Unmanned aerial vehicle end license plate recognition method and related device

Abstract

The embodiment of the invention provides an unmanned aerial vehicle end license plate recognition method and a related device, and relates to the technical field of license plate detection. Acquiring a plurality of ground images acquired by the unmanned aerial vehicle, and performing vehicle detection on each ground image; determining a license plate detection algorithm for the ground image according to whether the to-be-analyzed vehicle exists in the ground image, and performing license plate detection on the ground image through the license plate detection algorithm to obtain a license plate area in the ground image; intercepting the ground image according to the license plate area to obtain a license plate image, and performing multi-modal processing on the license plate image to obtain a plurality of license plate images to be analyzed; wherein the image modals of the license plate images to be analyzed are different; and performing text recognition on each to-be-analyzed license plate image to obtain the candidate license plate number corresponding to each to-be-analyzed license plate image, and determining the target license plate number of the to-be-analyzed vehicle from the plurality of candidate license plate numbers, thereby improving the reliability and recognition accuracy of unmanned aerial vehicle license plate detection.

Description

Technical Field

[0001] This application relates to the field of license plate detection technology, and more specifically, to a method and related device for license plate recognition on a drone. Background Technology

[0002] In practical scenarios such as urban traffic management, illegal parking patrols, and emergency law enforcement, there is often a need for rapid identification and evidence collection of vehicles within a large dynamic area. Due to the limited coverage and deployment locations of fixed cameras, it is difficult to meet the regulatory requirements of high mobility and rapid response. Therefore, drones, with their flexible flight capabilities, wide-area coverage, and real-time image transmission capabilities, are gradually becoming a key carrier in intelligent inspection systems. Equipped with high-definition visual sensors, drones can continuously collect images of ground roads and surrounding areas during aerial patrols, thereby achieving remote perception and target analysis of moving or parked vehicles. The license plate, as the unique identifier of a vehicle, directly determines the practical value of the entire system through its automatic recognition capability.

[0003] In traditional automatic license plate recognition technology systems, the methods are primarily designed for fixed monitoring equipment deployed at road checkpoints, toll stations, or residential entrances. In these scenarios, the camera's imaging conditions are relatively stable, resulting in high-resolution license plate images with clearly discernible character structures, thus achieving high recognition accuracy and system stability. However, for the unique challenges of drones—high-altitude, dynamic, tilted, and multi-interference scenarios—the license plate recognition technologies exhibit significant limitations in adaptability, failing to support highly reliable license plate recognition tasks and resulting in low recognition accuracy. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a method and related device for license plate recognition on a drone, so as to improve the reliability and recognition accuracy of drone license plate detection.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of this application are as follows: In a first aspect, this application provides a license plate recognition method for unmanned aerial vehicles (UAVs), applied to the edge computing unit of an UAV, the method comprising: Multiple ground images collected by the drone are acquired, and vehicle detection is performed on each of the ground images; Based on whether the vehicle to be analyzed exists in the ground image, a license plate detection algorithm is determined for the ground image. The license plate detection algorithm is then used to detect the license plate in the ground image to obtain the license plate region in the ground image. The ground image is cropped based on the license plate area to obtain a license plate image. The license plate image is then subjected to multimodal processing to obtain multiple license plate images to be analyzed. Each license plate image to be analyzed has a different image modality. Text recognition is performed on each of the license plate images to be analyzed to obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and the target license plate number of the vehicle to be analyzed is determined from the multiple candidate license plate numbers.

[0006] In an optional implementation, the step of determining a license plate detection algorithm for the ground image based on whether a vehicle to be analyzed exists in the ground image, and performing license plate detection on the ground image using the license plate detection algorithm to obtain the license plate region in the ground image includes: If a vehicle to be analyzed is present in the ground image, the license plate detection algorithm is determined to be a local compensation detection algorithm. The license plate of the vehicle to be analyzed is detected according to the local compensation detection algorithm to obtain the license plate region in the ground image. If the vehicle to be analyzed is not found in the ground image, the license plate detection algorithm is determined to be a global search detection algorithm. The license plate is then detected in the ground image according to the global search detection algorithm to obtain the license plate region in the ground image.

[0007] In an optional implementation, the step of detecting the license plate of the vehicle to be analyzed according to the local compensation detection algorithm to obtain the license plate region in the ground image includes: The region coordinates of the vehicle to be analyzed in the ground image are obtained, and the region coordinates are expanded according to a preset expansion value to obtain the expanded region coordinates of the vehicle to be analyzed in the ground image. Based on the coordinates of the extended region, the vehicle image corresponding to the vehicle to be analyzed is extracted from the ground image. The vehicle image is then input into a pre-trained license plate detection model for processing to obtain the license plate region of the vehicle to be analyzed in the ground image. The step of detecting license plates in the ground image according to the global search detection algorithm to obtain the license plate region in the ground image includes: The ground image is input into a pre-trained license plate detection model for processing to obtain the license plate region in the ground image.

[0008] In an optional implementation, acquiring multiple ground images collected by the UAV and performing vehicle detection on each of the ground images includes: The ground video collected by the UAV is acquired, and the ground video is parsed to obtain multiple initial ground images; Each of the initial ground images is subjected to denoising and color space correction to obtain multiple ground images; Each of the ground images is input into a pre-trained vehicle recognition model for processing to detect vehicles in each ground image.

[0009] In an optional implementation, the step of performing multimodal processing on the license plate image to obtain multiple license plate images to be analyzed includes: The original image of the license plate is determined as the first license plate image to be analyzed, and the license plate image is processed into grayscale to obtain the second license plate image to be analyzed. Histogram equalization, adaptive thresholding, and CLAHE processing are performed on the second license plate image to be analyzed to obtain the third, fourth, and fifth license plate images to be analyzed.

[0010] In an optional implementation, the step of performing text recognition on each of the license plate images to be analyzed to obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and determining the target license plate number of the vehicle to be analyzed from the multiple candidate license plate numbers, includes: Each of the license plate images to be analyzed is subjected to text region detection and orientation classification to obtain the text region and rotation direction in the license plate image to be analyzed. Based on the text region and rotation direction in the license plate image to be analyzed, text recognition is performed on the license plate image to obtain the candidate license plate number and analysis confidence level corresponding to the license plate image to be analyzed. The license plate number with the highest analysis confidence and exceeding a preset threshold among the multiple candidate license plate numbers is determined as the target license plate number of the vehicle to be analyzed; The method further includes: Based on the target license plate number, a pre-stored Chinese font file is called to generate a Chinese license plate label. The ground image is then labeled with the Chinese license plate label, and the labeled ground image is sent to the server.

[0011] Secondly, this application provides a license plate recognition device for unmanned aerial vehicles (UAVs), applied to the edge computing unit of a UAV, the device comprising: The detection module is used to acquire multiple ground images collected by the UAV and perform vehicle detection on each of the ground images. The detection module is further configured to determine a license plate detection algorithm for the ground image based on whether there is a vehicle to be analyzed in the ground image, and to perform license plate detection on the ground image through the license plate detection algorithm to obtain the license plate region in the ground image; The processing module is used to crop the ground image according to the license plate area to obtain a license plate image, and to perform multimodal processing on the license plate image to obtain multiple license plate images to be analyzed; wherein, each license plate image to be analyzed has a different image modality; The processing module is further configured to perform text recognition on each of the license plate images to be analyzed, obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and determine the target license plate number of the vehicle to be analyzed from the multiple candidate license plate numbers.

[0012] Thirdly, this application provides an edge computing unit, including a processor and a memory, wherein the memory stores a computer program that can be executed by the processor, and the processor can execute the computer program to implement the method described in any of the foregoing embodiments.

[0013] Fourthly, this application provides a drone that includes the edge computing unit described in the foregoing embodiments.

[0014] Fifthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in any of the foregoing embodiments.

