Method and device for identifying unmanned aerial vehicle based on motion features of high-point camera holder

Abstract

The invention relates to the technical field of unmanned aerial vehicles, and discloses a method and device for identifying an unmanned aerial vehicle based on high-point camera holder motion characteristics, and the device comprises a high-point installation block and an unmanned aerial vehicle body, and is characterized in that the top of the high-point installation block is provided with a high-point holder camera; a real-time motion trail tracking camera and an unmanned aerial vehicle recognition camera are installed on one side face of the high-point installation block, a storage module is installed in the high-point installation block, and a GPS locator is arranged at the top of the unmanned aerial vehicle body. According to the method and the device for identifying the unmanned aerial vehicle based on the motion characteristics of the high-point camera pan-tilt, the identification anti-interference capability of the unmanned aerial vehicle can be improved, and various interference factors, such as birds, balloons, kite and other objects in the air, illumination changes, weather changes and the like, exist in a low-altitude environment. According to the method and the device, the unmanned aerial vehicle and other interference objects can be distinguished through accurate analysis of motion features, and the recognition accuracy and reliability are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicles, and in particular to a method and a device for identifying unmanned aerial vehicles based on motion characteristics of a high-point camera gimbal. Background Art

[0002] With the rapid development of drone technology, drones are increasingly used in various fields, including aerial photography, logistics distribution, agricultural plant protection, environmental monitoring, etc. However, the widespread use of drones has also brought about a series of safety hazards and management issues.

[0003] On the one hand, unauthorized drones may infringe on personal privacy, interfere with aviation safety, and threaten the safety of important facilities. For example, drones appearing near airports may interfere with the normal takeoff and landing of aircraft, causing serious aviation accidents; drones appearing in sensitive areas such as important government agencies and military facilities may conduct illegal reconnaissance activities and leak state secrets.

[0004] On the other hand, effective management and monitoring of drones also face huge challenges. At present, common drone monitoring methods include radar detection, radio signal monitoring, and traditional photoelectric monitoring. Radar detection is costly and susceptible to environmental interference, and has limited detection effects on drones made of special materials; radio signal monitoring has problems such as limited monitoring types and susceptibility to signal interference; traditional photoelectric monitoring is susceptible to interference from drone-like targets such as birds and clouds; these traditional monitoring methods often make it difficult to achieve real-time and accurate identification of drones, and cannot meet the needs of modern society for drone safety management.

[0005] In this context, there is an urgent need for a new drone identification method and device that can achieve efficient and accurate monitoring of drones to ensure social security and the healthy development of the low-altitude economy. The method and device for identifying drones based on the motion characteristics of the high-point camera gimbal provide new ideas and solutions for solving the problem of drone monitoring, and lay a solid foundation for the integration of multiple detection devices to monitor drones. Summary of the invention

[0006] 1. Technical issues to be solved

[0007] In view of the deficiencies in the prior art, the present invention provides a method and device for identifying drones based on high-point camera gimbal motion features, which solves the above-mentioned problems.

[0008] (II) Technical solution

[0009] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a device for identifying drones based on the motion characteristics of a high-point camera pan-tilt, comprising a high-point mounting block and a drone body, characterized in that a high-point pan-tilt camera is installed on the top of the high-point mounting block, a real-time motion trajectory tracking camera and a drone identification camera are respectively installed on one side of the high-point mounting block, a storage module is installed inside the high-point mounting block, and a GPS locator is arranged on the top of the drone body.

[0010] Preferably, support legs and an image recognition camera are respectively provided at the bottom of the drone body, and the lowest position of the support legs is lower than the lowest position of the image recognition camera.

[0011] Preferably, the top of the high-point mounting block is in a horizontal state, and the high-point pan / tilt camera is in a parallel state with the top of the high-point mounting block.

[0012] Preferably, the high-point pan-tilt camera, the real-time motion trajectory tracking camera, the drone identification camera and the storage module are all electrically connected to an external power source via wires.

[0013] Another technical problem to be solved by the present invention is to provide a method for identifying drones based on high-point camera pan-tilt motion features, characterized in that it includes the following steps: high-point pan-tilt camera setting, model training and optimization, real-time detection and tracking, pan-tilt motion feature extraction and modeling, alarm rule engine and trajectory and video playback;

[0014] 1) High-point PTZ camera setting: Select a suitable communication tower high point to install the PTZ camera in the monitoring area (select different models of PTZ cameras according to monitoring requirements, with main parameters such as camera resolution, CMOS size, and ZOOM). During installation, ensure that the high-point PTZ installation surface remains horizontal. Configure the PTZ camera area scan according to the size of the monitored drone target and the monitored airspace. The PTZ camera continuously scans the monitored airspace.

