable weight matrix; b e is the bias vector for calculating the spatial attention weight. Then, the softmax function in formula (10) is used to normalize each attention weight coefficient to obtain the attention weight α of each feature. t =[α 1t , α 2t , …, α mb , …, α Mt ], where α mt Represents the attention weight value of the mth feature:
[0066]
[0067] And there is Then use α t The input feature vector x t Recalculate the weight vector x′ t , which is the output of the attention mechanism layer at time t. The output of the attention mechanism layer is recorded as X′=[x′ 1 ,…,x′ t ,…,x′ L ] T ∈R L×M ,in
[0068] x′ t =α t ⊙x t =[α 1t x 1t α 2t x 2t … α Mt x Mt ]
[0069] Spatial attention enables the model to focus on image regions that are critical for drone detection, while temporal attention strengthens the model's recognition of key frames in a time series. The combination of this dual mechanism enables accurate identification and tracking of drones under diverse environmental conditions, such as changing lighting and dynamic backgrounds. The backbone network using this spatiotemporal attention structure improves the quality and relevance of features extracted from video sequences, thereby significantly enhancing the overall performance of the drone detection system and ensuring high accuracy and low false alarm rates in a variety of scenarios.
[0070] S2.3: The CSPPAN network is used as a path aggregation network component. The network generates three head (HEAD) matrices based on the feature matrix of the channel information and performs a random weighted sum operation on the three matrices.
[0071] S2.4: The CA attention mechanism is used as the attention module of the neck part. This mechanism is responsible for weighting the merged results. The merged data is used as the input feature map. This feature map is divided into two dimensions, height and width, and a global average pooling operation is performed on each dimension. The specific calculation formula is as follows:
[0072]
[0073] in, is the output of the cth channel in the height direction, is the output of the cth channel in the width direction, x c () is the input feature map of the cth channel, h is the feature map height, and w is the feature map width.
[0074] The feature maps of these two dimensions are then concatenated and merged, and a 1×1 convolution kernel is used to reduce the dimension to 1 / 16 of the original. The feature fusion conversion operation is performed to obtain the fused feature representation:
[0075] f=δ(F 1 ([z h , z w ]))
[0076] Where δ is the ReLU nonlinear activation function, F 1 is the sum operation of the spatial dimension, z h , z w is an intermediate feature map that encodes spatial information in the horizontal and vertical directions. Subsequently, the fused feature f is decomposed into two independent tensors along the spatial dimension, and these tensors are converted into output tensors with the same number of channels through two different convolution operations:
[0077] g h =σ(F h (f h ))
[0078] g W =σ(F w (f W ))
[0079] where g h , g w are the attention weights in height and width directions, σ is the Sigmoid activation function, and F h , F w are the segmentation functions of the tensor along the height and width directions, respectively, h , f w are the mapping of fusion features in height and width respectively. The final result y output by the CA attention mechanism is as follows:
[0080]
[0081] Where c is the cth channel. CA attention is used to improve the image's ability to understand width and height.
[0082] The detailed steps of S3 are:
[0083] The classification and positioning part uses the initial parameters of the model to analyze the input features to generate results, and then calculates the loss based on the actual image data to evaluate the difference between the model prediction and the actual result. By referring to the loss value, the model will gradually adjust its internal parameters, and through repeated iterations, the model output will gradually approach the true value. After multiple rounds of optimization, a set of parameters that best match the actual situation is finally formed to form the final model. Use this trained model to test images outside the training set to verify its target detection capabilities in different environments. If the effect is not ideal, the model needs to be further optimized to enhance its generalization and robustness.
[0084] The classification and positioning part is divided into two submodules: classification and positioning. The classification submodule uses the VariFocal Loss (VFL) cross entropy loss function to weight the loss value of positive samples, thereby giving more attention to high-quality samples to increase the influence of these samples in model training.
[0085] The positioning submodule is optimized using the Distribution Focal Loss (DFL) and Generalized Intersection over Union (GIoU) loss functions, where the GIoU loss is calculated as follows:
[0086]
[0087] Among them, IoU is the intersection-over-union ratio of the prediction box and the target box. A and B represent the area of the prediction box and the target box in the minimum convex hull, respectively, and C is the area of the minimum convex hull. The GIoU loss function obtains a more accurate IoU determination method by subtracting the ratio of the area not covered by A and B in C from IoU. Setting Represents the standardized target score, and the overall loss Loss is expressed as follows:
[0088]
[0089] Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively, N pos is the number of positive samples, 10os VFL is the VFL cross entropy loss. 1oss GIOU =1-GIOU,10Ss DFL is the DFL cross entropy loss function, 1oss GIOU is the GIoU loss function, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
[0090] The detailed steps of S4 are:
[0091] S4.1: The present invention uses visible light images to achieve track association to enhance the accuracy of drone target detection. When tracking a target, three frames of visible light images are captured from the start, middle, and end of the target's track. These three frames of images are fused on the RGB channel, and an image reflecting the characteristics of the target track is generated by calculating the average value of the corresponding pixel positions.
