Unmanned aerial vehicle target identification method and device, computer equipment and storage medium

Abstract

The invention belongs to the technical field of unmanned aerial vehicles, and relates to an unmanned aerial vehicle target identification method and device, computer equipment and a storage medium, and the method comprises the steps: collecting image data through an unmanned aerial vehicle, and carrying out the preprocessing of the image data; extracting feature representation with distinction degree from the preprocessed image data through a deep learning model; according to the feature representation, positioning a target position in the image data, and carrying out category division on a target; carrying out parallel processing on each category group, and respectively carrying out target detection and tracking; and associating detection results at different time points according to target detection and tracking results to form a motion track of the target. Through constructing a complex neural network model, image features are automatically learned, efficient target detection, classification, tracking and other functions are realized, and stable and accurate target tracking can be realized. The method can be widely applied to the fields of video monitoring, automatic driving, man-machine interaction and the like.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a UAV target identification method, apparatus, computer equipment, and storage medium. Background Technology

[0002] Unmanned aerial vehicle (UAV) target recognition, as an important research direction in modern technology, refers to enabling UAVs to identify and classify specific targets through machine vision technology.

[0003] Despite the numerous advantages of drone target recognition technology, some shortcomings remain in practical applications. Due to limitations in the performance of the sensors or equipment carried by the drone, or unreasonable flight path planning, certain areas may fail to be effectively detected. Therefore, missed detections are a common problem. Summary of the Invention

[0004] The purpose of this invention is to provide a method, apparatus, computer device, and storage medium for unmanned aerial vehicle (UAV) target recognition, in order to solve the problem of missed detection in UAV target recognition in the prior art.

[0005] To address the aforementioned technical problems, this invention provides a method for unmanned aerial vehicle (UAV) target recognition, employing the following technical solution:

[0006] The drone collects image data and preprocesses the image data.

[0007] Using deep learning models, discriminative feature representations are extracted from preprocessed image data.

[0008] Based on the feature representation, the target location is located in the image data, and the target is classified.

[0009] Each category group is processed in parallel, performing target detection and tracking separately;

[0010] Based on the results of target detection and tracking, the detection results at different time points are correlated to form the target's motion trajectory.

[0011] Preferably, the step of acquiring image data by the UAV and preprocessing the image data specifically includes:

[0012] Equipped with high-definition cameras or sensors, it captures images and videos along a pre-set flight path;

[0013] Filtering algorithms are used to remove noise from the acquired images and videos;

[0014] Perform contrast adjustment, sharpening, and color correction on image and video data processed by filtering algorithms;

[0015] Annotate the image and video data that have undergone contrast adjustment, sharpening, and color correction to clarify the location, category, and other information of the target.

[0016] Preferably, the step of extracting discriminative feature representations from preprocessed image data using a deep learning model specifically includes:

[0017] Select a deep learning model based on task requirements;

[0018] Multi-layer convolution operations are used to extract low-level to high-level feature representations from the original image;

[0019] Pooling operations reduce the dimensionality of feature maps, thereby reducing computational cost while preserving feature information.

[0020] Adding a non-linear activation function between the convolutional and pooling layers increases the model's non-linear expressive power and improves feature extraction performance.

[0021] Preferably, the step of locating the target position in the image data based on the feature representation and classifying the target specifically includes:

[0022] Candidate regions are generated in the image using a sliding window or region proposal network method as potential target locations;

[0023] Feature extraction is performed on each candidate region to obtain its corresponding feature representation;

[0024] The extracted features are input into the classifier and regressor for classification and location regression.

[0025] For multiple detection boxes of the same category, the NMS algorithm is used to remove redundant boxes, and the detection box with the highest score is retained as the final result.

[0026] Preferably, the step of performing parallel processing on each category group, and separately performing target detection and tracking, specifically includes:

[0027] Based on the target category, the video stream or image data is divided into multiple independent processing units, each of which is responsible for processing the video stream or image data of a certain type of target.

[0028] Data synchronization between processing units can be achieved through shared memory or message queues.

[0029] Resource allocation is dynamically adjusted based on the workload and computational complexity of each processing unit;

[0030] Target detection and tracking are performed based on the target's motion characteristics.

[0031] Preferably, the step of correlating the detection results at different time points to form the target's motion trajectory based on the target detection and tracking results specifically includes:

[0032] A data association method is used to match detection results at different time points to determine the detection results belonging to the same target;

[0033] Based on data association, the target's state information is updated through state estimation;

[0034] Based on the updated target status information, the target is continuously tracked and its movement trajectory is constantly updated.

[0035] Preferably, after the step of correlating the detection results at different time points based on the target detection and tracking results to form the target's motion trajectory, the method further includes:

[0036] When multiple targets are detected to intersect or occlude, the accurate position and trajectory of each target are determined based on the target's motion characteristics and shape features.