[0015] 2) Image acquisition and annotation: To train an accurate and generalizable drone detection and recognition model, it is necessary to collect a large amount of drone image data of different types, postures, and environments. Data is collected through actual flight photography, collection of public data sets, or cooperation with third-party organizations. The data covers various common drone models (fixed-wing, rotary-wing, etc.), flight postures (such as horizontal flight, ascent, descent, hovering, etc.), different lighting conditions, background environments, etc. Preprocess and enhance the collected data, including random cropping, rotation, flipping, color transformation and other operations to increase data diversity and improve the generalization ability of the model; finely annotate image data and accurately annotate the collected image data, which is a key step in training the model. The annotation content should include information such as the type, location, size, and shape of the drone. A combination of manual annotation and automated annotation tools is used to improve efficiency;

[0016] 3) Model training and optimization: YOLO V8 is selected as the deep learning algorithm model training and detection framework based on the characteristics of drone targets and application scenarios. With its advanced backbone network, YOLO V8 is suitable for the extraction and detection of drone target features in small drone types and complex scenarios. Its optimized model architecture reduces the amount of calculation and parameters, and is suitable for real-time reasoning of large-scale video streams. The training and optimization of drone recognition models is divided into the following steps:

[0017] Dataset division: The dataset is divided into training set, validation set and test set in a ratio of 7:2:1;

[0018] Environment setup: Install the PyTorch deep learning framework and the YOLO V8 detection framework, and configure the GPU training environment;

[0019] Parameter setting: set the hyperparameters of training, including learning rate, batch size, number of training rounds, etc.

[0020] Model training: Choose to use the YOLOV8x pre-trained model for initialization, and input the training set data into the model for training. During the training process, the model will continuously adjust the parameters to minimize the loss function. Regularly use the validation set to evaluate the performance of the model so that the hyperparameters can be adjusted in time and the training progress can be monitored;

[0021] Model preservation and testing: The model with the best performance on the validation set is considered the final training result. The trained model is evaluated on the test set for final performance, including accuracy, recall, F1 value, and mean average precision (mAP).

[0022] 4) Real-time detection: The real-time inference engine is connected through the RTSP (Real Time Streaming Protocol) protocol or the GB / T28181 protocol. The inference engine parses the video stream and performs calculations and analysis on each frame of the image, extracts features and matches them with the feature model of the drone. When a drone target is detected, the drone type, target confidence, and the pixel coordinate bounding box (Bounding Box) representation of the target in the image are output (the pixel coordinates of the upper left and lower right corners of the target rectangular bounding box are used to determine the position and size of the target, usually expressed as (x1, y1, x2, y2), where (x1, y1) are the coordinates of the upper left corner and (x2, y2) are the coordinates of the lower right corner);

[0023] 5) Real-time tracking: through coordinate conversion, PTZ motion control and magnification control functions;

[0024] Coordinate conversion: Convert the pixel coordinates of the target in the image to the actual angle coordinates of the gimbal, and obtain the PTZ parameters (Pan, Tilt, Zoom) of the gimbal camera in real time through the camera SDK, combined with the camera hardware parameters such as focal length, image sensor size, etc. The pixel coordinates are converted into the horizontal and vertical angles (Pan, Tilt) of the gimbal through conversion;

[0025] Gimbal motion control: Based on the difference between the angle coordinates of the target pixel after conversion and the angle coordinates of the camera center point currently acquired, the camera SDK gimbal control system sends instructions to the gimbal motor to rotate the gimbal toward the direction of the drone target. Combined with the results of the target detection algorithm, the target's position information is continuously updated to ensure that the gimbal is always aimed at the drone target;

[0026] Magnification control: To achieve stable and continuous tracking, adjust the camera magnification according to the proportion of the detection frame in the image, ensure that the drone target remains at an appropriate size in the video screen when the drone moves quickly (away from or close to the camera), and meet the requirements of continuous detection and tracking;

