Unmanned aerial vehicle monitoring system based on computer vision

Abstract

The invention discloses an unmanned aerial vehicle monitoring system based on computer vision. The unmanned aerial vehicle monitoring system comprises the following steps: S1, constructing an unmanned aerial vehicle platform; s2, computer vision hardware composition; s3, image acquisition and transmission; s4, computer vision algorithm composition; s5, data processing and storage; the invention belongs to the technical field of unmanned aerial vehicle monitoring, and particularly relates to an unmanned aerial vehicle monitoring system based on computer vision. The system has the advantages of high flexibility, wide coverage, excellent accuracy and reliability, super-strong real-time performance, quick response, efficient data management and decision support, and good user experience and collaboration.

Description

Technical Field

[0001] The present invention belongs to the technical field of unmanned aerial vehicle monitoring, and specifically refers to a unmanned aerial vehicle monitoring system based on computer vision. Background Art

[0002] In today's digital age, the application scenarios of monitoring technology are constantly expanding, from daily urban security to the operation and maintenance of industrial facilities, from dynamic monitoring of the ecological environment to the command and dispatch of emergency rescue operations, all of which put forward higher requirements on the performance of the monitoring system. Although traditional fixed monitoring equipment, such as common cameras, have played a monitoring role to a certain extent, they have many inherent defects. Its monitoring range is limited by the installation location. For large areas or places with complex terrain, such as mountainous areas, forests, and large construction sites, it is difficult to achieve full coverage, and there are a large number of monitoring blind spots. At the same time, in the face of moving targets, fixed monitoring equipment cannot flexibly adjust the viewing angle and tracking strategy, and it is difficult to meet the needs of dynamic monitoring.

[0003] The rapid development of drone technology has brought new opportunities to the field of monitoring. With its excellent maneuverability, flexible deployment capabilities and the advantage of being able to obtain a high-altitude bird's-eye view, drones can quickly reach designated areas to perform monitoring tasks. However, existing drone monitoring systems still face many challenges in practical applications. On the one hand, the quality of images collected by drones is easily affected by environmental factors, such as drastic changes in light intensity and adverse weather conditions (such as rain, snow, fog, etc.). These factors can cause images to be overexposed, underexposed, blurred, and have increased noise, which seriously affects the accuracy of target recognition. On the other hand, the efficiency and accuracy of existing computer vision algorithms when processing massive image data collected by drones are difficult to meet the requirements of real-time and accuracy. In complex scenarios, it is difficult to quickly and accurately identify and track targets, and the analysis of target behavior is also relatively limited, which cannot provide sufficient support for decision-making. Summary of the invention

[0004] The present invention is committed to building a drone monitoring system based on computer vision, aiming to overcome the shortcomings of existing technologies and achieve all-round, real-time and accurate monitoring of target areas. The system can accurately identify and track various targets in complex and changing environments, conduct in-depth analysis of their behaviors and provide timely warnings, providing solid and reliable technical support for the monitoring needs of various industries, and effectively improving monitoring efficiency and management level.

[0005] A computer vision-based drone monitoring system consists of the following components:

[0006] S1. UAV platform construction;

[0007] S2, computer vision hardware components;

[0008] S3, image acquisition and transmission;

[0009] S4, computer vision algorithm composition;

[0010] S5. Data processing and storage;

[0011] S6. User interaction interface.

[0012] Preferably, the specific method in S1, UAV platform construction is as follows:

[0013] S1-1. Choose high-performance, long-endurance drones designed for industrial-grade monitoring applications.

[0014] S1-2 is equipped with an integrated high-precision GPS / Beidou dual-mode positioning module, combined with multiple sensors such as inertial measurement unit (IMU), barometric altimeter and geomagnetic sensor, and uses advanced sensor fusion algorithms to achieve accurate perception and control of drone parameters such as attitude, position and speed.

[0015] S1-3 is equipped with 360-degree surround laser radar and ultrasonic sensors to perceive the surrounding environment in real time, plan a safe flight path in advance, and effectively avoid collision accidents.

[0016] Preferably, the UAV body in the construction of the S1 UAV platform is made of high-strength, lightweight carbon fiber composite material.

