Unmanned aerial vehicle tracking method, tracking device, system and storage medium

Abstract

The invention relates to the field of unmanned aerial vehicle identification and tracking, and discloses an unmanned aerial vehicle tracking method, tracking device and system and a storage medium, and the method comprises the steps: searching an unmanned aerial vehicle through an unmanned aerial vehicle tracking device, and employing a deep neural network visual template detection mode when the unmanned aerial vehicle is not searched and is in an unmanned aerial vehicle searching state; searching a suspicious target in the view field through a deep neural network; wherein the unmanned aerial vehicle is used as a search target; after the target is searched, adopting an identification result to initialize a machine vision detection mode so as to track the target through machine vision; when a target is lost, a tracking result is inaccurate, a tracked target moves too fast or background interference is serious in the tracking process, the mode is switched to a mixed mode for mutual cooperation so as to realize accurate tracking of the target. The method can achieve the recognition and tracking of the unmanned plane under the edge calculation, is precise in recognition, is high in tracking precision, and is small in loss rate.

Description

Technical Field

[0001] The present invention relates to the technical field of unmanned aerial vehicle identification and tracking, and in particular to a method, a tracking device, a system and a storage medium for unmanned aerial vehicle tracking based on an improved YOLO algorithm and a KCF algorithm. Background Art

[0002] With the development of quality of life and drone technology, drones and other low-speed, slow and small aircraft are increasingly entering people's lives with extremely high use value and development prospects. In recent years, how to achieve the recognition and tracking of low-speed, slow and small aircraft in complex backgrounds has become a prominent focus. Drone recognition and tracking is a basic problem to solve problems such as drone laser power supply and drone countermeasures. At present, drone recognition and tracking often need to be combined with manual work, which is time-consuming, labor-intensive and inefficient. At present, there is no drone recognition and tracking system that is easy to operate and has high recognition and tracking accuracy. Summary of the invention

[0003] In view of the above problems, the purpose of the present invention is to provide a drone tracking method, tracking device, system and storage medium, which can realize the identification and tracking of drones under edge computing, and has accurate identification, high tracking accuracy and low loss rate.

[0004] To achieve the above-mentioned purpose, in the first aspect, the technical solution adopted by the present invention is: a drone tracking method, which comprises: searching for the drone through a drone tracking device, and when the drone is not found and is in a drone search state, using a deep neural network visual template detection mode to search for suspicious targets in the field of view through a deep neural network; wherein the drone is used as the search target; when the target is searched, the machine vision detection mode is initialized using the recognition result to track the target through machine vision, and the deep neural network visual template detection mode is exited; when the target is lost, the tracking result is inaccurate, the tracked target moves too fast or the background interference is serious during the tracking process, switching to a mixed mode for mutual cooperation to achieve accurate tracking of the target.

[0005] Furthermore, when no drone is found and the drone search state is in progress, the deep neural network visual template detection mode is used to search for suspicious targets in the field of view through the deep neural network, including:

[0006] The deep neural network uses improved YOLOv10;

[0007] Resize the pixels of each frame captured by the drone tracking device proportionally, convert the image color back to RGB color space, and convert the read image data into QImage format and pass it to Label for display;

[0008] Copy the image data to get the same image data, pass any image data to the improved YOLOv10 algorithm, obtain the video stream feedback, target confidence and IOU size after recognition, and display the recognition result in a new window;

[0009] The target confidence and IOU information are stored in an array for caching, and the identified target position, identification box size, etc. are extracted from the tensor through feedback and stored in the array to obtain the coordinate information of the target in the image;

[0010] Repeat the above steps to complete the target tracking and identification.

[0011] Further, after the target is searched, the recognition result is used to initialize the machine vision detection mode, so as to track the target through machine vision in the subsequent tracking process, including: the machine vision detection mode adopts the KCF+HOG detection mode;

[0012] The human-machine system delineates the range of the target to be tracked in the displayed image, and the machine vision algorithm extracts features from the target range area, captures the original feature information of the target, and describes the background features of the target area.