[0027] 6) Analysis of motion characteristics of various targets: UAV targets account for a small proportion of pixels and have few features in optoelectronic devices. The use scenarios of UAVs are complex. There are many UAV-like targets in low-altitude scenarios, which are mainly divided into three categories: one is fast-moving objects such as birds, the second is slow-moving objects such as clouds, kites, balloons and other airborne objects, and the third is stationary objects such as communication tower platforms, power tower top structures, treetops, and building edges. The small size and few features of the targets coupled with the complex detection scenarios make false detections difficult to avoid. Since the various targets have obvious differences in motion characteristics, the UAV motion characteristics have fast flight speeds and smooth trajectories, the bird motion characteristics have fast speed changes and large trajectory fluctuations, and the slow-moving objects have slow speeds and smooth trajectories. By analyzing the motion characteristics of various targets, the target types can be distinguished to reduce false alarms;

[0028] UAV target motion characteristics: From the perspective of behavior patterns, UAVs are typical man-made targets, and their flight is essentially to complete various designated operational purposes, such as surveillance, reconnaissance, cargo transportation, attack, etc. Therefore, the behavior pattern of UAVs is a comprehensive reflection of the driver's operating habits and operating mode. Although UAVs are currently developing in the direction of intelligent and autonomous operation, their behavior patterns are essentially still related to the operational purpose and artificial intelligence algorithms, and there is a certain degree of certainty and predictability.

[0029] Bird target movement characteristics: Birds are non-man-made targets with self-awareness and intelligence. The behavior patterns of birds are usually closely related to their living habits. For example, there are significant differences between migratory birds and local resident birds. The flight maneuvering patterns of birds of the same species when foraging, returning to the nest, avoiding natural enemies, etc., and the differences in body size, living habits, and maneuverability of different bird species make their movement patterns more varied, and the corresponding target track characteristics are also more diverse;

[0030] Slow-moving target motion characteristics: In low-altitude scenes, slow-moving objects mainly include kites, balloons, airships, hot air balloons, etc. This type of object has a low speed value, stable speed, consistent direction, and smooth and continuous trajectory. The biggest difference between the slow-moving target motion characteristics and the drone motion characteristics is that the movement speed is slow;

[0031] 7) Pan, Tilt, Zoom acquisition: The real-time inference engine outputs the detection results. After issuing the gimbal control command, the tracking module obtains the Pan, Tilt, and Zoom values ​​every second through the camera SDK, and finally forms a Pan, Tilt, and Zoom array with a timestamp;

[0032] 8) PTZ motion feature extraction: Low-altitude moving objects have three-dimensional motion features. The camera identifies and tracks the target, projects the three-dimensional coordinates into two-dimensional coordinates, and extracts features based on the acquired PTZ array to characterize the target's speed and trajectory characteristics;

[0033] ① Horizontal average speed: The horizontal average speed is defined as the P value movement degree of the gimbal per unit time;

[0034]

[0035] ②Vertical average speed: Horizontal average speed is defined as the T-value movement degree of the gimbal per unit time;

[0036]

[0037] The horizontal average speed and vertical average speed describe the average speed of the object per unit time expressed from the gimbal movement speed, and are the most basic descriptors of the target track motion characteristics;

[0038] ③ Oscillation factor: This descriptor analyzes the characteristics of each adjacent sampling point in the gimbal trajectory and realizes a quantitative description of the oscillation characteristics of the gimbal trajectory. By calculating the relative heading difference value τ between each adjacent sampling point in the gimbal trajectory h (k) = h k+1 -h k, when the relative difference value exceeds a certain threshold, it is counted as one oscillation. The oscillation factor characterizes the smoothness of the target object trajectory through the change in the direction of the gimbal motion trajectory, and the number of oscillations within a certain period of time is used to determine the type of the target object;

[0039] 9) Alarm rule engine: Detect drone targets based on high-point pan-tilt cameras. Due to the high speed of drone targets, small targets, few features, and complex backgrounds, the false detection rate of drones based on image detection methods such as single application of deep learning is high. It is necessary to introduce target trajectory feature analysis methods to exclude non-drone targets;

[0040] 10) Trajectory and video playback: After the alarm rule engine determines it as an alarm, the system automatically records the target discovery camera and motion PT value, and starts recording at the same time for event playback and tracing.

[0041] Compared with the prior art, the present invention provides a method and device for identifying drones based on high-point camera gimbal motion features, which has the following beneficial effects:

[0042] 1. The method and device for identifying drones based on the motion characteristics of the high-point camera gimbal can improve the anti-interference ability of drone identification. There are various interference factors in the low-altitude environment, such as birds, balloons, kites and other aerial objects, as well as changes in lighting and weather. The method and device can distinguish drones from other interfering objects through accurate analysis of motion characteristics, thereby improving the accuracy and reliability of identification.