[0015] The UAV-based license plate recognition method and related apparatus provided in this application first acquire multiple ground images collected by the UAV, and then perform vehicle detection operations on each ground image to determine whether a vehicle to be analyzed exists in the image. Based on the determination of whether a vehicle to be analyzed exists in the ground image, the corresponding license plate detection algorithm is adaptively determined, and license plate detection is performed on the ground image according to the corresponding license plate detection algorithm. In this way, different license plate detection paths are executed based on the detection of ground vehicles, which can reduce the false negative rate of small targets and improve the overall recall rate of the system. At the same time, after obtaining the license plate region, the corresponding region can be cropped from the original ground image according to the coordinates of the region to generate a standard-sized license plate image. This license plate image is used as input data for the next stage and subjected to multimodal processing operations to generate multiple license plate images to be analyzed with different image modalities. Finally, the target license plate number is determined based on the candidate license plate numbers identified from the license plate images to be analyzed under multiple different modalities. Since images in each modality highlight different visual features, they can provide diverse input samples for subsequent text recognition. By utilizing the information redundancy and complementarity brought by multimodal input, the stability and accuracy of recognition decisions are enhanced, and the robustness and recognition accuracy of license plate recognition in complex environments are improved.

[0016] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A block diagram of the recognition system is shown; Figure 2 A block diagram of an edge computing unit provided in an embodiment of this application is shown; Figure 3 This paper illustrates a flowchart of a drone-based license plate recognition method provided in an embodiment of this application. Figure 4 A schematic diagram showing the performance comparison of different license plate recognition methods from the perspective of a drone is presented. Figure 5 The diagram shows a functional block diagram of a drone-based license plate recognition device provided in an embodiment of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0020] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0021] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0022] Traditional automatic license plate recognition technology, namely vision-based automatic license plate recognition (ALPR), has long played a key role in the field of intelligent transportation, and is widely used in scenarios such as road checkpoints, parking lot entrances and exits, and community security. These systems typically rely on fixed-installation cameras for image acquisition. Their working environment is relatively stable, and they have advantages such as close shooting distance, near-vertical or side-facing viewing angle, uniform lighting conditions, and high license plate imaging resolution.

[0023] Thanks to these favorable conditions, the traditional ALPR system can adopt a four-stage serial processing flow of "license plate detection - image preprocessing - character segmentation - character recognition", and combine Haar features and HOG features with SVM classifier, or use deep learning models such as convolutional neural networks and CRNN to complete the tasks of each stage. Under controlled environment, it can achieve high recognition accuracy and system stability.

[0024] However, when these mature technologies are transferred to drone platforms, their performance deteriorates significantly. The fundamental reason lies in the essential change in imaging conditions. Drones typically fly at altitudes of 50 to 150 meters during urban patrols, resulting in ground vehicles and their license plates occupying a very small portion of the image, often less than 30×10 pixels – a typical small target detection scenario. Simultaneously, due to dynamic flight attitude adjustments and the use of wide-angle lenses, the shooting angle is often tilted or even dived, causing significant perspective distortion in the license plate area. The characters no longer maintain a horizontal linear structure, severely interfering with subsequent character segmentation. Furthermore, the lightweight design of drones makes them susceptible to airflow and vibration, causing image blurring. Combined with the effects of high-altitude operations such as sunlight direction, building shadows, and rain / fog, the license plate area often experiences localized overexposure or underexposure, further deteriorating image quality. These factors combined make it difficult for traditional ALPR algorithms, originally designed for fixed scenes, to effectively extract features, leading to high false negative rates and frequent recognition errors, failing to meet the accuracy and reliability requirements of practical applications.

[0025] Based on this, embodiments of this application provide a method and related apparatus for license plate recognition on a drone to solve the above problems.

[0026] Specifically, Figure 1 For a block diagram of the recognition system, please refer to Figure 1 The identification system includes a drone and a server, and the drone communicates with the server.

[0027] Optionally, the drone system can achieve multi-level communication and data interaction through ground controllers, AI platforms and remote servers, building a complete technology chain from edge perception to cloud management.

[0028] The drone, serving as the front-end data acquisition and processing unit, integrates a flight platform and intelligent computing capabilities. Its core components include the drone itself, a high-definition camera, a flight control system, a wireless image transmission module, and a scalable edge computing unit. The high-definition camera is responsible for capturing visual information of ground vehicle scenes in real time and transmitting this information directly to the edge computing unit. This computing unit is integrated into the drone platform via a pluggable interface (such as USB), possessing the ability to independently run deep learning models. It can complete license plate recognition tasks on-site during flight without relying on remote servers for real-time inference, thereby reducing communication latency and improving response speed.

[0029] After completing localized recognition, the edge computing unit generates multimodal output data including detection results, structured text, and visual annotations, and transmits this data back to the ground control station in real time for operators to monitor and make decisions. Simultaneously, the edge computing unit can also upload raw images and / or recognition results to a remote server for long-term data storage and archiving. Based on this, the ground control station uses an AI platform as an intermediate scheduling and query hub to access historical data stored on the server as needed, retrieving relevant records for specific time periods, areas, or vehicles, supporting subsequent applications such as traffic analysis, illegal parking determination, and law enforcement tracing.

[0030] In one possible implementation method Figure 2 Please refer to the block diagram of the edge computing unit provided in the embodiments of this application. Figure 2 The edge computing unit includes a memory, a processor, and a communication module. These components are electrically connected directly or indirectly to enable data transmission or interaction. For example, they can be electrically connected via one or more communication buses or signal lines.

[0031] Memory is used to store computer programs or data that can be executed by a processor. Memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc.

[0032] The processor is used to read / write data or computer programs stored in the memory, and execute the computer program to implement the UAV license plate recognition method provided in the embodiments of this application.

[0033] The communication module is used to establish communication connections between the edge computing unit and other communication terminals via the network, and to send and receive data via the network.

[0034] It should be understood that, Figure 2 The structure shown is only a schematic diagram of an edge computing unit. An edge computing unit may also include components that are larger than... Figure 2 The more or fewer components shown, or having the same Figure 2 The different configurations shown. Figure 2 The components shown can be implemented using hardware, software, or a combination thereof.

[0035] The following is based on the above. Figure 2 The edge computing unit in the process is the execution entity. The drone-based license plate recognition method provided in this application is illustrated with a flowchart. Specifically, Figure 3 For a flowchart illustrating the drone-based license plate recognition method provided in this application, please refer to [link / reference]. Figure 3 The method includes: Step S20: Acquire multiple ground images collected by the drone, and perform vehicle detection on each ground image.

[0036] During the above steps, the edge computing module can acquire multiple ground images containing ground vehicle information based on the ground visual information collected by the high-definition camera in the drone, and process each ground image through a pre-trained vehicle recognition model to detect vehicles in each ground image.

[0037] Step S21: Determine the license plate detection algorithm for the ground image based on whether there is a vehicle to be analyzed in the ground image, and perform license plate detection on the ground image using the license plate detection algorithm to obtain the license plate region in the ground image.

[0038] Optionally, in the ground image processing flow acquired by the UAV, the edge computing unit first performs vehicle detection to determine whether the image contains identifiable vehicle targets. The detection result serves as the basis for subsequent license plate detection path selection.

[0039] In this embodiment, the edge computing unit can dynamically adjust the detection strategy and select different license plate detection algorithms for license plate detection in two situations: the presence of identifiable vehicles to be analyzed in the image and the absence of identifiable vehicles to be analyzed, thereby improving the recall rate of small targets.

[0040] Optionally, the license plate area can be the spatial coordinates of the license plate in the entire ground image, usually represented in the form of a bounding box, which consists of the pixel coordinates of the upper left and lower right corners of the rectangular area, used to accurately define the geometric range of the license plate.

[0041] Step S22: The ground image is cropped according to the license plate area to obtain the license plate image. The license plate image is then subjected to multimodal processing to obtain multiple license plate images to be analyzed.

[0042] Each license plate image to be analyzed has a different image modality.