[0013] Mapping the original target features extracted from the first frame image into a high-dimensional space to distinguish different target areas; and using the original target features as training samples to train the relevant filters in the high-dimensional feature space;

[0014] By comparing adjacent frame images, the features of the target candidate area are extracted, and the similarity between the candidate area and the target is evaluated through the correlation filter trained with the original features of the target;

[0015] According to the similarity degree obtained by evaluation, the candidate area with the highest similarity in a frame of image is selected as the position of the target in this frame, and the position of the target in the previous frame is used as the tracking target in the next frame; repeat the above steps, continuously train the relevant filter, and evaluate the similarity degree;

[0016] The target location is stored in the cache, the target is continuously tracked and ensured not to exceed the image range.

[0017] Furthermore, the hybrid mode is: a mode in which the deep neural network visual template detection mode and the machine vision detection mode cooperate with each other;

[0018] Hybrid modes work together to achieve accurate tracking of targets, including:

[0019] The acquired image is transmitted to the improved YOLOv10 detection algorithm in the deep neural network visual template detection mode for target recognition, and the target location and recognition box size are obtained;

[0020] The target location and the size of the recognition frame are passed into the machine vision detection mode, and the KCF+HOG algorithm is used to extract the features of the target area, capture the appearance information of the target, and describe the features of the target area;

[0021] The deep vision algorithm re-detects the target location and recognition box size at set intervals, and compares them with the position tracked by the correlation filter obtained by training the original features of the target in the KCF+HOG algorithm;

[0022] The recognition results of the two algorithms are compared to obtain the confidence levels of tracking accuracy. The accuracy of the two algorithms is judged by the confidence levels to correct tracking deviations.

[0023] Furthermore, the accuracy of the two algorithms is judged by confidence level and the tracking deviation is corrected, including:

[0024] Improve the IoU size in the loss function of the YOLOv10 algorithm according to the coordinate calculation;

[0025] Use Old Position to store the first or last recognized target center position, calculate the moving speed of the tracked target in the horizontal and vertical directions, and judge it as an interference object when the moving speed and moving distance are higher than the set threshold;

[0026] Combine IoU and confidence to further determine the accuracy of the target recognition, calculate the flight speed and IoU value of the drone in the two recognition results, and convert the speed difference into a normalized value Divc to calculate the normalized value flag;

[0027] The flag value is used to determine the position deviation between this recognition and the previous recognition result. When the position deviation is greater than the threshold and the intersection-union ratio is less than the set threshold, the tracking deviation is corrected.

[0028] Further, correcting tracking bias includes:

[0029] When the fluctuation of the flag value is normal, there is no need to correct the tracking target;

[0030] When the flag value fluctuates beyond the set range, interference is considered to occur during the tracking process. At this time, the flag and the improved YOLOv10 algorithm are used to obtain the confidence conf, and the parameter Confidence is calculated based on this. Confidence is used to determine whether the algorithm recognition result is accurate. Based on the tracking result, it is comprehensively determined whether to re-transmit the new tracking target coordinates to the KCF algorithm for correction;

[0031] When interference or multiple targets gather, Divc and flag as conditions cannot accurately determine the correct target. The parameter Confidence is introduced to assist Divc and flag in making judgments. When the KCF tracking result needs to be corrected, the recognition result of the improved algorithm is passed into the KCF algorithm. When it is judged that the tracking result of a certain algorithm is wrong, another algorithm is used to correct it to ensure the accuracy of the tracking result.

[0032] In the second aspect, the technical solution adopted by the present invention is: a drone tracking device for implementing the above-mentioned drone tracking method, which includes: a bottom servo bracket, a control panel is arranged on the bottom servo bracket, and a program for executing the above-mentioned drone tracking method is arranged in the control panel; a single-axis bus servo servo, arranged below the bottom servo bracket, connected to the bottom servo bracket by bolts and nuts, and the output end of the single-axis bus servo servo passes through the top of the bottom servo bracket; a servo connecting bracket, connected to the output end of the single-axis servo bus servo through a connecting disk, and the servo connecting bracket is driven by the single-axis servo bus servo to rotate 360°; a servo bracket, the top of which is fixed to the bottom of the adapter plate, and one end of the bottom is movably connected to the servo connecting bracket; a bus servo servo, arranged on the servo connecting bracket by bolts and nuts, the output end of the bus servo servo is connected to the other end of the bottom of the servo bracket through the connecting disk, and the bus servo drives the servo bracket and the adapter plate to pitch and rotate; a camera and additional equipment are arranged at the center of the adapter plate, and are used to obtain drone target image information and transmit it to the control panel.