[0017] Preferably, the specific method in the S2, computer vision hardware composition is as follows:

[0018] S2-1. Install a customized high-resolution, low-light, wide dynamic range camera on the drone.

[0019] S2-2 is equipped with a dedicated image processor and uses NVIDIA's latest high-performance GPU chip.

[0020] Preferably, the camera resolution in the S2 computer vision hardware composition is ultra-high definition 12K, and the image processor has thousands of CUDA cores.

[0021] Preferably, the specific method in S3, image acquisition and transmission, is as follows:

[0022] S3-1. During the flight, the UAV collects images according to multiple preset strategies.

[0023] S3-2: The image data is first compressed and processed based on deep learning on the drone.

[0024] Preferably, an event-triggered intelligent acquisition mechanism is also introduced into the S3 image acquisition and transmission, and the compression processing method is a compression algorithm based on a generative adversarial network (GAN).

[0025] Preferably, the specific method in the composition of S4, computer vision algorithm is as follows:

[0026] S4-1. Target detection: An improved deep learning-based target detection algorithm is used in combination with an attention mechanism module. The attention mechanism model can focus more on the target area in the image and enhance the detection capability of small targets and targets under complex backgrounds. At the same time, multi-scale feature fusion technology is used to fuse feature maps of different scales, making full use of the contextual information of the image and improving the detection accuracy of targets of various sizes and shapes. During the training process, a large-scale and diverse drone image dataset is used, covering images of different seasons, weather, scenes, and various targets to enhance the generalization ability of the model. The target detection algorithm uses the YOLOv8 target detection model, and the attention mechanism model is the CBAM-Convolutional Block Attention Module.

[0027] S4-2, Target Tracking: A multi-target tracking algorithm based on multi-feature fusion and reinforcement learning is used. The appearance features, motion features and spatial position relationship features of the target are integrated to construct a comprehensive and accurate target descriptor. Using the reinforcement learning algorithm, based on the deep Q network (DQN) and the proximal policy optimization algorithm (PPO), the tracking strategy is dynamically adjusted according to the historical motion information of the target and the current environmental state. During the training process, various complex scenes are simulated, such as occlusion, cross motion, rapid motion, etc. of the target, so that the algorithm can learn the optimal tracking strategy under different circumstances. The appearance features are color histogram, texture features, and feature vectors extracted by deep learning. The motion features are speed, acceleration, and curvature of the motion trajectory.

[0028] S4-3. Behavior Analysis: Construct a behavior analysis model based on the combination of spatiotemporal graph convolutional network ST-GCN and long short-term memory network LSTM. ST-GCN is used to learn the movement patterns and relationships of targets in the spatial dimension, and to explore the movement patterns of target groups by modeling information such as the relative position, speed and direction between targets. LSTM focuses on capturing the behavior change trend of targets in time series, can handle long-term dependency problems, and memorize and analyze the historical behavior of targets.

[0029] Preferably, the specific method in S5, data processing and storage is as follows:

[0030] S5-1. The ground control center adopts a distributed cluster computing architecture, which consists of multiple high-performance servers. The high-performance servers are equipped with multi-core high-performance CPUs, large-capacity memory and high-speed solid-state hard drives, and high-speed data transmission within the cluster is achieved through high-speed network switches.

[0031] S5-2. Data storage uses a solution that combines the distributed file system Ceph with the distributed database TiDB. Ceph is used to store massive amounts of raw image data and has high reliability, high scalability, and high performance. Erasure coding technology and multi-copy strategies are used to ensure the security and fault tolerance of data during storage. TiDB is used to store analyzed structured data, such as the location of the target, behavior information, and event records.

[0032] Preferably, the specific method in S6, user interaction interface is as follows:

[0033] S6-1, using a user interface that supports desktop Windows, MacOS, Linux and mobile iOS, Android devices to display the UAV's flight status (including location, altitude, battery, flight speed, attitude, etc.), real-time video images, target detection and tracking results, and behavior analysis warning information in real time. Through high-definition map integration, the UAV's flight trajectory and the target's geographic location information are intuitively displayed.