[0043] Step S23: Perform text recognition on each license plate image to be analyzed to obtain the candidate license plate number corresponding to each license plate image to be analyzed, and determine the target license plate number of the vehicle to be analyzed from multiple candidate license plate numbers.

[0044] It should be understood that in the technical scenario of license plate recognition on drones, due to factors such as high-altitude viewing, tilted shooting angle, and complex lighting conditions, directly recognizing characters from the original image often fails to yield stable and reliable output results. To address this issue, this embodiment introduces a multimodal processing and multipath recognition fusion mechanism, aiming to improve the robustness and accuracy of license plate text recognition under low-resolution, distorted, and non-uniform lighting conditions.

[0045] During the above steps, the edge computing unit first uses the spatial coordinates of the license plate area to precisely extract a local image containing the license plate content from the ground image, thus obtaining the license plate image. This license plate image serves as input data for subsequent processing, carrying the character information to be recognized. However, its image quality may be affected by backlighting, shadows, or fog, making a single image modality unable to adapt to diverse environmental influences.

[0046] Based on this, in order to enhance the system's adaptability to diverse degradation modes, the edge computing unit can perform multimodal processing on license plate images to generate multiple license plate images that differ in visual features to be analyzed.

[0047] In this embodiment, each license plate image to be analyzed corresponds to a specific image enhancement strategy, thereby forming different image modalities. These images of different modalities together constitute a set of complementary input samples, which can cover more potential effective recognition paths.

[0048] Subsequently, the multiple license plate images to be analyzed were processed independently to obtain the candidate license plate numbers corresponding to each image. Due to the differences in the enhancement processing methods of each image modality, the recognition results of the same license plate may differ under different paths. Therefore, it is still necessary to finally determine the target license plate number with the highest confidence from multiple candidate license plate numbers.

[0049] Understandably, by performing multimodal image generation and parallel text recognition on license plate images, multiple candidate license plate numbers can be obtained. Finally, the unique target license plate number can be determined from the multiple candidate license plate numbers. This can effectively suppress the risk of misidentification caused by the deterioration of the quality of a single sample, thereby significantly improving the stability and accuracy of license plate recognition in complex imaging environments.

[0050] The UAV-based license plate recognition method provided in this application first acquires multiple ground images collected by the UAV, and performs vehicle detection operations on each ground image to determine whether a vehicle to be analyzed exists in the image. Based on the determination of whether a vehicle to be analyzed exists in the ground image, the corresponding license plate detection algorithm is adaptively determined, and license plate detection is performed on the ground image according to the corresponding license plate detection algorithm. In this way, different license plate detection paths are executed based on the detection of ground vehicles, which can reduce the false negative rate of small targets and improve the overall recall rate of the system. At the same time, after obtaining the license plate region, the corresponding region can be cropped from the original ground image according to the coordinates of the region to generate a standard-sized license plate image. This license plate image is used as input data for the next stage and subjected to multimodal processing operations to generate multiple license plate images to be analyzed with different image modalities. Finally, the target license plate number is determined based on the candidate license plate numbers identified from the license plate images to be analyzed under multiple different modalities. Since images in each modality highlight different visual features, they can provide diverse input samples for subsequent text recognition. By utilizing the information redundancy and complementarity brought by multimodal input, the stability and accuracy of recognition decisions are enhanced, and the robustness and recognition accuracy of license plate recognition in complex environments are improved.

[0051] In the process of UAVs performing ground vehicle identification tasks, the quality of image data acquisition and preprocessing directly affects the accuracy and stability of subsequent detection and identification stages. Because UAVs are affected by factors such as sensor noise, atmospheric disturbances, lighting changes, and motion blur during flight, directly using raw acquired images for analysis may lead to false detections or missed detections. Therefore, this embodiment aims to improve the quality and consistency of input data by constructing a complete front-end processing workflow from video stream parsing to image enhancement, providing a reliable data foundation for subsequent vehicle detection.

[0052] In this embodiment, the edge computing unit can acquire ground video collected by the UAV, perform video stream parsing on the ground video to obtain multiple initial ground images, perform denoising processing and color space correction on each initial ground image to obtain multiple ground images, and input each ground image into a pre-trained vehicle recognition model for processing to detect vehicles in each ground image.

[0053] During the execution of the above steps, the edge computing unit can first acquire ground video captured by a high-definition camera mounted on the drone platform. This high-definition camera can be a high-definition RGB camera, which records urban road scenes in the form of continuous frames, uses RGB color space encoding, and has fixed resolution and frame rate parameters.

[0054] Based on this, the edge computing unit can perform video stream parsing on the acquired ground video, decomposing the continuous video data into independent image frames arranged in a time sequence, thereby obtaining multiple initial ground images. ( Each initial ground image serves as the basic unit for subsequent processing, carrying visual information about ground features at a specific moment.

[0055] in, Characterizing the initial ground image It is a three-dimensional real matrix, belonging to the real number space. One of the elements in the image is the pixel dimension. The first and second dimensions correspond to the spatial resolution of the image, namely height H and width W, reflecting the image's size in pixel coordinates. The third dimension is the channel dimension; the value "3" indicates that the image uses the RGB (Red-Green-Blue) color space encoding, containing three color channels. Each pixel consists of three floating-point or integer values, representing the intensity values ​​of the three primary colors—red, green, and blue—at that location, collectively forming color visual information.

[0056] In this embodiment, H=1080 and W=1920, which means that each original image frame has a pixel array of 1080 rows and 1920 columns, forming a 1080P high-definition image.

[0057] As can be seen, this mathematical expression not only clearly defines the data format of the input image, but also provides a clear input format basis for subsequent denoising processing, model inference and other operations, ensuring that the entire license plate recognition process runs stably under a unified data representation framework.

[0058] Furthermore, to suppress random noise introduced during the imaging process and correct any possible color distortion, the edge computing unit can perform denoising and color space correction on each initial ground image.

[0059] The denoising process is achieved by applying a Gaussian filter kernel through convolution operations. This kernel generates weight coefficients based on a two-dimensional normal distribution function, smoothing high-frequency noise while preserving image edge structures as much as possible. Color space correction is used to compensate for color shifts caused by changes in lighting color temperature or camera white balance deviations, ensuring consistency in color representation across images acquired at different times and under different weather conditions. After these processes, the initial ground image is transformed into a higher-quality image format more suitable for model inference, resulting in multiple ground images.

[0060] In one possible implementation, noise reduction can be achieved using the following formula:

[0061] in, Characterizes the image after denoising. The high-speed filter kernel is represented by x and y, which are local coordinate offsets with the center of the filter kernel as the origin. They represent the horizontal and vertical distances of each position in the neighborhood of the currently calculated pixel relative to the center point. denoted as the standard deviation of the Gaussian distribution.

[0062] Each processed ground image is then input into a pre-trained vehicle recognition model for inference calculations.

[0063] Optionally, the vehicle recognition model can be designed based on the YOLOv11s architecture, with its backbone network using CSPDarknet53 to extract multi-level semantic features and output multi-scale feature maps to adapt to targets of different sizes. The detection head adopts a decoupled structure, independently predicting the probability of the bounding box belonging to a specific category. and the regression offset of the bounding box relative to the anchor point This improves the precision of classification and location tasks.

[0064] During the inference phase, only detection results classified as "car" are retained, i.e., target instances satisfying c=car, and their bounding boxes are set accordingly. Filtering: Based on preset category confidence thresholds The detection results are screened, and low-confidence candidate boxes below the threshold are removed; non-maximum suppression (NMS) technology is further applied to set the intersection-union ratio (IU) threshold. This process eliminates redundant detection boxes within overlapping areas, ultimately outputting a structured set of vehicles. Each element This indicates a detected vehicle to be analyzed.

[0065] In one possible implementation method ,in, The bounding box of the vehicle to be detected is used to represent the vehicle's coordinate position in the ground image. The detection score represents the overall confidence level of the target detection model in the existence of a target vehicle category within a candidate bounding box. The higher the detection score, the greater the probability that the bounding box is identified as a real vehicle target.