[0043] 2. The method and device for identifying drones based on the motion characteristics of the high-point camera gimbal improve the generalization ability of drone identification types. The method and device can effectively identify drones of different types, sizes, flight speeds and altitudes. By analyzing the motion characteristics of drones, various types of drones can be identified and can adapt to different flight states of drones, such as flying, climbing, descending, etc.

[0044] 3. The method and device for identifying drones based on the motion features of the high-point camera gimbal can effectively reduce the training cost of the drone recognition model. The introduction of precise analysis of motion features improves the anti-interference ability of drone recognition and the generalization ability of drone recognition types, thereby reducing the dependence on the accuracy of the recognition model and avoiding large amounts of data collection and labeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 It is a schematic diagram of the process structure of the present invention;

[0046] Figure 2 This is a schematic diagram of the flight mechanism of the rotary-wing UAV of the present invention;

[0047] Figure 3 This is a schematic diagram of the horizontal plane projection of the flight track of a typical UAV of the present invention;

[0048] Figure 4 This is a schematic diagram of the bird flight mechanism of the present invention;

[0049] Figure 5 This is a schematic diagram of the horizontal plane projection of the bird flight track of the present invention;

[0050] Figure 6 It is a schematic diagram of the Pan, Tilt and Zoom array of the present invention;

[0051] Figure 7 This is a flow chart of the high-point camera gimbal motion feature recognition of drones of the present invention;

[0052] Figure 8 This is a three-dimensional structural schematic diagram of the high-point installation block of the present invention;

[0053] Fig. 9 It is a three-dimensional structural schematic diagram of the UAV of the present invention;

[0054] Fig.10 It is a schematic structural diagram of the front cross section of the present invention.

[0055] In the figure: 1. High-point mounting block; 2. UAV body; 3. High-point gimbal camera; 4. Real-time motion trajectory tracking camera; 5. UAV identification camera; 6. Storage module; 7. GPS locator; 8. Support legs; 9. Image recognition camera. DETAILED DESCRIPTION

[0056] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] Embodiment 1:

[0058] A device for identifying a drone based on a high-point camera gimbal motion feature comprises a high-point installation block 1 and a drone body 2. Support legs and an image recognition camera are respectively arranged at the bottom of the drone body 2. The lowest position of the support legs is lower than the lowest position of the image recognition camera. The top of the high-point installation block 1 is in a horizontal state. A high-point gimbal camera 3 is parallel to the top of the high-point installation block 1. The high-point gimbal camera 3 is installed on the top of the high-point installation block 1. The high-point gimbal camera 3, a real-time motion trajectory tracking camera 4, a drone identification camera 5 and a storage module 6 are all electrically connected to an external power supply through wires. The real-time motion trajectory tracking camera 4 and the drone identification camera 5 are respectively installed on one side of the high-point installation block 1. A storage module 6 is installed inside the high-point installation block 1. A GPS locator 7 is arranged on the top of the drone body 2.

[0059] A method for identifying drones based on high-point camera pan-tilt motion features, characterized by comprising the following steps: high-point pan-tilt camera setting, model training and optimization, real-time detection and tracking, pan-tilt motion feature extraction and modeling, alarm rule engine and trajectory and video playback;

[0060] 1) High-point PTZ camera setting: Select a suitable communication tower high point to install the PTZ camera in the monitoring area (select different models of PTZ cameras according to monitoring requirements, with main parameters such as camera resolution, CMOS size, and ZOOM). During installation, ensure that the high-point PTZ installation surface remains horizontal. Configure the PTZ camera area scan according to the size of the monitored drone target and the monitored airspace. The PTZ camera continuously scans the monitored airspace.

[0061] 2) Image acquisition and annotation: To train an accurate and generalizable drone detection and recognition model, it is necessary to collect a large amount of drone image data of different types, postures, and environments. Data is collected through actual flight photography, collection of public data sets, or cooperation with third-party organizations. The data covers various common drone models (fixed-wing, rotary-wing, etc.), flight postures (such as horizontal flight, ascent, descent, hovering, etc.), different lighting conditions, background environments, etc. Preprocess and enhance the collected data, including random cropping, rotation, flipping, color transformation and other operations to increase data diversity and improve the generalization ability of the model; finely annotate image data and accurately annotate the collected image data, which is a key step in training the model. The annotation content should include information such as the type, location, size, and shape of the drone. A combination of manual annotation and automated annotation tools is used to improve efficiency;