[0092] S4.2: The track association network of the present invention is mainly based on image processing, and includes a convolution module, which has a three-layer feature pyramid structure, each layer is followed by an activation function. The convolution kernel size is 3x3, the step size is 2, and the size of the feature map is gradually reduced. The feature maps after processing the three frames of image data are merged in channel order, and then the features are enhanced through the downsampling convolution layer and the channel attention mechanism, and finally a comprehensive feature map P is generated.
[0093] S4.3: The comprehensive feature map P is dimensionally adjusted and feature mapped through the fully connected layer. Finally, a correlation coefficient is calculated by comparing the feature differences of the feature maps. If the correlation coefficient exceeds a preset threshold, such as 0.9, it can be determined that the target and the known track belong to the same drone. This method improves the continuity and accuracy of drone target detection, and is particularly suitable for dynamic monitoring environments.
[0094] Compared with the prior art, this embodiment has at least the following effects:
[0095] 1) Compared with the traditional single image processing method, the visible light image processing strategy of the present invention can more effectively cope with the challenges of target detection in different scenarios and significantly improve the recognition efficiency and reliability of the results;
[0096] 2) The present invention can effectively determine whether it is multiple tracking of the same target by combining track correlation technology, thereby reducing misjudgment and repeated tracking and optimizing system performance. This not only reduces the workload for drone systems that perform complex monitoring and search and rescue tasks, but also improves the efficiency and accuracy of task execution.
[0097] The second embodiment of the present invention corresponds to the first embodiment, referring to Figure 6 This embodiment introduces a UAV target detection device based on spatiotemporal attention mechanism and track matching, including the following components:
[0098] An acquisition unit is configured to acquire a known UAV visible light video sequence and process the video sequence into a single frame image;
[0099] A feature unit configured to determine corresponding image features according to an image sample consisting of a single frame image;
[0100] A detection unit is configured to determine a drone target detection result corresponding to the image sample based on the image features and using a configured algorithm model;
[0101] The matching unit is configured to capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
[0102] In this embodiment, the feature unit is further configured to: process the image sample using the spatiotemporal attention mechanism, specifically including:
[0103] Acquire track points represented by multiple single-frame images, which include input feature vectors of corresponding multiple features, and map the input feature vectors through a single-layer neural network to obtain an attention weight vector;
[0104] Normalize each attention weight coefficient to obtain the attention weight of each feature;
[0105] Using the attention weights, the input feature vector is re-determined as a weighted vector, that is, the output of the spatiotemporal attention mechanism at the current moment.
[0106] In this embodiment, the detection unit is further configured to optimize the algorithm model, and the optimization process is specifically implemented by the following formula:
[0107]
[0108] Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively; N pos is the number of positive samples, loss VFL is the VFL cross entropy loss, 1oss GIOU =1-GIOU,loss DFL is the DFL cross entropy loss function, 1oss GIOU is the loss function of GIoU, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
[0109] In this embodiment, the matching unit is further configured as follows:
[0110] Three frames of visible light images are captured from the beginning, middle and end of the target's track respectively;
[0111] Perform convolution processing on the visible light image, merge the feature maps of the three frames of image data after processing in channel order, enhance the features through the downsampling convolution layer and channel attention mechanism, and generate a feature map;
[0112] The feature map is dimensionally adjusted and feature mapped through a fully connected layer, and a correlation coefficient is determined by comparing the feature difference between the feature map and the drone target detection result;
[0113] If the correlation coefficient exceeds the preset threshold, it is determined that the target and the known track belong to the same UAV.
[0114] A third embodiment of the present invention is an electronic device, which can be understood as a physical device, including a processor and a memory storing instructions executable by the processor. When the instructions are executed by the processor, the following operations are performed:
[0115] Step S1, obtaining a known UAV visible light video sequence and processing the video sequence into a single frame image;
[0116] Step S2, determining corresponding image features according to image samples consisting of a single frame image;
[0117] Step S3, based on the image features, using the configured algorithm model, determining the drone target detection result corresponding to the image sample;
[0118] Step S4, capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
[0119] The fourth embodiment of the present invention, the process of the drone target detection method based on the spatiotemporal attention mechanism and track matching in this embodiment is the same as that of the first, second or third embodiment, except that, in terms of engineering implementation, this embodiment can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the method of the present invention can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, disk, CD), including a number of instructions for a device to execute the method described in the embodiment of the present invention.
[0120] Through the description of the specific implementation methods, a deeper and more specific understanding of the technical means and effects adopted by the present invention to achieve the predetermined purpose should be obtained. However, the accompanying drawings are only for reference and illustration purposes and are not intended to limit the present invention.