[0066] It should be noted that the bounding box of the vehicle to be detected refers to a rectangular representation method used to accurately locate the spatial position of the vehicle target in a two-dimensional image coordinate system.

[0067] In this embodiment, the bounding box is defined as an axis-aligned minimum bounding rectangle, whose four parameters correspond to the extreme points in the horizontal and vertical directions of the image pixel coordinate system: where and This represents the x and y coordinates of the top-left vertex of the rectangle, while and This represents the x and y coordinates of the lower right corner vertex.

[0068] In one possible implementation, the aforementioned category confidence threshold... It can be set to 0.3, the cross-union threshold. It can be set to 0.5.

[0069] In this embodiment, the edge computing unit needs to process the ground image frame by frame to complete the vehicle detection task for all relevant frames in the entire video, ensuring continuous perception capability in dynamic scenes.

[0070] Therefore, in this embodiment, the edge computing model effectively improves the quality of the input image and the interpretability of the model by systematically analyzing and enhancing the video stream captured by the drone. This technical approach not only enhances the adaptability to complex environmental factors but also provides a stable and reliable prerequisite for the subsequent determination of license plate detection algorithms and license plate recognition.

[0071] The following section addresses how to determine a license plate detection algorithm for a ground image based on the presence of the vehicle to be analyzed. It then proposes a possible implementation method for detecting license plates in the ground image using this algorithm to obtain the license plate region.

[0072] In one possible implementation, if the vehicle to be analyzed is present in the ground image, the edge computing unit can determine that the license plate detection algorithm is a local compensation detection algorithm, and perform license plate detection on the vehicle to be analyzed according to the local compensation detection algorithm to obtain the license plate region in the ground image; if the vehicle to be analyzed is not present in the ground image, the edge computing unit can determine that the license plate detection algorithm is a global search detection algorithm, and perform license plate detection on the ground image according to the global search detection algorithm to obtain the license plate region in the ground image.

[0073] It should be understood that ground images captured by drones during high-altitude flight are characterized by tilted viewpoints, imaging distortion, and complex lighting conditions. Furthermore, vehicles and license plates occupy extremely small pixel areas within these images, making traditional single-path-based license plate detection methods prone to missed detections. Research has found that relying solely on license plate detection models to directly locate license plates globally results in a significant decrease in recall when faced with low-resolution, blurry, or partially occluded license plates, thus impacting the integrity of subsequent recognition processes and the overall system performance.

[0074] The core concept of this embodiment is to improve the license plate localization capability under different visual conditions by dynamically selecting a license plate detection algorithm that adapts to the current image content.

[0075] Specifically, the edge computing unit can adaptively switch to the appropriate detection strategy based on whether a vehicle has been identified in the image: when the vehicle to be analyzed is present in the ground image, it indicates that the system has completed the localization of the vehicle. At this time, the spatial position of the vehicle can be used as prior information to focus the detection on a local area, thereby improving the sensitivity to small license plates. In this case, the edge computing unit can determine that the license plate detection algorithm is a local compensation detection algorithm, and perform license plate detection on the vehicle to be analyzed based on this algorithm to obtain more accurate license plate area coordinates.

[0076] Understandably, since the vehicle area itself has a larger spatial scale and more obvious texture features than the license plate, its detection stability is much higher than that of directly detecting tiny license plates. Therefore, using the known vehicle position to guide the license plate location can effectively improve the recall rate of the system without increasing the computational burden of the whole map.

[0077] Conversely, if the vehicle to be analyzed is not present in the ground image, it means that prior vehicle information cannot be used to assist in localization. In this case, it is necessary to maintain attention on the entire image range to avoid missing isolated license plate instances. In this situation, the edge computing unit can determine that the license plate detection algorithm is a global search detection algorithm and perform a comprehensive scan of the ground image based on this algorithm, ensuring that the basic license plate detection task can still be completed even in scenarios without any prior conditions. This global search process covers the entire frame of the image and is suitable for situations where the vehicle is not identified or the image only contains isolated license plates, ensuring the completeness of the detection logic.

[0078] Therefore, the above steps, by understanding the semantic content of the image—that is, whether the vehicle to be analyzed exists—determine which detection path to adopt, thus achieving conditional traffic splitting of the detection strategy. This context-aware algorithm switching mechanism retains the universality of global detection while integrating the high sensitivity of local compensation detection, forming a dual-modal detection framework with context adaptability. By introducing this setting, the aim is to address the diverse imaging environments faced by UAV platforms during dynamic inspections, thereby improving the overall robustness and reliability of license plate detection in complex urban road scenarios.

[0079] In one possible implementation, the edge computing unit can perform license plate detection on the vehicle to be analyzed based on a local compensation detection algorithm to obtain the license plate region in the ground image through the following steps: The system obtains the region coordinates of the vehicle to be analyzed in the ground image, expands the region coordinates according to a preset expansion value, and obtains the expanded region coordinates of the vehicle to be analyzed in the ground image. Based on the expanded region coordinates, the system extracts the vehicle image corresponding to the vehicle to be analyzed from the ground image, and inputs the vehicle image into a pre-trained license plate detection model for processing to obtain the license plate region of the vehicle to be analyzed in the ground image.

[0080] In this embodiment, the edge computing unit can improve its ability to capture low-resolution, distorted, or partially obscured license plates by narrowing the detection range and enhancing local attention, thereby improving the recall rate of license plate recognition.

[0081] During the above steps, the edge computing unit can first obtain the region coordinates of the vehicle to be analyzed in the ground image. These coordinates represent the spatial range of the vehicle's main body in the image. In one possible implementation, these region coordinates refer to the bounding box of the vehicle to be analyzed. .

[0082] To ensure that no license plate area is missed, especially when the license plate is near the vehicle boundary, has an installation offset, or there is an error in the coverage of the area coordinates for vehicle positioning, the area coordinates need to be extended. Based on this, the edge computing unit can extend the area coordinates according to a preset extension value to obtain the extended area coordinates of the vehicle to be analyzed in the ground image.

[0083] In one possible implementation, the expansion operation refers to the expansion within the original vehicle bounding box. Based on this, the search window is appropriately expanded outwards in all directions to form a local search window that includes the vehicle and its surrounding neighborhood, effectively covering possible license plate locations. This can be represented by the following formula:

[0084] in, Representing the coordinates of the extended region, The extended value represents the application and can be preset according to the actual application, for example, set to... .

[0085] In this process, the determination of the expanded region coordinates is not an indiscriminate amplification, but rather the expansion amplitude is controlled based on empirical scale parameters to balance detection accuracy and computational cost. Subsequently, the edge computing unit can extract the vehicle image corresponding to the vehicle to be analyzed from the ground image based on the expanded region coordinates. This vehicle image is a sub-region of the original image, focusing on the local scene associated with the specific vehicle. Then, the edge computing unit can input the vehicle image into a pre-trained license plate detection model for processing, allowing the model to concentrate computational resources in high signal-to-noise ratio local areas to focus on identifying license plate instances associated with that vehicle. Due to the reduced input image size and lower background interference, the model's responsiveness to subtle textures and structural features is enhanced, thereby improving detection sensitivity under complex imaging conditions.

[0086] In one possible implementation, the license plate detection model could be YOLO-LP.

[0087] In this embodiment, a single ground image may contain multiple vehicles to be analyzed. The edge computing unit can perform the above operations for each vehicle to be analyzed, ultimately obtaining the set of license plates in the ground image. ,and , where each element This represents a detected license plate. Represents the license plate region in a ground image. The detection score characterizing the license plate. The vehicle to be analyzed is represented by the license plate.

[0088] Optionally, It can be the rectangular bounding box of the license plate, which is composed of the x and y coordinates of the top left and bottom right vertices of the rectangular bounding box.