[0062] 3) Model training and optimization: YOLO V8 is selected as the deep learning algorithm model training and detection framework based on the characteristics of drone targets and application scenarios. With its advanced backbone network, YOLO V8 is suitable for the extraction and detection of drone target features in small drone types and complex scenarios. Its optimized model architecture reduces the amount of calculation and parameters, and is suitable for real-time reasoning of large-scale video streams. The training and optimization of drone recognition models is divided into the following steps:

[0063] Dataset division: The dataset is divided into training set, validation set and test set in a ratio of 7:2:1;

[0064] Environment setup: Install the PyTorch deep learning framework and the YOLO V8 detection framework, and configure the GPU training environment;

[0065] Parameter setting: set the hyperparameters of training, including learning rate, batch size, number of training rounds, etc.

[0066] Model training: Choose to use the YOLOV8x pre-trained model for initialization, and input the training set data into the model for training. During the training process, the model will continuously adjust the parameters to minimize the loss function. Regularly use the validation set to evaluate the performance of the model so that the hyperparameters can be adjusted in time and the training progress can be monitored;

[0067] Model preservation and testing: The model with the best performance on the validation set is considered the final training result. The trained model is evaluated on the test set for final performance, including accuracy, recall, F1 value, and mean average precision (mAP).

[0068] 4) Real-time detection: The real-time inference engine is connected through the RTSP (Real Time Streaming Protocol) protocol or the GB / T28181 protocol. The inference engine parses the video stream and performs calculations and analysis on each frame of the image, extracts features and matches them with the feature model of the drone. When a drone target is detected, the drone type, target confidence, and the pixel coordinate bounding box (Bounding Box) representation of the target in the image are output (the pixel coordinates of the upper left and lower right corners of the target rectangular bounding box are used to determine the position and size of the target, usually expressed as (x1, y1, x2, y2), where (x1, y1) are the coordinates of the upper left corner and (x2, y2) are the coordinates of the lower right corner);

[0069] 5) Real-time tracking: through coordinate conversion, PTZ motion control and magnification control functions;

[0070] Coordinate conversion: Convert the pixel coordinates of the target in the image to the actual angle coordinates of the gimbal, and obtain the PTZ parameters (Pan, Tilt, Zoom) of the gimbal camera in real time through the camera SDK, combined with the camera hardware parameters such as focal length, image sensor size, etc. The pixel coordinates are converted into the horizontal and vertical angles (Pan, Tilt) of the gimbal through conversion;

[0071] Gimbal motion control: Based on the difference between the angle coordinates of the target pixel after conversion and the angle coordinates of the camera center point currently acquired, the camera SDK gimbal control system sends instructions to the gimbal motor to rotate the gimbal toward the direction of the drone target. Combined with the results of the target detection algorithm, the target's position information is continuously updated to ensure that the gimbal is always aimed at the drone target;

[0072] Magnification control: To achieve stable and continuous tracking, adjust the camera magnification according to the proportion of the detection frame in the image, ensure that the drone target remains at an appropriate size in the video screen when the drone moves quickly (away from or close to the camera), and meet the requirements of continuous detection and tracking;

[0073] 6) Analysis of motion characteristics of various targets: UAV targets account for a small proportion of pixels and have few features in optoelectronic devices. The use scenarios of UAVs are complex. There are many UAV-like targets in low-altitude scenarios, which are mainly divided into three categories: one is fast-moving objects such as birds, the second is slow-moving objects such as clouds, kites, balloons and other airborne objects, and the third is stationary objects such as communication tower platforms, power tower top structures, treetops, and building edges. The small size and few features of the targets coupled with the complex detection scenarios make false detections difficult to avoid. Since the various targets have obvious differences in motion characteristics, the UAV motion characteristics have fast flight speeds and smooth trajectories, the bird motion characteristics have fast speed changes and large trajectory fluctuations, and the slow-moving objects have slow speeds and smooth trajectories. By analyzing the motion characteristics of various targets, the target types can be distinguished to reduce false alarms;

[0074] UAV target motion characteristics: From the perspective of behavior patterns, UAVs are typical man-made targets, and their flight is essentially to complete various designated operational purposes, such as surveillance, reconnaissance, cargo transportation, attack, etc. Therefore, the behavior pattern of UAVs is a comprehensive reflection of the driver's operating habits and operating mode. Although UAVs are currently developing in the direction of intelligent and autonomous operation, their behavior patterns are essentially still related to the operational purpose and artificial intelligence algorithms, and there is a certain degree of certainty and predictability.