[0034] S6-2. Users can perform various operations through the interface, such as remote control of drone takeoff, landing, hovering, route planning and adjustment. The route planning function supports manual drawing, importing GPX files, and automatically generating routes based on target areas. Manually mark and classify targets to facilitate focus and management of specific targets. Set warning rules and thresholds, and customize warning conditions according to different target types, behavior patterns and scenario requirements. Query historical monitoring data and visualize it in the form of charts (bar charts, line charts, pie charts, etc.), reports, etc., to facilitate users to analyze data and judge trends. In addition, it also supports multi-user collaborative operations. Through the authority management system, people from different departments and different authorities can jointly participate in monitoring and management work to achieve information sharing and collaborative decision-making.

[0035] The beneficial effects achieved by the present invention using the above structure are as follows:

[0036] 1. High flexibility and wide coverage: The maneuverability of drones enables them to quickly reach any designated area, whether it is a resource monitoring point in a remote mountainous area, every corner of a large construction site, or an ecological protection area on the coastline. It can achieve all-round monitoring, effectively make up for the coverage blind spots of traditional fixed monitoring equipment, and meet the monitoring needs in various complex environments.

[0037] 2. Excellent accuracy and reliability: The combination of advanced computer vision algorithms and high-performance hardware devices achieves high-precision recognition, stable tracking and accurate behavior analysis of targets. Even in complex environments, such as direct sunlight, low light, rain and snow, and frequent occlusion and rapid movement of targets, it can still maintain extremely high accuracy, provide reliable data support for decision-making, and significantly improve the credibility and application value of the monitoring system.

[0038] 3. Super real-time performance and rapid response: 5G communication technology and efficient data processing procedures ensure that monitoring data can be transmitted and analyzed in real time. The time delay from image acquisition to the release of warning information is controlled within 1 second, which greatly improves the emergency response speed, enables relevant personnel to obtain key information at the first time, take effective measures to deal with emergencies, and protect the safety of life and property and social stability.

[0039] 4. Efficient data management and decision support: Distributed data storage and processing architecture, combined with advanced big data analysis and machine learning technology, achieves efficient management and deep mining of massive monitoring data. Through the analysis of historical data and trend prediction, it can discover potential security risks and problems in advance, provide managers with forward-looking decision support, help realize intelligent and refined management, and improve management efficiency and scientific decision-making.

[0040] 5. Good user experience and collaboration: The cross-platform user interaction interface is simple and intuitive, easy to operate, and supports multi-user collaborative operation and authority management. Personnel from different departments and with different permissions can share monitoring information in real time through this interface, participate in monitoring management work together, achieve efficient information flow and collaborative decision-making, improve work efficiency and teamwork capabilities, and lay a good foundation for the widespread application and promotion of the monitoring system. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 Schematic diagram of a flow chart of an embodiment of the present invention. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0043] like Figure 1 As shown in the figure, a drone monitoring system based on computer vision consists of the following parts:

[0044] S1. UAV platform construction;

[0045] S2, computer vision hardware components;

[0046] S3, image acquisition and transmission;

[0047] S4, computer vision algorithm composition;

[0048] S5. Data processing and storage;

[0049] S6. User interaction interface.

[0050] like Figure 1 As shown in S1, the specific methods for building the drone platform are as follows:

[0051] S1-1. Choose high-performance, long-endurance drones designed for industrial-grade monitoring applications.

[0052] S1-2 is equipped with an integrated high-precision GPS / Beidou dual-mode positioning module, combined with multiple sensors such as inertial measurement unit (IMU), barometric altimeter and geomagnetic sensor, and uses advanced sensor fusion algorithms to achieve accurate perception and control of drone parameters such as attitude, position and speed.

[0053] S1-3 is equipped with 360-degree surround laser radar and ultrasonic sensors to perceive the surrounding environment in real time, plan a safe flight path in advance, and effectively avoid collision accidents.

[0054] The drone body used in the construction of the S1 drone platform is made of high-strength, lightweight carbon fiber composite materials, which reduces the weight of the fuselage and improves the endurance while ensuring the stability of the structure. By optimizing the aerodynamic design, the drone can still fly stably in a wind environment of level 6.

[0055] like Figure 1 As shown, the specific method in S2 and computer vision hardware composition is as follows:

[0056] S2-1. Install a customized high-resolution, low-light, wide dynamic range camera on the drone.