[0089] Therefore, the above steps construct a compensatory detection path by effectively reusing vehicle location information. This strategy of using "vehicle positioning" to assist "license plate lookup" is essentially a phased detection mechanism based on spatial correlation. It avoids the efficiency loss caused by blindly searching the entire map and compensates for the shortcomings of single global detection in small target recognition. By introducing this setting, the aim is to address the issue of target scale changes caused by the flight altitude and dynamic perspective of the UAV platform, thereby improving the system's overall recall capability of license plate targets in real urban environments.

[0090] In another possible implementation, the edge computing unit can perform license plate detection on the ground image using a global search detection algorithm to obtain the license plate region in the ground image through the following steps: The ground image is input into a pre-trained license plate detection model for processing to obtain the license plate region in the ground image.

[0091] It should be understood that, considering the possibility that a vehicle may be present in the ground image but is partially obscured or not detected due to detection errors, an independent detection mechanism that does not rely on the location results of the preceding vehicle is needed to ensure that the license plates of isolated or difficult-to-detect vehicles can still be detected.

[0092] Specifically, when it is determined that there is no vehicle to be analyzed in the ground image, the edge computing unit can enable a global search detection algorithm. The core of this algorithm is to directly scan and identify the license plate area within the entire range of the ground image.

[0093] During the above steps, the edge computing unit can directly input the pre-processed ground image into a pre-trained license plate detection model, such as YOLO-LP. This model is based on a lightweight network structure design and has been optimized for license plate targets under small-scale, distorted, and long-distance imaging. It can effectively extract the unique texture and geometric features of license plates under complex background interference.

[0094] In this process, the model can output a set of license plates from the ground image. Understandably, , where each element This represents a detected license plate. Represents the license plate region in a ground image. The detection score characterizing the license plate. The vehicle to be analyzed is represented by the license plate.

[0095] Optionally, It can be the rectangular bounding box of the license plate, which is composed of the x and y coordinates of the top left and bottom right vertices of the rectangular bounding box.

[0096] The UAV license plate recognition method provided in this application uses a global search detection algorithm to achieve indiscriminate coverage detection of the entire image without relying on prior vehicle location information, thereby ensuring basic detection capability in edge cases such as vehicle detection failure or sparse scene.

[0097] This step can be understood as constructing a fallback detection path, complementing the vehicle-region-based local compensation detection, and together forming the technical foundation of the multi-path detection mechanism. Therefore, by directly feeding the entire ground image into a dedicated detection model for processing, not only are the system's applicable scenarios expanded, but its robustness in handling abnormal or challenging images is also enhanced.

[0098] The following section provides a possible approach to performing multimodal processing on license plate images to obtain multiple license plate images for analysis.

[0099] Specifically, the edge computing unit can determine the original image of the license plate as the first license plate image to be analyzed, and perform grayscale processing on the license plate image to obtain the second license plate image to be analyzed; and perform histogram equalization processing, adaptive thresholding processing and CLAHE processing on the second license plate image to be analyzed to obtain the third, fourth and fifth license plate images to be analyzed.

[0100] It should be understood that in the high-altitude shooting environment of drones, due to the long imaging distance, tilted viewing angle, and complex and variable lighting conditions, the captured license plate images often suffer from degradation phenomena such as low contrast, local overexposure or underexposure, and blurred characters, making it difficult for a single image form to stably support subsequent text recognition tasks. To address such uncertainties, the drone-based license plate recognition method provided in this application requires diversified enhancement processing of the original license plate image to improve its readability and feature integrity under different degradation modes.

[0101] In this embodiment, the edge computing unit can generate multiple image versions with different visual characteristics in parallel, providing multi-angle information input for subsequent recognition processes.

[0102] During the above steps, the edge computing unit can use the license plate image extracted from the ground image as the initial data source and directly retain its original image as the first license plate image to be analyzed, ensuring that the original color and texture information is not lost. On this basis, grayscale processing is performed on the same license plate image to obtain the second license plate image to be analyzed. Histogram equalization, adaptive thresholding, and CLAHE (Contrast Limited Adaptive Histogram Equalization) processing are then performed on the second license plate image to be analyzed to generate the third to fifth license plate images to be analyzed.

[0103] In one possible implementation, the original license plate image is directly retained as the first license plate image to be analyzed. ,in, Characterize the first license plate image to be analyzed. This is the original image of the license plate.

[0104] Grayscale processing is performed using the following formula:

[0105] in, Representing license plate image The corresponding second license plate image to be analyzed is a grayscale image. This is the original image of the license plate.

[0106] Histogram equalization is performed using the following formula:

[0107] in, Representing license plate image The corresponding pixels in the third license plate image to be analyzed, i.e., the histogram equalized image, Characterize the pixels in the second license plate image to be analyzed. , Image size, grayscale The number of pixels.

[0108] Adaptive thresholding is performed using the following formula:

[0109] in, The pixels in the fourth license plate image to be analyzed are represented by the adaptive threshold image. Characterize the pixels in the second license plate image to be analyzed. It is the local neighborhood mean. .

[0110] CLAHE processing is performed using the following formula:

[0111] in, The fifth image to be analyzed is the CLAHE image. Characterize the second license plate image to be analyzed.

[0112] In this process, grayscale processing converts the three-channel RGB image into a single-channel grayscale image, eliminating color interference and reducing computational complexity, providing a basic input for subsequent enhancement operations based on brightness distribution; histogram equalization processing expands the overall dynamic range of the image by remapping the pixel intensity distribution, effectively improving the global low contrast problem caused by backlighting or shadows; adaptive thresholding sets a dynamic segmentation threshold based on the local neighborhood mean, which can highlight character outlines and suppress the influence of background noise in uneven lighting scenes; CLAHE processing further refines the contrast enhancement capability of local areas, and significantly improves the visibility of edge details while avoiding excessive amplification of noise through a block histogram cropping and remapping mechanism.

[0113] The core concept of this embodiment lies in constructing five image modalities—original, grayscale, global enhancement, local binarization, and local contrast optimization—to form a set of highly complementary and widely covered input samples. Based on this, even if a preprocessing path suffers character information distortion or breakage due to specific environmental factors (such as strong reflection or rain / fog obstruction), other paths may still retain complete or identifiable structural features, thus providing diverse recovery opportunities for license plate text recognition. In essence, this multimodal processing mechanism introduces redundant and heterogeneous data representation paths into the recognition system, enhancing its adaptability to complex imaging conditions.

[0114] Optionally, the edge computing unit can perform text region detection and orientation classification on each license plate image to be analyzed, obtain the text region and rotation direction in the license plate image to be analyzed, and perform character recognition on the license plate image to be analyzed based on the text region and rotation direction in the license plate image to be analyzed, obtain the candidate license plate number and analysis confidence level corresponding to the license plate image to be analyzed, and determine the license plate number with the highest analysis confidence level and exceeding the preset threshold among multiple candidate license plate numbers as the target license plate number of the vehicle to be analyzed.

[0115] In this embodiment, in order to improve the stability and reliability of the recognition output, a multi-path parallel recognition mechanism needs to be introduced, and the optimal result can be selected by quantitatively evaluating the credibility of the results of each path.

[0116] Specifically, the edge computing unit can first use multiple license plate images generated through multimodal processing as input data. Each image represents a specific enhancement mode. Text region detection and orientation classification operations are performed independently for each license plate image. Text region detection is used to locate the spatial range of character distribution in the image to form text regions. Orientation classification is used to determine whether the text is rotated or tilted, and outputs the corresponding rotation angle information for correction in the subsequent recognition stage.

[0117] Based on this, the edge computing unit can perform text recognition processing on each license plate image to be analyzed based on the acquired text region and rotation direction. This process can use an OCR engine with end-to-end recognition capabilities to complete character sequence decoding and output the candidate license plate number corresponding to the image modality and its associated analysis confidence. This value reflects the internal probability score of the recognition result and reflects the degree of certainty of the model for this recognition result.