[0075] Bird target movement characteristics: Birds are non-man-made targets with self-awareness and intelligence. The behavior patterns of birds are usually closely related to their living habits. For example, there are significant differences between migratory birds and local resident birds. The flight maneuvering patterns of birds of the same species when foraging, returning to the nest, avoiding natural enemies, etc., and the differences in body size, living habits, and maneuverability of different bird species make their movement patterns more varied, and the corresponding target track characteristics are also more diverse;

[0076] Slow-moving target motion characteristics: In low-altitude scenes, slow-moving objects mainly include kites, balloons, airships, hot air balloons, etc. This type of object has a low speed value, stable speed, consistent direction, and smooth and continuous trajectory. The biggest difference between the slow-moving target motion characteristics and the drone motion characteristics is that the movement speed is slow;

[0077] 7) Pan, Tilt, Zoom acquisition: The real-time inference engine outputs the detection results. After issuing the gimbal control command, the tracking module obtains the Pan, Tilt, and Zoom values ​​every second through the camera SDK, and finally forms a Pan, Tilt, and Zoom array with a timestamp;

[0078] 8) PTZ motion feature extraction: Low-altitude moving objects have three-dimensional motion features. The camera identifies and tracks the target, projects the three-dimensional coordinates into two-dimensional coordinates, and extracts features based on the acquired PTZ array to characterize the target's speed and trajectory characteristics;

[0079] ① Horizontal average speed: The horizontal average speed is defined as the P value movement degree of the gimbal per unit time;

[0080]

[0081] ②Vertical average speed: Horizontal average speed is defined as the T-value movement degree of the gimbal per unit time;

[0082]

[0083] The horizontal average speed and vertical average speed describe the average speed of the object per unit time expressed from the gimbal movement speed, and are the most basic descriptors of the target track motion characteristics;

[0084] ③ Oscillation factor: This descriptor analyzes the characteristics of each adjacent sampling point in the gimbal trajectory and realizes a quantitative description of the oscillation characteristics of the gimbal trajectory. By calculating the relative heading difference value τ between each adjacent sampling point in the gimbal trajectory h (k) = h k+1 -h k, when the relative difference value exceeds a certain threshold, it is counted as one oscillation. The oscillation factor characterizes the smoothness of the target object trajectory through the change in the direction of the gimbal motion trajectory, and the number of oscillations within a certain period of time is used to determine the type of the target object;

[0085] 9) Alarm rule engine: Detect drone targets based on high-point pan-tilt cameras. Due to the high speed of drone targets, small targets, few features, and complex backgrounds, the false detection rate of drones based on image detection methods such as single application of deep learning is high. It is necessary to introduce target trajectory feature analysis methods to exclude non-drone targets;

[0086] 10) Trajectory and video playback: After the alarm rule engine determines it as an alarm, the system automatically records the target discovery camera and motion PT value, and starts recording at the same time for event playback and tracing.

[0087] The beneficial effects of the present invention are as follows: the method and device for identifying drones based on the motion characteristics of the high-point camera gimbal can improve the anti-interference ability of drone identification. There are various interference factors in the low-altitude environment, such as other aerial objects such as flying birds, balloons, kites, as well as changes in light and weather. The method and device can distinguish drones from other interfering objects through accurate analysis of motion characteristics, improve the accuracy and reliability of identification, and enhance the generalization ability of drone identification types. For drones of different types, sizes, flight speeds and altitudes, the method and device can effectively identify drones. By analyzing the motion characteristics of drones, various types of drones can be identified, and they can adapt to different flight states of drones, such as flying, climbing, descending, etc., which can effectively reduce the training cost of drone identification models. The introduction of accurate analysis of motion characteristics improves the anti-interference ability of drone identification and the generalization ability of drone identification types, thereby reducing the reliance on the accuracy of the identification model and avoiding large amounts of data collection and labeling.

[0088] 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.

Claims

1. A device for identifying drones based on the motion characteristics of a high-point camera gimbal, comprising a high-point mounting block and a drone body, characterized in that: A high-point pan-tilt camera is installed on the top of the high-point mounting block, a real-time motion trajectory tracking camera and a drone identification camera are installed on one side of the high-point mounting block, a storage module is installed inside the high-point mounting block, and a GPS locator is set on the top of the drone body.

2. The device for identifying drones based on high-point camera gimbal motion features according to claim 1, characterized in that: Support legs and an image recognition camera are respectively arranged at the bottom of the drone body, and the lowest position of the support legs is lower than the lowest position of the image recognition camera.