[0037] To address the aforementioned technical problems, the present invention also provides a drone target identification device, which employs the following technical solution, including:

[0038] The acquisition and preprocessing module is used to acquire image data by the UAV and to preprocess the image data.

[0039] The feature extraction module is used to extract discriminative feature representations from preprocessed image data using a deep learning model;

[0040] The category segmentation module is used to locate the target position in the image data based on the feature representation and to classify the target into categories.

[0041] The detection and tracking module is used to process each category group in parallel, performing target detection and tracking separately;

[0042] The association module is used to associate the detection results at different time points based on the target detection and tracking results to form the target's motion trajectory.

[0043] To address the aforementioned technical problems, the present invention also provides a computer device that employs the technical solution described below, comprising a memory and a processor. The memory stores computer-readable instructions, and the processor executes the computer-readable instructions to implement the steps of the aforementioned UAV target recognition method.

[0044] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium, which employs the technical solution described below. The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of the aforementioned UAV target recognition method.

[0045] Compared with existing technologies, the present invention has the following advantages: it can collect a large amount of image and video data in the air and can quickly and accurately identify targets from massive amounts of data; it adopts an artificial intelligence-based target recognition method, especially deep learning technology, and automatically learns image features by constructing a complex neural network model to achieve efficient target detection, classification, tracking and other functions, and can achieve stable and accurate target tracking; it can be widely applied to fields such as video surveillance, autonomous driving, and human-computer interaction. Attached Figure Description

[0046] To more clearly illustrate the solutions in this invention, the accompanying drawings used in the description of the embodiments of this invention will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0047] Figure 1 This is an exemplary system architecture diagram in which the present invention can be applied;

[0048] Figure 2 This is a flowchart of an embodiment of a UAV target recognition method according to the present invention;

[0049] Figure 3 This is a schematic diagram of one embodiment of a drone target recognition device according to the present invention;

[0050] Figure 4 This is a schematic diagram of the structure of an embodiment of the computer device of the present invention. Detailed Implementation

[0051] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains; the terminology used herein in the specification is for the purpose of describing particular embodiments only and is not intended to limit the invention; the terms "comprising" and "having," and any variations thereof, in the specification, claims, and foregoing drawings are intended to cover non-exclusive inclusion. The terms "first," "second," etc., in the specification, claims, or foregoing drawings are used to distinguish different objects and not to describe a particular order.

[0052] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0054] like Figure 1 As shown, the system architecture 100 may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing communication links between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links, or fiber optic cables, etc.

[0055] Users can use the first terminal device 101, the second terminal device 102, or the third terminal device 103 to interact with the server 105 via the network 104 to receive or send messages, etc. Various communication client applications can be installed on the first terminal device 101, the second terminal device 102, and the third terminal device 103, such as web browser applications, shopping applications, search applications, instant messaging tools, email clients, social media platform software, etc.

[0056] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be various electronic devices with displays and support web browsing, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), MP4 players (Moving Picture Experts Group Audio Layer IV), laptops, and desktop computers, etc.

[0057] Server 105 can be a server that provides various services, such as a backend server that supports the pages displayed on the first terminal device 101, the second terminal device 102, and the third terminal device 103.

[0058] It should be noted that the UAV target recognition method provided in this embodiment of the invention is generally executed by a server / terminal device, and correspondingly, a UAV target recognition device is generally installed in the server / terminal device.

[0059] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0060] Example 1

[0061] Continue to refer to Figure 2 The diagram illustrates a flowchart of an embodiment of a drone target recognition method according to the present invention. The drone target recognition method includes the following steps:

[0062] Step S1: The drone collects image data and preprocesses the image data.

[0063] In this embodiment, a drone target recognition method operates on an electronic device (e.g., Figure 1 The server / terminal device shown can receive a drone target identification request via wired or wireless connection. It should be noted that the aforementioned wireless connection methods may include, but are not limited to, 3G / 4G / 5G connections, Wi-Fi connections, Bluetooth connections, Wi-Fi Maxx connections, Zigbee connections, UWB (ultra-wireless) connections, and other currently known or future wireless connection methods.

[0064] In this embodiment, step S1, which involves the drone collecting image data and preprocessing the image data, may further include the following steps:

[0065] The S11 is equipped with a high-definition camera or sensor to capture images and videos along a preset flight path.

[0066] Depending on mission requirements, different flight altitudes, speeds, and angles can be set to obtain diverse data. High-definition cameras or sensors of different specifications can be selected based on actual needs.

[0067] S12 uses a filtering algorithm to remove noise from the acquired images and videos.

[0068] Filtering algorithms reduce or eliminate noise interference by performing local or global numerical operations on pixels in an image or video frame. For example, Gaussian filtering can be used, which smooths the image based on a Gaussian function as the weighting criterion. Gaussian filters assign different weights to neighboring pixels based on their distance from the center pixel; pixels closer to the center pixel have a greater impact, thus preserving image details while removing noise.