[0057] S2-2 is equipped with a dedicated image processor and uses NVIDIA's latest high-performance GPU chip.

[0058] The camera resolution of the S2 computer vision hardware is UHD 12K, which can capture subtle image details. In low-light environments, such as at night or in dimly lit areas, clear images can still be obtained. The wide dynamic range technology ensures that in scenes with strong light and shadows, the bright and dark details in the image can be clearly presented. In addition, the camera has optical image stabilization and autofocus functions to ensure that the images taken during the flight of the drone are stable and clear. The image processor has thousands of CUDA cores and has powerful parallel computing capabilities. With high-speed memory and dedicated image acceleration hardware, it can complete complex processing tasks for high-resolution images in milliseconds, including image denoising, enhancement, feature extraction, etc.

[0059] like Figure 1 As shown, the specific methods in S3, image acquisition and transmission are as follows:

[0060] S3-1. During the flight, the UAV collects images according to multiple preset strategies.

[0061] S3-2: The image data is first compressed and processed based on deep learning on the drone.

[0062] S3 also introduces an event-triggered intelligent acquisition mechanism in image acquisition and transmission. For example, when the thermal imaging sensor on the drone detects an abnormal heat source, or captures a specific abnormal sound through the sound sensor, it automatically triggers high-definition image acquisition. At the same time, the image is analyzed in real time using a deep learning algorithm. When a specific target (such as a person, vehicle, animal, etc.) is identified to enter a preset sensitive area, the acquisition function is immediately started. The compression processing method is a compression algorithm based on a generative adversarial network (GAN). On the premise of ensuring that the key information of the image is not lost, the data volume is compressed to less than 1 / 10 of the original size. Then, through the high-speed 5G communication module, the compressed image data is transmitted to the ground control center in real time at a transmission rate of up to 10Gbps. To ensure the reliability of data transmission, redundant transmission and error correction coding technology are used to effectively deal with signal interference in complex electromagnetic environments.

[0063] S4. The specific methods in the composition of computer vision algorithms are as follows:

[0064] S4-1. Target detection: Use the improved deep learning-based target detection algorithm in combination with the attention mechanism module. The attention mechanism model can focus more on the target area in the image and enhance the detection ability of small targets and targets in complex backgrounds. At the same time, multi-scale feature fusion technology is used to fuse feature maps of different scales, make full use of the context information of the image, and improve the detection accuracy of targets of various sizes and shapes. During the training process, a large-scale and diverse drone image dataset is used, covering images of different seasons, weather, scenes, and various types of targets to enhance the generalization ability of the model. The target detection algorithm uses the YOLOv8 target detection model, and the attention mechanism model is CBAM-Convolutional Block Attention Module. After optimized training, the model has a detection accuracy of more than 99% for common targets in complex scenes.

[0065] S4-2, Target Tracking: A multi-target tracking algorithm based on multi-feature fusion and reinforcement learning is adopted. The appearance features, motion features and spatial position relationship features of the target are integrated to construct a comprehensive and accurate target descriptor. Using the reinforcement learning algorithm, based on the deep Q network (DQN) and the proximal policy optimization algorithm (PPO), the tracking strategy is dynamically adjusted according to the historical motion information of the target and the current environmental state. During the training process, various complex scenes are simulated, such as occlusion, cross-motion, and rapid motion of the target, so that the algorithm can learn the optimal tracking strategy under different circumstances. The appearance features are color histograms, texture features, and feature vectors extracted by deep learning. The motion features are speed, acceleration, and curvature of the motion trajectory. In actual tests, the algorithm can still maintain a tracking accuracy of more than 95% in complex scenes with frequent occlusion and cross-motion of the target, and can quickly resume tracking of the target.