3. The device for identifying drones based on high-point camera gimbal motion features according to claim 1, characterized in that: The top of the high-point mounting block is in a horizontal state, and the high-point pan / tilt camera is in a parallel state with the top of the high-point mounting block.

4. The device for identifying drones based on high-point camera gimbal motion features according to claim 1, characterized in that: The high-point pan-tilt camera, the real-time motion trajectory tracking camera, the drone identification camera and the storage module are all electrically connected to an external power source through wires.

5. A method for identifying drones based on high-point camera gimbal motion features, characterized in that: The following steps are involved: High-point PTZ camera setup, model training and optimization, real-time detection and tracking, PTZ motion feature extraction and modeling, alarm rule engine, and trajectory and video playback; 1) High-point PTZ camera setting: Select a suitable communication tower high point to install the PTZ camera in the monitoring area (select different models of PTZ cameras according to monitoring requirements, with main parameters such as camera resolution, CMOS size, and ZOOM). During installation, ensure that the high-point PTZ installation surface remains horizontal. Configure the PTZ camera area scan according to the size of the monitored drone target and the monitored airspace. The PTZ camera continuously scans the monitored airspace. 2) Image acquisition and annotation: To train an accurate and generalizable drone detection and recognition model, it is necessary to collect a large amount of drone image data of different types, postures, and environments. Collect data through actual flight photography, collection of public data sets, or cooperation with third-party organizations. The data covers various common drone models (fixed-wing, rotary-wing, etc.), flight postures (such as horizontal flight, ascent, descent, hovering, etc.), different lighting conditions, background environments, etc. Preprocess and enhance the collected data, including random cropping, rotation, flipping, color transformation and other operations, to increase data diversity and improve the generalization ability of the model; Finely label image data. Accurately labeling the collected image data is a key step in training the model. The labeling content should include information such as the type, location, size, and shape of the drone. Combining manual labeling with automated labeling tools can improve efficiency; 3) Model training and optimization: YOLO V8 is selected as the deep learning algorithm model training and detection framework based on the characteristics of drone targets and application scenarios. With its advanced backbone network, YOLO V8 is suitable for the extraction and detection of drone target features in small drone types and complex scenarios. Its optimized model architecture reduces the amount of calculation and parameters, and is suitable for real-time reasoning of large-scale video streams. The training and optimization of drone recognition models is divided into the following steps: Dataset division: The dataset is divided into training set, validation set and test set in a ratio of 7:2:1; Environment setup: Install the PyTorch deep learning framework and the YOLO V8 detection framework, and configure the GPU training environment; Parameter setting: set the hyperparameters of training, including learning rate, batch size, number of training rounds, etc. Model training: Choose to use the YOLOV8x pre-trained model for initialization, and input the training set data into the model for training. During the training process, the model will continuously adjust the parameters to minimize the loss function. Regularly use the validation set to evaluate the performance of the model so that the hyperparameters can be adjusted in time and the training progress can be monitored; Model preservation and testing: The model with the best performance on the validation set is considered the final training result. The trained model is evaluated on the test set for final performance, including accuracy, recall, F1 value, and mean average precision (mAP). 4) Real-time detection: The real-time inference engine is connected through the RTSP (Real Time Streaming Protocol) protocol or the GB / T28181 protocol. The inference engine parses the video stream and performs calculations and analysis on each frame of the image, extracts features and matches them with the feature model of the drone. When a drone target is detected, the drone type, target confidence, and the pixel coordinate bounding box (Bounding Box) representation of the target in the image are output (the pixel coordinates of the upper left and lower right corners of the target rectangular bounding box are used to determine the position and size of the target, usually expressed as (x1, y1, x2, y2), where (x1, y1) are the coordinates of the upper left corner and (x2, y2) are the coordinates of the lower right corner); 5) Real-time tracking: through coordinate conversion, PTZ motion control and magnification control functions; Coordinate conversion: Convert the pixel coordinates of the target in the image to the actual angle coordinates of the gimbal, and obtain the PTZ parameters (Pan, Tilt, Zoom) of the gimbal camera in real time through the camera SDK, combined with the camera hardware parameters such as focal length, image sensor size, etc. The pixel coordinates are converted into the horizontal and vertical angles (Pan, Tilt) of the gimbal through conversion; Gimbal motion control: Based on the difference between the angle coordinates of the target pixel after conversion and the angle coordinates of the camera center point currently acquired, the camera SDK gimbal control system sends instructions to the gimbal motor