[0069] In addition, bilateral filtering can be used. As a nonlinear filter, it combines the spatial proximity of the image with the gray-level similarity between pixels for filtering. Bilateral filters consider not only the distance between pixels but also their gray-level differences, thus effectively avoiding edge blurring while removing noise, making them particularly suitable for scenarios where image edge information needs to be preserved.

[0070] Step S12 can remove noise from the image and improve image quality.

[0071] S13 performs contrast adjustment, sharpening, and color correction on the image and video data after filtering algorithm processing.

[0072] Contrast adjustment is achieved by modifying the grayscale values ​​at various locations in the image, which can enhance the contrast between light and dark areas, making the image more vivid. Sharpening focuses on image edges and details, and can be achieved using high-pass filtering or unsharp masking, by enhancing the high-frequency components of the image to make edges and textures clearer.

[0073] Color correction involves steps such as white balance adjustment, color correction, and color enhancement, aiming to eliminate the influence of light sources and restore the true colors of an image. In practice, deep learning algorithms can also be used to learn the color change patterns under different lighting conditions, thereby adjusting image colors more accurately.

[0074] Step S13 can enhance useful information in the image and improve the target recognition rate.

[0075] S14 annotates the image and video data that has undergone contrast adjustment, sharpening, and color correction, clarifying information such as the location and category of the target.

[0076] Image details are enhanced through adaptive contrast stretching, edge sharpening is achieved using a high-pass filter to improve image clarity, and color correction algorithms are applied to restore or enhance color fidelity and saturation. Subsequently, during annotation, a combination of high-precision manual annotation and automated auxiliary tools is used to clearly define the location boundaries of target objects, employing polygonal bounding boxes or pixel-level segmentation techniques for precise depiction. Simultaneously, accurate category labels, such as "pedestrian," "vehicle," and "building," are assigned to each target according to a predefined classification system, ensuring the accuracy of category information. This process can be facilitated by annotation software platforms to achieve efficient management and quality control of annotation data, laying a solid foundation for subsequent model training and optimization.

[0077] Step S14 provides labeled data for model training. Labeling can be done manually or automatically, the latter relying on a pre-trained model.

[0078] Step S2: Extract discriminative feature representations from the preprocessed image data using a deep learning model.

[0079] In this embodiment, step S2, extracting discriminative feature representations from the preprocessed image data using a deep learning model, may further include the following steps:

[0080] S21. Select a deep learning model based on task requirements.

[0081] The specific nature of the task, including image recognition, needs to be analyzed. The characteristics of the data, such as data size, feature dimensionality, data distribution, and the presence of imbalances, must also be considered. Based on these analyses, the applicability of the deep learning model can be evaluated from multiple dimensions. For image recognition tasks, Convolutional Neural Networks (CNNs) can be chosen. The advantage of CNNs lies in their ability to efficiently extract spatial hierarchical features from images. For example, a suitable deep learning model, such as ResNet, VGG, or Mobil eNet, can be selected based on the task requirements.

[0082] S22 uses multi-layer convolution operations to extract low-level to high-level feature representations from the original image.

[0083] In Convolutional Neural Networks (CNNs), multi-layer convolutional operations are crucial for extracting low- to high-level feature representations from raw images. By progressively extracting and combining abstract features of the image layer by layer, the network gradually achieves this. The first convolutional layer focuses on extracting low-level features such as edges, colors, and textures. As the number of network layers increases, intermediate convolutional layers further combine these low-level features to extract more complex image structures, such as object parts or patterns. In higher-level convolutions, the network can capture high-level semantic information, such as object categories, scene representations, or representations of the entire object.

[0084] Each convolutional layer extracts local features by performing a sliding window operation on the input feature map using a convolutional kernel (also called a filter). As the network depth increases, the resolution of the feature map gradually decreases, but the semantic information it contains becomes increasingly richer. This feature extraction process from low to high layers enables CNNs to learn data representations more effectively and perform complex image recognition and processing tasks. Convolutional layers effectively extract local features from images through sliding windows and weight sharing mechanisms.

[0085] S23 reduces the dimensionality of the feature map through pooling operations, thereby reducing computational cost while preserving feature information.

[0086] Pooling is a crucial step in convolutional neural networks (CNNs). It reduces computation by lowering the dimensionality of the feature map while effectively preserving key feature information. This operation is typically implemented by sliding a fixed-size window (such as 2x2 or 3x3) across the input feature map, where the maximum value (max pooling) or the average value (average pooling) can be selected as the output.

[0087] Max pooling captures salient features in the feature map, such as edges and corners, which typically have high response values, thus helping to preserve important image information. Average pooling, on the other hand, smooths the output by averaging all values ​​within a window, helping to maintain the consistency of background features.