Claims
1. A UAV target detection method based on spatiotemporal attention mechanism and track matching, characterized in that: include: Step S1, obtaining a known UAV visible light video sequence, and processing the video sequence into a single frame image; Step S2, determining corresponding image features according to the image samples formed by the single frame image; Step S3, based on the image features, using the configured algorithm model, determining the drone target detection result corresponding to the image sample; Step S4, capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
2. The method for detecting unmanned aerial vehicle targets based on spatiotemporal attention mechanism and track matching according to claim 1, characterized in that: The step S2 includes: processing the image sample using a spatiotemporal attention mechanism, specifically including: Acquire a plurality of track points represented by the single-frame images, which include input feature vectors of corresponding multiple features, and obtain an attention weight vector by mapping the input feature vectors through a single-layer neural network; Normalize each attention weight coefficient to obtain the attention weight of each feature; Using the attention weights, the input feature vector is re-determined as a weighted vector, that is, the output of the spatiotemporal attention mechanism at the current moment.
3. The method for detecting unmanned aerial vehicle targets based on spatiotemporal attention mechanism and track matching according to claim 2 is characterized in that: The step S3 also includes an optimization process for the algorithm model, and the optimization process is specifically implemented by the following formula: Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively; N pos is the number of positive samples, loss VFL is the VFL cross entropy loss, 1oss GIOU =1-GIOU,loss DFL is the DFL cross entropy loss function, 1oss GIOU is the loss function of GIoU, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
4. The method for detecting unmanned aerial vehicle targets based on spatiotemporal attention mechanism and track matching according to claim 3 is characterized in that: The step S4 comprises: Three frames of visible light images are captured from the beginning, middle and end of the target's track respectively; Perform convolution processing on the visible light image, merge the feature maps of the three frames of image data after processing in channel order, enhance the features through the downsampling convolution layer and channel attention mechanism, and generate a feature map; The feature map is dimensionally adjusted and feature mapped through a fully connected layer, and a correlation coefficient is determined by comparing the feature difference between the feature map and the drone target detection result; If the correlation coefficient exceeds the preset threshold, it is determined that the target and the known track belong to the same UAV.
5. A UAV target detection device based on spatiotemporal attention mechanism and track matching, characterized in that: include: An acquisition unit is configured to acquire a known UAV visible light video sequence and process the video sequence into a single frame image; A feature unit, configured to determine corresponding image features according to the image sample formed by the single frame image; A detection unit configured to determine a drone target detection result corresponding to the image sample based on the image feature and using a configured algorithm model; The matching unit is configured to capture three consecutive frames of visible light images in the video of the target track, determine the corresponding feature map, and determine the target correlation with the drone target detection result. If it is greater than a preset value, the target and the known track belong to the same drone.
6. The unmanned aerial vehicle target detection device based on spatiotemporal attention mechanism and track matching according to claim 5, characterized in that: The feature unit is further configured to process the image sample using a spatiotemporal attention mechanism, specifically including: Acquire a plurality of track points represented by the single-frame images, which include input feature vectors of corresponding multiple features, and obtain an attention weight vector by mapping the input feature vectors through a single-layer neural network; Normalize each attention weight coefficient to obtain the attention weight of each feature; Using the attention weights, the input feature vector is re-determined as a weighted vector, that is, the output of the spatiotemporal attention mechanism at the current moment.
7. The unmanned aerial vehicle target detection device based on spatiotemporal attention mechanism and track matching according to claim 6, characterized in that: The detection unit is further configured to optimize the algorithm model, and the optimization process is specifically implemented by the following formula: Among them, α, β and γ are the weights of VFL cross entropy loss function, DFL loss function and GIoU loss function respectively; N pos is the number of positive samples, loss VFL is the VFL cross entropy loss, 1oss GIOU =1-GIOU,loss DFL is the DFL cross entropy loss function, 1oss GIOU is the loss function of GIoU, 1oss GIoU and 1oss DFL Together they form the cross entropy loss for pin frame positioning.
8. The unmanned aerial vehicle target detection device based on spatiotemporal attention mechanism and track matching according to claim 7, characterized in that: The matching unit is further configured as: Three frames of visible light images are captured from the beginning, middle and end of the target's track respectively; Perform convolution processing on the visible light image, merge the feature maps of the three frames of image data after processing in channel order, enhance the features through the downsampling convolution layer and channel attention mechanism, and generate a feature map; The feature map is dimensionally adjusted and feature mapped through a fully connected layer, and a correlation coefficient is determined by comparing the feature difference between the feature map and the drone target detection result; If the correlation coefficient exceeds the preset threshold, it is determined that the target and the known track belong to the same UAV.
9. An electronic device, characterized in that: The electronic device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is executed by the processor, the steps of the drone target detection method based on spatiotemporal attention mechanism and track matching as described in any one of claims 1 to 4 are implemented.
10. A computer storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the drone target detection method based on spatiotemporal attention mechanism and track matching as described in any one of claims 1 to 4 are implemented.
Citation Information
Cited By
Ship type identification method and device based on motion features of track point set
CN121479529A