[0118] In one possible implementation, the edge computing unit can perform the above operations in parallel on five paths: the original image, the grayscale image, the histogram equalized image, the adaptive thresholded image, and the CLAHE-processed image, ultimately generating a result set containing multiple candidate license plate numbers and their respective analysis confidence scores. ,and ,in, Characterize the candidate license plate number corresponding to the license plate image k to be analyzed. Characterizes the analysis confidence level corresponding to the license plate image k to be analyzed. The license plate image k to be analyzed is represented.

[0119] Understandably, the above candidate license plate numbers refer to the text sequence corresponding to the license plate.

[0120] Furthermore, the edge computing unit can select candidate license plate numbers with the highest analysis confidence and exceeding a preset threshold from this set as the target license plate numbers of the vehicle to be analyzed. ,in, Represents the target license plate number. The confidence level of the analysis characterizing the target license plate number.

[0121] Optionally, to facilitate users' subsequent viewing of the analysis results, a multimodal processing method identifier can be added to each candidate license plate number. ,and Wherein, raw represents the original license plate image and grayscale processing, eq represents histogram equalization processing, adap represents adaptive thresholding processing, and clahe represents CLAHE processing.

[0122] Optionally, the preset threshold here can be used to filter out low-confidence outputs and avoid adopting incorrect results. In addition, when the highest confidence level does not reach the threshold, a review mechanism can be activated for supplementary judgment.

[0123] In a possible implementation manner, the review mechanism can be manual review, that is, both the above analysis results and process data are sent to the corresponding processing personnel, and the processing personnel confirm the target license plate number.

[0124] The core idea of this embodiment is that by mapping the same license plate object to multiple image modalities for parallel solution, an identification decision framework supported by multiple evidences is constructed, thus breaking through the limitation of traditional methods relying on single-sample discrimination. In this way, not only the compensation advantages of different image enhancement methods for specific degradation modes are fully utilized, but also the dynamic evaluation of the recognition quality is achieved through the confidence index. It can be understood that this method essentially establishes a fusion decision mechanism based on the principle of confidence priority, effectively improving the accuracy and robustness of the license plate recognition result in complex real scenarios.

[0125] Optionally, to ensure that the recognized license plate text conforms to the basic coding specification of motor vehicle license plates, the edge computing unit needs to perform format consistency verification on the candidate license plate numbers before generating the final result.

[0126] The core purpose of this step is to exclude obviously incorrect recognition results that do not conform to the actual license plate format from the semantic structure level, thereby improving the basic reliability of the output result. A typical license plate number usually consists of seven or eight characters, where the first character is a Chinese character representing the abbreviation of the province where the vehicle belongs, and the subsequent characters are依次 combinations of letters and numbers, and the whole has a fixed arrangement pattern. If the recognition result deviates from this structure, it is very likely to be misrecognized.

[0127] To achieve the above goal, the edge computing unit can compare the candidate license plate numbers output by each path with the preset rule database one by one according to the character position. In specific implementation, first verify whether the first character of the character sequence belongs to the set of Chinese characters corresponding to provincial administrative regions, such as "Jing", "Hu", "Yue", etc., to confirm the legality of its use as a provincial abbreviation; then check whether the second character is a capital English letter, representing the license-issuing authority code; and then determine the type of the remaining characters to ensure that they are numbers or within the specified letter range. This process is completed by querying the built-in structured database, which stores all legal provincial abbreviations, permitted letter ranges, and character length combinations, forming a complete basis for format matching.

[0128] After completing the format verification, only the candidate license plate numbers that meet the conditions are retained and enter the subsequent confidence screening stage. In this stage, the result with the highest analysis confidence that exceeds the preset threshold is preferentially selected as the target output to ensure the optimal recognition quality. When multiple candidate results with compliant formats have the same highest confidence, the character length can be further introduced as a discriminant basis, and the result with a length of seven digits is selected as the final target license plate number.

[0129] It should be understood that the standard length of civilian motor vehicle license plates is usually 7 characters, consisting of 1 Chinese character (provincial abbreviation), 1 English letter (issuing authority code), and a combination of 5 letters or numbers, which is applicable to the vast majority of small cars. Although there are some special vehicles using 8-digit license plates, such as new energy vehicles adopting the green license plate coding rule (e.g., "Beijing A·D12345"), the proportion of such vehicle types in the overall traffic flow is relatively low, and their license plate structures have obvious separator characteristics, which can be distinguished through dedicated recognition logic. In the actual UAV inspection scenario, the vehicles captured are mainly traditional fuel passenger cars, and the vast majority of their license plates are in the 7-digit format.

[0130] In this context, if there are several candidate license plate numbers with the same highest analysis confidence, the edge computing unit needs to introduce additional criteria to determine the unique output and avoid decision ambiguity. That is, the result with a character length of 7 is selected as the preferred output, not to deny the existence of 8-digit license plates, but based on the comprehensive consideration of statistical prior knowledge and engineering practicality.

[0131] This strategy can effectively reduce the risk of misjudgment of abnormal long sequences caused by character adhesion, breakage, or noise interference, and at the same time conforms to the coding rules of mainstream vehicles, improving the output stability of the system under high-concurrency and low-quality image conditions.

[0132] In addition, it should be noted that this length preference mechanism does not mean that the system cannot handle 8-digit new energy license plates. In actual deployment, the license plate color (blue plate / green plate) or structural characteristics (whether there are separation dots) can be identified through a pre-classification module to divert the independent processing path for new energy vehicles. In the unified decision-making stage, using the 7-digit length as the default preference standard under the same confidence level helps to simplify the decision-making logic, improve the processing efficiency, and meet the actual operation requirements of resource-constrained UAV devices in the edge computing environment.

[0133] It can be seen that this design is a technical trade-off to achieve a balance between ensuring recognition accuracy and system robustness, rather than a forced normalization for all license plate types.

[0134] In the typical application scenario of UAVs performing urban traffic inspection tasks, the final output visual image is often used as a key basis for law enforcement evidence collection, event recording, or monitoring display, and needs to contain directly readable semantic information for manual verification or system discrimination.

[0135] However, traditional license plate recognition systems have significant limitations in the result presentation phase, especially when dealing with the rendering of Chinese characters, facing inherent defects in the underlying technical architecture. It should be noted that most general optical character recognition (OCR) engines and image processing libraries, such as Tesseract OCR and OpenCV (Open Source Computer Vision Library), do not have built-in support for the Unicode Chinese character set in their default graphics rendering modules. Especially without an additional font resource loading mechanism, they are unable to correctly generate Chinese text pixel data. Therefore, when the edge computing unit attempts to overlay the recognized license plate number (such as "Beijing A12345") and prompt text (such as "License Plate:") onto the original ground image, character garbling, box substitution, or complete absence of characters are very likely to occur.

[0136] In this process, without native Chinese annotation capabilities, technicians often need to rely on external image editing tools or post-processing scripts to manually add Chinese labels offline. This operation process not only breaks the real-time linkage between recognition and visualization but also introduces additional manual intervention nodes and processing delays.

[0137] Obviously, due to the adaptation gap between the image rendering module and the local language environment in the related technologies, the overall architecture tends to be closed, making it difficult to meet the requirements for the immediacy, integrity, and standardization of the result images in domestic intelligent transportation management scenarios. Especially during the dynamic cruise of drones, a large number of video frames need to be continuously processed to form a structured evidence chain. If each frame of the image requires subsequent manual annotation, it will greatly reduce the automation level and practical usability of the system.

[0138] Based on this, the edge computing unit can also call the pre-stored Chinese font file according to the target license plate number to generate a Chinese license plate label, annotate the ground image according to the Chinese license plate label, and send the annotated ground image to the server.

[0139] It can be understood that in this embodiment, by introducing a font rendering mechanism independent of the image processing process, it is ensured that Chinese semantic information can be accurately overlaid on the original image.