[0088] Pooling significantly reduces the spatial size of feature maps, which not only reduces the number of parameters and computational complexity of subsequent layers but also improves the model's robustness to positional changes (i.e., position invariance). Even if objects in the image undergo slight displacement, the pooled features remain stable, helping the model to maintain sensitivity to these small changes.

[0089] S24 introduces a non-linear activation function between the convolutional and pooling layers to enhance the model's non-linear expressive power and improve feature extraction performance.

[0090] In Convolutional Neural Networks (CNNs), adding a non-linear activation function between convolutional and pooling layers is a crucial step in enhancing the model's non-linear expressive power and feature extraction performance. This operation, by introducing a non-linear element, allows the model to learn and approximate more complex function mappings. Specifically, convolutional layers are responsible for extracting local features of an image, such as edges and textures; however, their output is typically linear or linearly transformed. To enhance the model's expressive power, a non-linear activation function, such as ReLU (Rectified Linear Unit), follows the convolutional layer.

[0091] The ReLU function introduces non-linearity into the model by setting negative values ​​to zero while leaving positive values ​​unchanged. This non-linear activation allows the network to learn complex non-linear combinations of input data, thereby improving the richness and accuracy of feature extraction. Adding an activation function before the pooling layer ensures that the data passed to the pooling layer has undergone a non-linear transformation, thus retaining more representative features and reducing data redundancy. In this way, the model can more effectively extract key features from the input data, laying a solid foundation for subsequent tasks such as classification and detection.

[0092] Step S3: Based on the feature representation, locate the target position in the image data and classify the target into categories.

[0093] In this embodiment, step S3, locating the target position in the image data based on feature representation and classifying the target, may further include the following steps:

[0094] S31. A sliding window or region proposal network method is used to generate candidate regions in the image as potential target locations.

[0095] Generating candidate regions in an image as potential target locations can be achieved using sliding window or Region Proposal Network (RPN) methods. The sliding window method extracts the image region within each window as a candidate region by sliding a fixed-size window across the image at certain steps and proportions.

[0096] Region Proposal Networks (RPNs) offer advantages in efficiency and flexibility. RPNs generate a series of anchor boxes on the feature map, covering regions of varying locations and sizes, selected based on the average size and distribution of the target. Subsequently, RPNs utilize convolutional neural networks to classify and regress these anchor boxes, determining whether they contain the target and adjusting their positions and sizes for more accurate target matching. In this process, RPNs can generate candidate regions of multiple scales and shapes, effectively adapting to target objects of different sizes and shapes.

[0097] S32, extract features from each candidate region to obtain its corresponding feature representation.

[0098] Preprocessing steps, such as image cropping, scaling, and normalization, ensure the consistency of the input data. Then, a convolutional neural network (CNN) is used to automatically learn hierarchical features of candidate regions. Through sliding operations of multiple convolutional kernels, CNNs can capture local features ranging from edges and textures to complex shapes, and gradually abstract these features into global semantic features as the network depth increases.

[0099] S33, the extracted features are input into the classifier and regressor for classification and location regression.

[0100] The raw data is processed using feature engineering methods (such as standardization, normalization, and PCA dimensionality reduction) to eliminate noise and improve model efficiency. Then, for classification tasks, the processed feature vectors are used as input to classifiers such as logistic regression, decision trees, random forests, or deep learning models (such as CNNs and RNNs). These classifiers learn the mapping relationship between features and target categories, enabling them to accurately classify new samples into predefined categories.

[0101] For location regression tasks, features need to be input into a regression-enabled model, such as linear regression, ridge regression, Lasso regression, or gradient boosting trees (GBDT). These models aim to learn the functional relationship between feature vectors and continuous target values ​​(such as location coordinates), optimizing model parameters by minimizing the error between predicted and actual values. Ultimately, the trained regression model can predict the specific location of the target object based on the input features, achieving accurate location regression. This entire process combines the advantages of feature extraction and model training, ensuring efficiency and accuracy in both classification and regression tasks.

[0102] S34. For multiple detection boxes of the same category, the NMS algorithm is used to remove redundant boxes and retain the detection box with the highest score as the final result.

[0103] The detection boxes are sorted in descending order based on their confidence score (score), and the boxes with the highest confidence scores are selected as the retained boxes. For the sorted list of detection boxes, the overlap between each box and other boxes is checked one by one, which can be achieved by calculating the Intersection over Union (IoU). If the IoU between a box and a selected detection box exceeds a preset threshold (e.g., 0.5), the box is considered redundant and is removed. This process is repeated iteratively until all candidate detection boxes have been processed.

[0104] The core of the NMS algorithm lies in determining the overlap relationship between bounding boxes through IoU calculation, thereby eliminating suboptimal bounding boxes and ensuring that each target is represented by only one detection box with the highest score. This strategy not only reduces false positives and duplicate detections but also significantly improves the accuracy and efficiency of target detection.

[0105] Step S4: Perform parallel processing on each category group, and perform target detection and tracking respectively.