[0066] S4-3, Behavior Analysis: Construct a behavior analysis model based on the combination of spatiotemporal graph convolution network ST-GCN and long short-term memory network LSTM. ST-GCN is used to learn the movement patterns and relationships of targets in the spatial dimension, and to explore the movement patterns of target groups by modeling information such as the relative position, speed and direction between targets. LSTM focuses on capturing the behavior change trend of targets in time series, can handle long-term dependency problems, and memorize and analyze the historical behavior of targets. By learning a large number of normal and abnormal behavior samples, the model can accurately identify a variety of complex abnormal behaviors, such as abnormal gathering of people (more than 5 people gather in a specific area for a short time and behave abnormally), illegal parking of vehicles (staying in a prohibited parking area for more than 3 minutes without turning on the warning light), abnormal migration of animals (deviating from the regular migration route for more than a certain distance and with abnormal speed), etc. The accuracy rate of abnormal behavior detection reaches more than 98%. At the same time, combined with natural language processing technology, it can semantically describe abnormal behaviors and provide intuitive and accurate information for managers.

[0067] S5. The specific methods for data processing and storage are as follows:

[0068] S5-1. The ground control center adopts a distributed cluster computing architecture, which consists of multiple high-performance servers equipped with multi-core high-performance CPUs, large-capacity memory and high-speed solid-state hard drives. High-speed data transmission within the cluster is achieved through high-speed network switches. The cluster has a powerful computing ability to process millions of images per second. After receiving the image data transmitted by the drone, it first performs rapid decompression and preprocessing. Then, using parallel computing technology, the image data is input in parallel to multiple computer vision algorithm modules for analysis. Through the load balancing algorithm, the computing tasks are reasonably allocated to ensure the efficient operation of the system under high concurrency.

[0069] S5-2. Data storage uses a solution that combines the distributed file system Ceph with the distributed database TiDB. Ceph is used to store massive amounts of raw image data and has high reliability, high scalability, and high performance. Through erasure coding technology and multi-copy strategies, the security and fault tolerance of data during storage are ensured. TiDB is used to store analyzed structured data, such as the location, behavior information, and event records of the target. By establishing multidimensional indexes, such as time index, space index, target category index, behavior type index, etc., it can complete the query and retrieval of specific data within milliseconds. At the same time, using big data analysis and machine learning technologies, such as cluster analysis, association rule mining, deep learning prediction models, etc., historical data is deeply mined to predict potential security risks and event trends, providing managers with forward-looking decision support.

[0070] S6. The specific methods in the user interaction interface are as follows:

[0071] S6-1, using a user interface that supports desktop Windows, MacOS, Linux and mobile iOS, Android devices, real-time display of the drone's flight status (including location, altitude, power, flight speed, attitude, etc.), real-time video images, target detection and tracking results, and behavior analysis warning information. Through high-definition map integration, the drone's flight trajectory and the target's geographic location information are intuitively displayed. The interface adopts a simple, intuitive, beautiful and generous design style, and is optimized based on user experience design principles to ensure that users can quickly get started.

[0072] S6-2. Users can perform various operations through the interface, such as remote control of drone takeoff, landing, hovering, route planning and adjustment. The route planning function supports manual drawing, importing GPX files, and automatically generating routes based on target areas. Manually mark and classify targets to facilitate focus and management of specific targets. Set warning rules and thresholds, and customize warning conditions according to different target types, behavior patterns and scenario requirements. Query historical monitoring data and visualize it in the form of charts (bar charts, line charts, pie charts, etc.), reports, etc., to facilitate users to analyze data and judge trends. In addition, it also supports multi-user collaborative operations. Through the authority management system, people from different departments and different authorities can jointly participate in monitoring and management work to achieve information sharing and collaborative decision-making.

[0073] When used specifically, 1. System deployment: around the area that needs to be monitored, use professional signal detection equipment and geographic information analysis tools to select a location with open terrain, good communication signals and easy maintenance to set up a ground control center. According to the area, topography, monitoring focus, and flight performance and endurance of the monitored area, use professional route planning software, combined with actual needs and restrictions, to plan multiple efficient and reasonable flight routes. In the planning process, fully consider factors such as no-fly zones, restricted-fly zones, and obstacles to ensure the safety of drone flights. Assemble and debug drones and computer vision hardware equipment to ensure the normal operation of the equipment. At the same time, install data processing and storage equipment, as well as user interface software in the ground control center, and configure network connections and system parameters;

[0074] 2. System initialization: Before the system is started, the flight control system of the drone is fully calibrated, including compass calibration, accelerometer calibration, gyroscope calibration, etc., to ensure the accuracy of sensor data. Set the flight parameters of the drone, such as the maximum flight altitude (set to 100-500 meters according to different scenarios), flight speed (5-20 meters / second), hovering accuracy (±0.5 meters), etc. Initialize the computer vision algorithm and load the trained model weights. In the ground control center, initialize the database connection, set the data storage path, index strategy and data backup plan. At the same time, according to the actual needs of the user, configure the warning rules and notification methods (such as SMS, email, pop-up reminders, etc.) in the user interaction interface.