[0140] Specifically, during the execution of the above steps, after obtaining the target license plate number of the vehicle to be analyzed, the edge computing unit first calls the pre-stored Chinese font file to generate a Chinese license plate label. The Chinese font file here refers to a font resource file that conforms to the TrueType format specification, such as "SimHei.ttf", which is pre-embedded in the system runtime environment and can be loaded and parsed by an image processing library (such as the Python Imaging Library, PIL).

[0141] In this process, the generation of Chinese license plate labels not only includes the target license plate number itself, but also includes prefix text with semantic prompts, such as "license plate:", as well as optional additional information such as "confidence level: X.XXX", thereby enhancing the information density and readability of the labeled content.

[0142] Furthermore, the edge computing unit annotates the ground image with license plates based on the generated Chinese license plate labels. This annotation operation is performed in the spatial coordinate system of the original ground image, and the annotation position is dynamically determined based on the geometric distribution of the license plate area in the image. It is usually located below or to the side of the license plate bounding box to avoid covering key visual features.

[0143] To ensure text clarity and visual consistency, the edge computing unit can be configured with uniform rendering parameters, including a 24-pixel font size, a red text color (RGB value: 255, 0, 0), and anti-aliasing technology to smooth character edges. The rendering process employs a transparent background overlay, allowing the annotation layer to blend naturally with the original image content, preserving the authenticity of the scene while achieving an intuitive overlay of semantic information.

[0144] The UAV-based license plate recognition method provided in this application allows the edge computing unit to directly call Chinese font files to complete Chinese annotation within the recognition process, without relying on external tools or post-processing steps, significantly improving the real-time performance and completeness of the output. Finally, the annotated ground image is sent to the server as a visualization result, which can be used for various application scenarios such as remote monitoring, manual review, or evidence archiving.

[0145] After completing license plate recognition and Chinese character annotation, the edge computing unit can also generate two types of complementary result data and send them to the server to adapt to the needs of different application scenarios.

[0146] One type is the output of a visualized image, which integrates Chinese semantic annotation information, fully preserving the content of the original ground image and the spatial correspondence of the recognition results. This annotated image can then be encoded in JPEG format, with a compression quality factor set to 95, effectively controlling data size while ensuring visual clarity, facilitating storage and transmission. This image can be saved locally or pushed to remote terminals in real time via a network interface, and is widely used in scenarios such as manual review, preservation of law enforcement evidence, or real-time display on urban surveillance screens.

[0147] Another type is structured data output, designed to support automated processing and integration of subsequent systems. This data object encapsulates key information in the recognition process in a standardized form, including a timestamp recording the completion time of processing, using the ISO8601 standard format (YYYY-MM-DDTHH:MM:SS.sssZ) to ensure the uniformity and traceability of time information; vehicle bounding boxes and license plate bounding boxes are described in the form of quaternion integer tuples to describe their precise positions in the image, providing basic coordinate data for spatial analysis and target association; the license plate text field carries the final identified 7 or 8-character sequence, conforming to the encoding specifications of motor vehicle license plates; the text confidence score reflects the reliability level of the recognition result as a floating-point value, providing a quantitative reference for the decision-making system; the detection path identifier clearly records that the result comes from "vehicle_search" (local compensation detection algorithm) or "global_detection" (global search detection algorithm), which helps to analyze the behavioral characteristics and performance of the detection process; the source image identifier serves as a unique label to associate the input image with the output result, ensuring the integrity of the data chain.

[0148] In the time representation format (YYYY-MM-DDTHH:MM:SS.sssZ), each character segment corresponds to a time element defined by the International Organization for Standardization (ISO 8601). "YYYY" represents a four-digit Gregorian calendar year, such as 2025; "MM" represents a two-digit month, ranging from 01 (January) to 12 (December); "DD" corresponds to a two-digit date, indicating the day of the month, valid from 01 to 31; and the letter "T" is a standard separator used to mark the boundary between the date and time portions, ensuring the clarity of the format parsing. Subsequently, "HH" represents the hour in 24-hour format, ranging from 00 to 23; "MM" represents the minute, ranging from 00 to 59; "SS" represents the second, also between 00 and 59; ".sss" is the millisecond portion, consisting of three digits, representing the precise millisecond value within the current second, for example, "123" corresponds to 0.123 seconds; finally, "Z" is the zero time zone identifier (Zulu time), indicating that the timestamp uses Coordinated Universal Time (UTC), a standard time without time zone offset, ensuring the consistency and comparability of time records globally. This complete format can accurately and unambiguously describe the instant an event occurs, suitable for log recording, data synchronization, and event tracing in cross-regional, distributed systems.

[0149] The structured information is serialized in JSON (JavaScript Object Notation) format, which has good readability and cross-platform compatibility. It can be persistently stored through the local file system, or efficiently transmitted to the server through RESTful API interfaces or message middleware such as RabbitMQ and Kafka. This allows for seamless integration with traffic management platforms, dispatch centers or big data platforms, supporting collaborative operation and data sharing among multiple systems.

[0150] To further illustrate the effectiveness of the drone-based license plate recognition method provided in this application embodiment, the following section will further explain the effectiveness of the drone-based license plate recognition method provided in this application embodiment by combining several different license plate recognition methods in related technologies.

[0151] Specifically, Figure 4 This diagram illustrates the performance comparison of different license plate recognition methods from a drone's perspective. Table 1 shows the experimental results comparing the performance of different license plate recognition methods from a drone's perspective. Please refer to Table 1 and... Figure 4 .

[0152] Table 1

[0153] As can be seen, the drone-based license plate recognition method provided in this application, compared with other methods, achieves an accuracy of 96.3%, a recall rate of 94.7%, and an F1-score of 95.5%, with all indicators significantly outperforming the comparative solutions. Although the processing latency is 320 milliseconds, slightly higher than some lightweight solutions, it is still within an acceptable range, especially considering that it can complete the entire closed-loop processing on edge computing devices without relying on external networks, thus possessing higher system autonomy and data security.

[0154] To perform the corresponding steps in the above embodiments and various possible methods, an implementation of a drone-based license plate recognition device is given below. Optionally, the drone-based license plate recognition device can adopt the above-described... Figure 2 The device structure of the edge computing unit is shown. Further, please refer to... Figure 5 , Figure 5 This is a functional block diagram of a drone-based license plate recognition device provided in this application embodiment. It should be noted that the basic principle and technical effects of the drone-based license plate recognition device provided in this embodiment are the same as those in the above embodiments. For the sake of brevity, any parts not mentioned in this embodiment can be referred to the corresponding content in the above embodiments. The drone-based license plate recognition device includes: a detection module and a processing module.

[0155] This detection module is used to acquire multiple ground images collected by the drone and perform vehicle detection on each ground image.

[0156] Understandably, this detection module can also be used to perform the above step S20.

[0157] The detection module is also used to determine the license plate detection algorithm for the ground image based on whether there is a vehicle to be analyzed in the ground image, and to perform license plate detection on the ground image through the license plate detection algorithm to obtain the license plate area in the ground image.

[0158] Understandably, this detection module can also be used to perform the above step S21.

[0159] This processing module is used to crop the ground image based on the license plate area to obtain the license plate image, and to perform multimodal processing on the license plate image to obtain multiple license plate images to be analyzed; wherein, each license plate image to be analyzed has a different image modality.

[0160] Understandably, this processing module can also be used to perform the above step S22.

[0161] The processing module is also used to perform text recognition on each license plate image to be analyzed, obtain the candidate license plate number corresponding to each license plate image to be analyzed, and determine the target license plate number of the vehicle to be analyzed from multiple candidate license plate numbers.

[0162] Understandably, this processing module can also be used to perform the above step S23.