[0106] In some optional implementations of this embodiment, step S4, which involves parallel processing of each category group to perform target detection and tracking respectively, may further include the following steps:

[0107] S41, based on the target category, divide the video stream or image data into multiple independent processing units, each unit being responsible for processing the video stream or image data of a certain type of target.

[0108] Object detection algorithms (such as YOLO and SSD) are used to locate and classify objects in video streams or images. Localization and classification determine which objects belong to which category, and segmentation is performed accordingly. Object detection algorithms not only output the object's category information but also its location (such as a bounding box) within the image or video frame.

[0109] Based on the object detection results, video streams or images can be segmented into multiple independent processing units. Each unit contains a set of objects belonging to the same category and their surrounding contextual information. This segmentation ensures that each processing unit focuses on processing a specific class of objects, improving processing efficiency and accuracy.

[0110] Each processing unit will independently perform further analysis or processing, such as target tracking, behavior analysis, or feature extraction, to meet specific application requirements. Throughout the process, data flow and processing are automated, and processing speed and efficiency can be further improved through algorithm optimization and hardware acceleration.

[0111] S42 synchronizes data between processing units through shared memory or message queues.

[0112] Shared memory is an efficient inter-process communication method that allows different processes to directly access the same memory region without data copying, thus improving data transfer efficiency. To achieve data synchronization, synchronization mechanisms such as mutexes or semaphores are typically used. For example, when a process needs to write data, it first attempts to acquire the mutex. If the lock is held by another process, the current process will wait until the lock is released. This ensures that only one process can access the shared memory region at a time, thereby avoiding data conflicts and inconsistencies.

[0113] Message queues are an asynchronous communication mechanism used to transfer data between different systems or components. In data synchronization scenarios, system A can encapsulate the data to be synchronized into messages and place them in a queue. System B then consumes these messages from the queue and processes them. After processing, system B can either put the result message back into the queue or directly feed it back to system A. In this way, loosely coupled data exchange is achieved between the various processing units, improving the system's scalability and fault tolerance. Furthermore, message queues support persistence and message acknowledgment mechanisms, ensuring data reliability and consistency.

[0114] Data synchronization is crucial in distributed systems or concurrent environments to ensure data consistency and accuracy among processing units. Shared memory or message queues can be used to effectively achieve data synchronization.

[0115] S43 dynamically adjusts resource allocation based on the workload and computational complexity of each processing unit.

[0116] The system needs to monitor the task queues and current load of each processing unit in real time, including key indicators such as the number and type of tasks and the execution time required. Through data collection and analysis, the system can gain a comprehensive understanding of the real-time status of each unit.

[0117] Scheduling algorithms, such as priority scheduling or Completely Fair Scheduler (CFS), are used to dynamically adjust resource allocation based on the importance and urgency of tasks. Priority scheduling allocates resources according to task priority, ensuring that critical tasks are processed first; while CFS dynamically adjusts the virtual runtime of processes to ensure that each process receives a roughly equal share of CPU time over a long period of time, thus achieving both fairness and efficiency.

[0118] In multi-core or multi-processor systems, load balancing algorithms are further applied to intelligently migrate tasks to relatively idle CPU cores for execution, thereby reducing waiting time and improving overall processing power. This process not only relies on accurate assessment of task load and computational complexity, but also requires efficient communication and coordination mechanisms to ensure smooth task migration and seamless integration.

[0119] S44 performs target detection and tracking based on the target's motion characteristics.

[0120] In the object detection stage, deep learning techniques, especially convolutional neural networks (CNNs), are used to extract features from the input image or video frames. These features include color, texture, and shape, which can effectively distinguish target objects from the background. By training on a large amount of labeled data, CNNs can learn the feature representations of target objects and achieve accurate identification and localization of them. During this process, algorithms such as non-maximum suppression (NMS) are used to remove detected overlapping regions, improving detection accuracy.

[0121] During the target tracking phase, appropriate tracking algorithms are employed for real-time tracking based on the target's motion characteristics, such as velocity, acceleration, and direction. Examples of such algorithms include Kalman filtering and particle filtering. These algorithms can predict the target's position in future frames and perform a search and matching process based on the prediction results in the new frame, achieving continuous target tracking. Simultaneously, techniques such as optical flow and feature point matching are also widely used in target tracking to address complex situations such as target deformation and occlusion.

[0122] Step S5: Based on the target detection and tracking results, the detection results at different time points are correlated to form the target's motion trajectory.

[0123] In some optional implementations of this embodiment, step S5, which involves associating the detection results at different time points to form the target's motion trajectory based on the target detection and tracking results, may further include the following steps:

[0124] S51 uses a data association method to match detection results at different time points to determine the detection results belonging to the same target.