to rotate the gimbal toward the direction of the drone target. Combined with the results of the target detection algorithm, the target's position information is continuously updated to ensure that the gimbal is always aimed at the drone target; Magnification control: To achieve stable and continuous tracking, adjust the camera magnification according to the proportion of the detection frame in the image, ensure that the drone target remains at an appropriate size in the video screen when the drone moves quickly (away from or close to the camera), and meet the requirements of continuous detection and tracking; 6) Analysis of motion characteristics of various targets: UAV targets account for a small proportion of pixels and have few features in optoelectronic devices. The usage scenarios of UAVs are complex. There are many UAV-like targets in low-altitude scenarios, which can be divided into three categories: the first category is fast-moving objects such as birds; the second category is slow-moving objects such as clouds, kites, balloons and other airborne objects; the third category is stationary objects such as communication tower platforms, power tower top structures, treetops, and building edges. Target Few small features and complex detection scenarios make false detection difficult to avoid. Since various targets have obvious differences in motion characteristics, drones have fast flight speeds and smooth trajectories, birds have fast speed changes and large trajectory fluctuations, and slow-moving objects have slow speeds and smooth trajectories. By analyzing the motion characteristics of various targets, target types can be distinguished to reduce false alarms. UAV target motion characteristics: From the perspective of behavior patterns, UAVs are typical man-made targets, and their flight is essentially to complete various designated operational purposes, such as surveillance, reconnaissance, cargo transportation, attack, etc. Therefore, the behavior pattern of UAVs is a comprehensive reflection of the driver's operating habits and operating mode. Although UAVs are currently developing in the direction of intelligent and autonomous operation, their behavior patterns are essentially still related to the operational purpose and artificial intelligence algorithms, and there is a certain degree of certainty and predictability. Bird target movement characteristics: Birds are non-man-made targets with self-awareness and intelligence. The behavior patterns of birds are usually closely related to their living habits. For example, there are significant differences between migratory birds and local resident birds. The flight maneuvering patterns of birds of the same species when foraging, returning to the nest, avoiding natural enemies, etc., and the differences in body size, living habits, and maneuverability of different bird species make their movement patterns more varied, and the corresponding target track characteristics are also more diverse; Slow-moving target motion characteristics: In low-altitude scenes, slow-moving objects mainly include kites, balloons, airships, hot air balloons, etc. This type of object has a low speed value, stable speed, consistent direction, and smooth and continuous trajectory. The biggest difference between the slow-moving target motion characteristics and the drone motion characteristics is that the movement speed is slow; 7) Pan, Tilt, Zoom acquisition: The real-time inference engine outputs the detection results. After issuing the gimbal control command, the tracking module obtains the Pan, Tilt, and Zoom values ​​every second through the camera SDK, and finally forms a Pan, Tilt, and Zoom array with a timestamp; 8) PTZ motion feature extraction: Low-altitude moving objects have three-dimensional motion features. The camera identifies and tracks the target, projects the three-dimensional coordinates into two-dimensional coordinates, and extracts features based on the acquired PTZ array to characterize the target's speed and trajectory characteristics; ① Horizontal average speed: The horizontal average speed is defined as the P value movement degree of the gimbal per unit time; ②Vertical average speed: Horizontal average speed is defined as the T-value movement degree of the gimbal per unit time; The horizontal average speed and vertical average speed describe the average speed of the object per unit time expressed from the gimbal movement speed, and are the most basic descriptors of the target track motion characteristics; ③ Oscillation factor: This descriptor analyzes the characteristics of each adjacent sampling point in the gimbal trajectory and realizes a quantitative description of the oscillation characteristics of the gimbal trajectory. By calculating the relative heading difference value τ between each adjacent sampling point in the gimbal trajectory h (k) = h k+1 -h k , when the relative difference value exceeds a certain threshold, it is counted as one oscillation. The oscillation factor characterizes the smoothness of the target object trajectory through the change in the direction of the gimbal motion trajectory, and the number of oscillations within a certain period of time is used to determine the type of the target object; 9) Alarm rule engine: Detect drone targets based on high-point pan-tilt cameras. Due to the high speed of drone targets, small targets, few features, and complex backgrounds, the false detection rate of drone identification based on image detection methods such as single application of deep learning is high. It is necessary to introduce target trajectory feature analysis methods to exclude non-drone targets; 10) Trajectory and video playback: After the alarm rule engine determines it as an alarm, the system automatically records the target discovery camera and motion PT value, and starts recording at the same time for event playback and tracing.

Images