[0163] Optionally, the detection module is further configured to: if the vehicle to be analyzed exists in the ground image, determine that the license plate detection algorithm is a local compensation detection algorithm, and perform license plate detection on the vehicle to be analyzed according to the local compensation detection algorithm to obtain the license plate region in the ground image; if the vehicle to be analyzed does not exist in the ground image, determine that the license plate detection algorithm is a global search detection algorithm, and perform license plate detection on the ground image according to the global search detection algorithm to obtain the license plate region in the ground image.

[0164] Optionally, the detection module is further configured to obtain the region coordinates of the vehicle to be analyzed in the ground image, extend the region coordinates according to a preset extension value to obtain the extended region coordinates of the vehicle to be analyzed in the ground image, extract the vehicle image corresponding to the vehicle to be analyzed from the ground image according to the extended region coordinates, input the vehicle image into a pre-trained license plate detection model for processing, and obtain the license plate region of the license plate of the vehicle to be analyzed in the ground image.

[0165] Optionally, the detection module is also used to input the ground image into a pre-trained license plate detection model for processing to obtain the license plate region in the ground image.

[0166] Optionally, the detection module is also used to acquire ground video collected by the UAV, perform video stream parsing on the ground video to obtain multiple initial ground images; perform denoising processing and color space correction on each initial ground image to obtain multiple ground images; and input each ground image into a pre-trained vehicle recognition model for processing to detect vehicles in each ground image.

[0167] Optionally, the processing module is further configured to determine the original image of the license plate as the first license plate image to be analyzed, and to perform grayscale processing on the license plate image to obtain the second license plate image to be analyzed; and to perform histogram equalization processing, adaptive threshold processing and CLAHE processing on the second license plate image to be analyzed to obtain the third, fourth and fifth license plate images to be analyzed.

[0168] Optionally, the processing module is further configured to perform text region detection and orientation classification on each license plate image to be analyzed, obtain the text region and rotation direction in the license plate image to be analyzed, and perform character recognition on the license plate image to be analyzed based on the text region and rotation direction in the license plate image to be analyzed, to obtain the candidate license plate number and analysis confidence level corresponding to the license plate image to be analyzed; and determine the license plate number with the highest analysis confidence level and exceeding a preset threshold among multiple candidate license plate numbers as the target license plate number of the vehicle to be analyzed.

[0169] Optionally, the processing module is also used to generate Chinese license plate labels by calling a pre-stored Chinese font file based on the target license plate number, to annotate the ground image with the license plate labels based on the Chinese license plate labels, and to send the annotated ground image to the server.

[0170] Optionally, the above modules can be stored in the form of software or firmware. Figure 2 The memory shown is either stored in or embedded in the operating system (OS) of the edge computing unit, and can be used by... Figure 2 The processor executes the commands. Meanwhile, the data and program code required to execute these modules can be stored in memory.

[0171] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, can implement the drone-based license plate recognition method provided in this application.

[0172] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0173] In addition, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0174] If a function is implemented as a software module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0175] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for license plate recognition on a drone, characterized in that, The method, applied to an edge computing unit of a drone, includes: Multiple ground images collected by the drone are acquired, and vehicle detection is performed on each of the ground images; Based on whether the vehicle to be analyzed exists in the ground image, a license plate detection algorithm is determined for the ground image. The license plate detection algorithm is then used to detect the license plate in the ground image to obtain the license plate region in the ground image. The ground image is cropped based on the license plate area to obtain a license plate image. The license plate image is then subjected to multimodal processing to obtain multiple license plate images to be analyzed. Each license plate image to be analyzed has a different image modality. Text recognition is performed on each of the license plate images to be analyzed to obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and the target license plate number of the vehicle to be analyzed is determined from the multiple candidate license plate numbers.

2. The method according to claim 1, characterized in that, The step of determining a license plate detection algorithm for the ground image based on whether a vehicle to be analyzed exists in the ground image, and performing license plate detection on the ground image using the license plate detection algorithm to obtain the license plate region in the ground image includes: If a vehicle to be analyzed is present in the ground image, the license plate detection algorithm is determined to be a local compensation detection algorithm. The license plate of the vehicle to be analyzed is detected according to the local compensation detection algorithm to obtain the license plate region in the ground image. If the vehicle to be analyzed is not found in the ground image, the license plate detection algorithm is determined to be a global search detection algorithm. The license plate is then detected in the ground image according to the global search detection algorithm to obtain the license plate region in the ground image.

3. The method according to claim 2, characterized in that, The step of detecting the license plate of the vehicle to be analyzed according to the local compensation detection algorithm to obtain the license plate region in the ground image includes: The region coordinates of the vehicle to be analyzed in the ground image are obtained, and the region coordinates are expanded according to a preset expansion value to obtain the expanded region coordinates of the vehicle to be analyzed in the ground image. Based on the coordinates of the extended region, the vehicle image corresponding to the vehicle to be analyzed is extracted from the ground image. The vehicle image is then input into a pre-trained license plate detection model for processing to obtain the license plate region of the vehicle to be analyzed in the ground image. The step of detecting license plates in the ground image according to the global search detection algorithm to obtain the license plate region in the ground image includes: The ground image is input into a pre-trained license plate detection model for processing to obtain the license plate region in the ground image.

4. The method according to claim 1, characterized in that, The process of acquiring multiple ground images collected by the drone and performing vehicle detection on each of the ground images includes: The ground video collected by the UAV is acquired, and the ground video is parsed to obtain multiple initial ground images; Each of the initial ground images is subjected to denoising and color space correction to obtain multiple ground images; Each of the ground images is input into a pre-trained vehicle recognition model for processing to detect vehicles in each ground image.

5. The method according to claim 1, characterized in that, The process of performing multimodal processing on the license plate image to obtain multiple license plate images to be analyzed includes: The original image of the license plate is determined as the first license plate image to be analyzed, and the license plate image is processed into grayscale to obtain the second license plate image to be analyzed. Histogram equalization, adaptive thresholding, and CLAHE processing are performed on the second license plate image to be analyzed to obtain the third, fourth, and fifth license plate images to be analyzed.

6. The method according to claim 1, characterized in that, The step of performing text recognition on each of the license plate images to be analyzed to obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and determining the target license plate number of the vehicle to be analyzed from the multiple candidate license plate numbers, includes: Each of the license plate images to be analyzed is subjected to text region detection and orientation classification to obtain the text region and rotation direction in the license plate image to be analyzed. Based on the text region and rotation direction in the license plate image to be analyzed, text recognition is performed on the license plate image to obtain the candidate license plate number and analysis confidence level corresponding to the license plate image to be analyzed. The license plate number with the highest analysis confidence and exceeding a preset threshold among the multiple candidate license plate numbers is determined as the target license plate number of the vehicle to be analyzed; The method further includes: Based on the target license plate number, a pre-stored Chinese font file is called to generate a Chinese license plate label. The ground image is then labeled with the Chinese license plate label, and the labeled ground image is sent to the server.

7. A license plate recognition device for unmanned aerial vehicles (UAVs), characterized in that, An edge computing unit for use in drones, the device comprising: The detection module is used to acquire multiple ground images collected by the UAV and perform vehicle detection on each of the ground images. The detection module is further configured to determine a license plate detection algorithm for the ground image based on whether there is a vehicle to be analyzed in the ground image, and to perform license plate detection on the ground image through the license plate detection algorithm to obtain the license plate region in the ground image; The processing module is used to crop the ground image according to the license plate area to obtain a license plate image, and to perform multimodal processing on the license plate image to obtain multiple license plate images to be analyzed; wherein, each license plate image to be analyzed has a different image modality; The processing module is further configured to perform text recognition on each of the license plate images to be analyzed, obtain candidate license plate numbers corresponding to each license plate image to be analyzed, and determine the target license plate number of the vehicle to be analyzed from the multiple candidate license plate numbers.

8. An edge computing unit, characterized in that, It includes a processor and a memory, the memory storing a computer program executable by the processor, the processor being able to execute the computer program to implement the method of any one of claims 1-6.

9. A drone, characterized in that, Includes the edge computing unit as described in claim 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-6.

Images