[0125] Each detection result is characterized by a well-defined set of features, which may include location coordinates, velocity, acceleration, appearance features (such as color and texture), etc. Similarity metrics (such as Euclidean distance, cosine similarity, or feature-based matching algorithms) are used to calculate the similarity between detection results at different time points.

[0126] Data association algorithms, such as the Hungarian algorithm (KM algorithm), can be used to solve the assignment problem, or more complex probabilistic data association filters (such as JPDAF, MHT, etc.) can be used. These methods can handle detection uncertainty and occlusion problems between targets. By calculating the association probability between each detection result and the previous target trajectory, the optimal matching can be achieved.

[0127] Introducing time windows and state prediction mechanisms is also an important means to improve matching accuracy. By predicting the possible state of the target at future points in time through historical data and combining it with real-time detection results, the matching search range can be narrowed, further improving the efficiency and accuracy of association.

[0128] S52 updates the target's state information through state estimation based on data association.

[0129] Data association involves matching sensor-acquired measurement data with existing target track information in a database to determine which measurement data originate from the same target. This process is typically achieved by establishing association thresholds, threshold filtering, determining similarity metrics, building association matrices, and setting decision criteria, such as using nearest neighbor association algorithms or global nearest neighbor association algorithms.

[0130] State estimation relies on filtering algorithms, such as Kalman filtering and its extensions, which can estimate the current state of a target based on its dynamic model and observation data. Kalman filtering works through two steps: prediction and update. First, it predicts the current state based on the model's state at the previous time step; then, it uses the observation data at the current time step to refine this prediction, obtaining the optimal estimate for the current time step.

[0131] During the update process, Kalman filtering considers observation noise and model prediction uncertainties, balancing their effects by calculating the Kalman gain to obtain the optimal state estimate. This estimate not only reflects the target's position and velocity at the current moment but also predicts its future trajectory, providing support for subsequent tracking and decision-making.

[0132] S53 continuously tracks the target and updates its trajectory based on the updated target status information.

[0133] Determining the target's location and features during the initialization phase is fundamental for subsequent tracking. In the initial frames of the video, the target is identified using object detection technology or manual calibration, and its features, such as color, texture, and edge information, are extracted.

[0134] In the detection phase, image processing or deep learning-based target detection models (such as convolutional neural networks, CNNs) are used to locate the target in subsequent frames. Once the target is detected, the tracking phase begins, using motion and appearance models to predict the target's position in the current frame. For example, Kalman filtering or particle filtering methods can be used, which estimate the target's new state by combining information from previous frames.

[0135] During tracking, the system needs to continuously update the target model to adapt to changes in the target's shape and appearance, ensuring tracking accuracy. For the target's trajectory, the system continuously records and updates the target's position information in each frame. This position information is typically represented in the form of bounding boxes or pixel masks and stored in a data structure for subsequent analysis and processing.

[0136] Furthermore, when the target is occluded or interacts with other targets, MeanShift, CamShift algorithms, or deep learning-based models can be used to maintain tracking continuity. These algorithms can effectively handle occlusion and continue to accurately track the target.

[0137] In some alternative implementations, after step S5 above, the electronic device may perform the following steps:

[0138] S6, when multiple targets are detected to be intersecting or occluding, the accurate position and trajectory of each target are determined based on the target's motion characteristics and shape features.

[0139] Based on the target's motion characteristics, a prediction and update mechanism can be employed to address occlusion. Specifically, motion information of the target in historical frames, such as velocity and acceleration, is used to predict motion through a Kalman filter or a more advanced deep learning model. This prediction mechanism can estimate the possible position and velocity of the target when it is partially or completely occluded, thus maintaining the continuity of tracking. When the occlusion is removed, the algorithm can quickly adjust and re-lock onto the target, ensuring the accuracy and stability of tracking.

[0140] Shape features also play a crucial role in resolving the problem of intersection occlusion. During target tracking, the algorithm continuously extracts and updates the shape features of the target, such as edges, textures, and corners. When multiple targets intersect, this feature information helps distinguish between different targets and reduces the possibility of false matches. Especially for targets with significant shape differences, precise matching of shape features can lead to more accurate identification and tracking.

[0141] To further improve the robustness of tracking, multiple feature information can be combined for joint decision-making. For example, when the target is occluded, more reliance can be placed on motion characteristics for prediction; while when the target is clearly visible, more emphasis can be placed on shape features for accurate matching. This multi-feature fusion strategy can fully utilize the advantages of different features and improve the overall performance of the tracking algorithm.

[0142] The beneficial effects of implementing this embodiment are: it can collect a large amount of image and video data in the air and can quickly and accurately identify targets from massive amounts of data; it adopts an artificial intelligence-based target recognition method, especially deep learning technology, and automatically learns image features by constructing a complex neural network model to achieve efficient target detection, classification, tracking and other functions, and can achieve stable and accurate target tracking; it can be widely applied to fields such as video surveillance, autonomous driving, and human-computer interaction.