[0075] 3. Monitoring process: The user issues monitoring task instructions through the user interaction interface, and the drone automatically takes off according to the preset route and quickly cruises to the target monitoring area. During the flight, the drone collects high-definition image data in real time through the camera according to the preset collection strategy. After the image data is compressed, it is transmitted to the ground control center in real time through the 5G communication link. After receiving the data, the control center quickly decompresses and preprocesses it, and then inputs the image data into the computer vision algorithm module for analysis. The algorithm module quickly detects and tracks the target objects in the image and analyzes their behavior in real time. Once abnormal behavior is detected, the system immediately issues an early warning signal on the user interaction interface and informs relevant personnel through the preset notification method. Users can view the detailed information of abnormal events in real time on the interface, including the time, location, target type and behavior description, and remotely control the drone to adjust the flight perspective as needed to further track and evaluate abnormal events.

[0076] 4. System maintenance and updates: Regularly conduct comprehensive hardware inspections and maintenance on drones, including battery capacity testing and replacement (every 100 flights or 3 months of use), fuselage structure inspection and tightening, camera lens cleaning and calibration, sensor accuracy verification, etc. Monitor and optimize the performance of the server cluster at the ground control center, regularly clean up the system cache, update the operating system and software patches, and ensure the stable operation of the system. According to new monitoring requirements and scene changes, continue to collect new image data, and optimize and update computer vision algorithm models. For example, retrain the target detection model once a quarter, adding new target categories and scene data to improve the model's generalization ability and detection accuracy. At the same time, continuously improve the system's functions and user experience, and promptly fix software vulnerabilities and optimize operating procedures based on user feedback.

[0077] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0078] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A computer vision-based drone monitoring system consists of the following components: S1. UAV platform construction; S2, computer vision hardware components; S3, image acquisition and transmission; S4, computer vision algorithm composition; S5. Data processing and storage; S6. User interaction interface.

2. The computer vision-based drone monitoring system according to claim 1, characterized in that: The specific method of S1, UAV platform construction is as follows: S1-1. Choose high-performance, long-endurance drones designed for industrial-grade monitoring applications. S1-2 is equipped with an integrated high-precision GPS / Beidou dual-mode positioning module, combined with multiple sensors such as inertial measurement unit (IMU), barometric altimeter and geomagnetic sensor, and uses advanced sensor fusion algorithms to achieve accurate perception and control of drone parameters such as attitude, position and speed. S1-3 is equipped with 360-degree surround laser radar and ultrasonic sensors to perceive the surrounding environment in real time, plan a safe flight path in advance, and effectively avoid collision accidents.

3. The computer vision-based drone monitoring system according to claim 2, characterized in that: The UAV body in the construction of the S1 UAV platform is made of high-strength, lightweight carbon fiber composite materials.

4. The computer vision-based drone monitoring system according to claim 3, characterized in that: The specific method in S2, the computer vision hardware composition is as follows: S2-1. Install a customized high-resolution, low-light, wide dynamic range camera on the drone. S2-2 is equipped with a dedicated image processor and uses NVIDIA's latest high-performance GPU chip.

5. The computer vision-based drone monitoring system according to claim 4, characterized in that: The S2 computer vision hardware consists of a camera with an ultra-high-definition 12K resolution and an image processor with thousands of CUDA cores.

6. The computer vision-based drone monitoring system according to claim 5, characterized in that: The specific method in S3, image acquisition and transmission is as follows: S3-1. During the flight, the UAV collects images according to multiple preset strategies. S3-2: The image data is first compressed and processed based on deep learning on the drone.