[0143] This invention can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0144] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by instructing related hardware through computer-readable instructions. These computer-readable instructions can be stored in a computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the methods described above. The aforementioned storage medium can be a non-volatile storage medium such as a magnetic disk, optical disk, or read-only memory (ROM), or random access memory (RAM).

[0145] It should be understood that although the steps in the flowcharts of the accompanying figures are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the accompanying figures may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times, and their execution order is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0146] Example 2

[0147] Further reference Figure 3 As a response to the above Figure 2 The present invention provides an embodiment of a drone target recognition device based on the method shown, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0148] like Figure 3 As shown, the UAV target recognition device 60 described in this embodiment includes: a data acquisition and preprocessing module 61, a feature extraction module 62, a category classification module 63, a detection and tracking module 64, and an association module 65. Wherein:

[0149] The acquisition and preprocessing module 61 is used to acquire image data by the UAV and preprocess the image data.

[0150] The feature extraction module 62 is used to extract discriminative feature representations from preprocessed image data through a deep learning model;

[0151] The category segmentation module 63 is used to locate the target position in the image data according to the feature representation and to classify the target into categories.

[0152] The detection and tracking module 64 is used to perform parallel processing on each category group, performing target detection and tracking respectively;

[0153] The association module 65 is used to associate the detection results at different time points based on the results of target detection and tracking to form the target's motion trajectory.

[0154] The beneficial effects of implementing this embodiment are: it can collect a large amount of image and video data in the air and can quickly and accurately identify targets from massive amounts of data; it adopts an artificial intelligence-based target recognition method, especially deep learning technology, and automatically learns image features by constructing a complex neural network model to achieve efficient target detection, classification, tracking and other functions, and can achieve stable and accurate target tracking; it can be widely applied to fields such as video surveillance, autonomous driving, and human-computer interaction.

[0155] Example 3

[0156] To address the aforementioned technical problems, embodiments of the present invention also provide a computer device. Please refer to [link / reference needed]. Figure 4 , Figure 4 This is a basic structural block diagram of the computer device in this embodiment.

[0157] The aforementioned computer device 7 includes a memory 71, a processor 72, and a network interface 73 that are interconnected via a system bus. It should be noted that only the computer device 7 with components 71, 72, and 73 is shown in the figure; however, it should be understood that it is not required to implement all the components shown, and more or fewer components can be implemented alternatively. Those skilled in the art will understand that the computer device described here is a device capable of automatically performing numerical calculations and / or information processing according to pre-set or stored instructions, and its hardware includes, but is not limited to, microprocessors, application-specific integrated circuits (ASICs), programmable gate arrays (FPGAs), digital digital processors (DSPs), embedded devices, etc.

[0158] The aforementioned computer devices can be desktop computers, laptops, handheld computers, and cloud servers, among other computing devices. These devices can facilitate human-computer interaction with users through keyboards, mice, remote controls, touchpads, or voice-activated devices.

[0159] The aforementioned memory 71 includes at least one type of readable storage medium, including flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the aforementioned memory 71 may be an internal storage unit of the aforementioned computer device 7, such as the hard disk or memory of the computer device 7. In other embodiments, the aforementioned memory 71 may also be an external storage device of the aforementioned computer device 7, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 7. Of course, the aforementioned memory 71 may also include both the internal storage unit and its external storage device of the aforementioned computer device 7. In this embodiment, the aforementioned memory 71 is typically used to store the operating system and various application software installed on the aforementioned computer device 7, such as computer-readable instructions for a drone target recognition method. In addition, the aforementioned memory 71 can also be used to temporarily store various types of data that have been output or will be output.

[0160] In some embodiments, the processor 72 may be a central processing unit (CPU), controller, microcontroller, microprocessor, or other data processing chip. The processor 72 is typically used to control the overall operation of the computer device 7. In this embodiment, the processor 72 is used to execute computer-readable instructions stored in the memory 71 or to process data, for example, to execute computer-readable instructions for a UAV target recognition method.

[0161] The network interface 73 may include a wireless network interface or a wired network interface, which is typically used to establish a communication connection between the computer device 7 and other electronic devices.

[0162] The beneficial effects of implementing this embodiment are: it can collect a large amount of image and video data in the air and can quickly and accurately identify targets from massive amounts of data; it adopts an artificial intelligence-based target recognition method, especially deep learning technology, and automatically learns image features by constructing a complex neural network model to achieve efficient target detection, classification, tracking and other functions, and can achieve stable and accurate target tracking; it can be widely applied to fields such as video surveillance, autonomous driving, and human-computer interaction.

[0163] Example 4

[0164] The present invention also provides another embodiment, namely, providing a computer-readable storage medium storing computer-readable instructions that can be executed by at least one processor to cause the at least one processor to perform the steps of the above-described UAV target identification method.

[0165] The beneficial effects of implementing this embodiment are: it can collect a large amount of image and video data in the air and can quickly and accurately identify targets from massive amounts of data; it adopts an artificial intelligence-based target recognition method, especially deep learning technology, and automatically learns image features by constructing a complex neural network model to achieve efficient target detection, classification, tracking and other functions, and can achieve stable and accurate target tracking; it can be widely applied to fields such as video surveillance, autonomous driving, and human-computer interaction.

[0166] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0167] Obviously, the embodiments described above are merely some embodiments of the present invention, not all embodiments. The accompanying drawings show preferred embodiments of the present invention, but do not limit the patent scope of the present invention. The present invention can be implemented in many different forms; rather, these embodiments are provided to provide a more thorough and complete understanding of the disclosure of the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing specific embodiments, or make equivalent substitutions for some of the technical features. Any equivalent structures made using the content of this specification and drawings, directly or indirectly applied to other related technical fields, are similarly within the patent protection scope of this invention.

Claims

1. A method for identifying unmanned aerial vehicle (UAV) targets, characterized in that, Includes the following steps: The drone collects image data and preprocesses the image data. Using deep learning models, discriminative feature representations are extracted from preprocessed image data. Based on the feature representation, the target location is located in the image data, and the target is classified. Each category group is processed in parallel, performing target detection and tracking separately; Based on the results of target detection and tracking, the detection results at different time points are correlated to form the target's motion trajectory.

2. The UAV target recognition method according to claim 1, characterized in that, The step of acquiring image data by the drone and preprocessing the image data specifically includes: Equipped with high-definition cameras or sensors, it captures images and videos along a pre-set flight path; Filtering algorithms are used to remove noise from the acquired images and videos; Perform contrast adjustment, sharpening, and color correction on image and video data processed by filtering algorithms; Annotate the image and video data that have undergone contrast adjustment, sharpening, and color correction to clarify the location, category, and other information of the target.

3. The UAV target recognition method according to claim 1, characterized in that, The step of extracting discriminative feature representations from preprocessed image data using a deep learning model specifically includes: Select a deep learning model based on task requirements; Multi-layer convolution operations are used to extract low-level to high-level feature representations from the original image; Pooling operations reduce the dimensionality of feature maps, thereby reducing computational cost while preserving feature information. Adding a non-linear activation function between the convolutional and pooling layers increases the model's non-linear expressive power and improves feature extraction performance.

4. The UAV target recognition method according to claim 1, characterized in that, The step of locating the target position in the image data based on the feature representation and classifying the target specifically includes: Candidate regions are generated in the image using a sliding window or region proposal network method as potential target locations; Feature extraction is performed on each candidate region to obtain its corresponding feature representation; The extracted features are input into the classifier and regressor for classification and location regression. For multiple detection boxes of the same category, the NMS algorithm is used to remove redundant boxes, and the detection box with the highest score is retained as the final result.

5. The UAV target recognition method according to claim 1, characterized in that, The step of parallel processing for each category group, performing target detection and tracking separately, specifically includes: Based on the target category, the video stream or image data is divided into multiple independent processing units, each of which is responsible for processing the video stream or image data of a certain type of target. Data synchronization between processing units can be achieved through shared memory or message queues. Resource allocation is dynamically adjusted based on the workload and computational complexity of each processing unit; Target detection and tracking are performed based on the target's motion characteristics.

6. The UAV target recognition method according to claim 1, characterized in that, The step of correlating the detection results at different time points based on the target detection and tracking results to form the target's motion trajectory specifically includes: A data association method is used to match detection results at different time points to determine the detection results belonging to the same target; Based on data association, the target's state information is updated through state estimation; Based on the updated target status information, the target is continuously tracked and its movement trajectory is constantly updated.

7. A method for identifying unmanned aerial vehicle (UAV) targets according to any one of claims 1 to 6, characterized in that, After the step of correlating the detection results at different time points based on the target detection and tracking results to form the target's motion trajectory, the method further includes: When multiple targets are detected to intersect or occlude, the accurate position and trajectory of each target are determined based on the target's motion characteristics and shape features.

8. A target identification device for unmanned aerial vehicles (UAVs), characterized in that, include: The acquisition and preprocessing module is used to acquire image data by the UAV and to preprocess the image data. The feature extraction module is used to extract discriminative feature representations from preprocessed image data using a deep learning model; The category segmentation module is used to locate the target position in the image data based on the feature representation and to classify the target into categories. The detection and tracking module is used to process each category group in parallel, performing target detection and tracking separately; The association module is used to associate the detection results at different time points based on the target detection and tracking results to form the target's motion trajectory.

9. A computer device comprising a memory and a processor, wherein the memory stores computer-readable instructions, and the processor, when executing the computer-readable instructions, implements the steps of a UAV target identification method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-readable instructions, which, when executed by a processor, implement the steps of a UAV target recognition method as described in any one of claims 1 to 7.

Images