7. The computer vision-based drone monitoring system according to claim 6, characterized in that: The S3 image acquisition and transmission also introduces an event-triggered intelligent acquisition mechanism, and the compression processing method is a compression algorithm based on a generative adversarial network (GAN).

8. The computer vision-based drone monitoring system according to claim 7, characterized in that: The specific method in the composition of S4, computer vision algorithm is as follows: S4-1. Target detection: An improved deep learning-based target detection algorithm is used in combination with an attention mechanism module. The attention mechanism model can focus more on the target area in the image and enhance the detection capability of small targets and targets under complex backgrounds. At the same time, multi-scale feature fusion technology is used to fuse feature maps of different scales, making full use of the contextual information of the image and improving the detection accuracy of targets of various sizes and shapes. During the training process, a large-scale and diverse drone image dataset is used, covering images of different seasons, weather, scenes, and various targets to enhance the generalization ability of the model. The target detection algorithm uses the YOLOv8 target detection model, and the attention mechanism model is the CBAM-Convolutional Block Attention Module. S4-2, Target Tracking: A multi-target tracking algorithm based on multi-feature fusion and reinforcement learning is used. The appearance features, motion features and spatial position relationship features of the target are integrated to construct a comprehensive and accurate target descriptor. Using the reinforcement learning algorithm, based on the deep Q network (DQN) and the proximal policy optimization algorithm (PPO), the tracking strategy is dynamically adjusted according to the historical motion information of the target and the current environmental state. During the training process, various complex scenes are simulated, such as occlusion, cross motion, rapid motion, etc. of the target, so that the algorithm can learn the optimal tracking strategy under different circumstances. The appearance features are color histogram, texture features, and feature vectors extracted by deep learning. The motion features are speed, acceleration, and curvature of the motion trajectory. S4-3. Behavior Analysis: Construct a behavior analysis model based on the combination of spatiotemporal graph convolutional network ST-GCN and long short-term memory network LSTM. ST-GCN is used to learn the movement patterns and relationships of targets in the spatial dimension, and to explore the movement patterns of target groups by modeling information such as the relative position, speed and direction between targets. LSTM focuses on capturing the behavior change trend of targets in time series, can handle long-term dependency problems, and memorize and analyze the historical behavior of targets.

9. The computer vision-based drone monitoring system according to claim 8, characterized in that: The specific method in S5, data processing and storage is as follows: S5-1. The ground control center adopts a distributed cluster computing architecture, which consists of multiple high-performance servers. The high-performance servers are equipped with multi-core high-performance CPUs, large-capacity memory and high-speed solid-state hard drives, and high-speed data transmission within the cluster is achieved through high-speed network switches. S5-2. Data storage uses a solution that combines the distributed file system Ceph with the distributed database TiDB. Ceph is used to store massive amounts of raw image data and has high reliability, high scalability, and high performance. Erasure coding technology and multi-copy strategies are used to ensure the security and fault tolerance of data during storage. TiDB is used to store analyzed structured data, such as the location of the target, behavior information, and event records.

10. The computer vision-based drone monitoring system according to claim 9, characterized in that: The specific method in S6, user interaction interface is as follows: S6-1, using a user interface that supports desktop Windows, MacOS, Linux and mobile iOS, Android devices to display the UAV's flight status (including location, altitude, battery, flight speed, attitude, etc.), real-time video images, target detection and tracking results, and behavior analysis warning information in real time. Through high-definition map integration, the UAV's flight trajectory and the target's geographic location information are intuitively displayed. S6-2. Users can perform various operations through the interface, such as remote control of drone takeoff, landing, hovering, route planning and adjustment. The route planning function supports manual drawing, importing GPX files, and automatically generating routes based on target areas. Manually mark and classify targets to facilitate focus and management of specific targets. Set warning rules and thresholds, and customize warning conditions according to different target types, behavior patterns and scenario requirements. Query historical monitoring data and visualize it in the form of charts (bar charts, line charts, pie charts, etc.), reports, etc., to facilitate users to analyze data and judge trends. In addition, it also supports multi-user collaborative operations. Through the authority management system, people from different departments and different authorities can jointly participate in monitoring and management work to achieve information sharing and collaborative decision-making.

Images