[0033] Further: the bus servo steering gear adopts a bus servo steering gear, and its control modes include: single-turn angle mode, multi-turn angle mode and rotation mode;

[0034] Single-turn angle control mode: means that the rotation angle of the bus servo is limited to the range of 0 to 360 degrees or -180 to +180 degrees;

[0035] Multi-turn angle control mode: allows the bus servo to rotate more than 360 degrees;

[0036] Wheel control mode: can keep the bus servo rotating without being restricted to a fixed angle range, can accurately specify the number of turns of the servo, and can adjust the rotation direction and speed.

[0037] In the third aspect, the technical solution adopted by the present invention is: a drone tracking system, comprising: a deep neural network search module, which searches for drones through a drone tracking device, and when the drone is not found and is in a drone search state, a deep neural network visual template detection mode is adopted to search for suspicious targets in the field of view through the deep neural network; wherein the drone is used as the search target; a machine vision detection module, when the target is searched, the machine vision detection mode is initialized using the recognition result to track the target through machine vision, and the deep neural network visual template detection mode is exited; a hybrid detection module, when the target is lost, the tracking result is inaccurate, the tracking target moves too fast or the background interference is serious during the tracking process, it switches to a hybrid mode to cooperate with each other to achieve accurate tracking of the target.

[0038] In a fourth aspect, the technical solution adopted by the present invention is: a computer-readable storage medium storing one or more programs, wherein the one or more programs include instructions, and when the instructions are executed by a computing device, the computing device executes any one of the above methods.

[0039] The present invention adopts the above technical solution, which has the following advantages:

[0040] 1. The present invention can automatically track and identify drones, facilitate observation and maintenance after deployment, have accurate identification, high tracking accuracy, low loss rate, and can replace and install equipment according to different needs to meet different usage needs and adapt to different environments.

[0041] 2. The present invention has the characteristics of high accuracy, high efficiency and high precision in identification and tracking.

[0042] 3. The present invention can realize multi-target recognition and tracking through the improved YOLOv10 algorithm.

[0043] 4. The present invention has strong scalability, and different devices can be replaced according to different usage scenarios.

[0044] 5. The present invention has strong adaptability and can track and identify UAVs in complex backgrounds, and has good robustness.

[0045] 6. The present invention has good flexibility and can detect and track different targets by replacing different data sets and models.

[0046] In summary, the present invention can be widely used in related fields such as drone identification and tracking, moving object detection, drone countermeasures, and drone tethering power supply. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 is an overall flow chart of the drone tracking method in an embodiment of the present invention;

[0048] Figure 2 is a flow chart of a method for identifying and tracking using a deep neural network visual target detection algorithm according to an embodiment of the present invention;

[0049] Figure 3 is a flow chart of a method for implementing tracking and positioning using a machine vision algorithm provided by an embodiment of the present invention;

[0050] Figure 4 It is a flow chart of a method for achieving tracking and positioning by the combined action of a deep neural network visual target detection algorithm and a machine vision algorithm provided by an embodiment of the present invention;

[0051] Figure 5 is a schematic diagram of the overall structure of the drone tracking device in an embodiment of the present invention;

[0052] Figure 6 is a front view of a drone tracking device according to an embodiment of the present invention;

[0053] Figure 7 is a side view of a drone tracking device according to an embodiment of the present invention;

[0054] Figure 8 It is a flow chart of a method for tracking a target and a zoom camera using a bus servo provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to solve the problems of time-consuming, labor-intensive and inefficient drone identification and tracking, the present invention provides a drone tracking method, tracking device, system and storage medium, which are easy to operate and have high identification and tracking accuracy. By improving the deep neural network visual target detection algorithm YOLOv10 and combining it with the computer vision KCF algorithm, the identification and tracking of drones under edge computing are realized, which effectively solves the identification accuracy and tracking loss problems of conventional algorithms under complex backgrounds.

[0056] In order to make the purpose, technical solution and advantages of the embodiment of the present invention clearer, the technical solution of the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiment of the present invention. Obviously, the described embodiment is a part of the embodiment of the present invention, not all of the embodiments. Based on the described embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field belong to the scope of protection of the present invention.

[0057] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0058] Embodiment 1, in this embodiment of the present invention, a method for tracking drones is provided. In this embodiment, the present invention can track multi-purpose and multi-field drones, including but not limited to drone identification and tracking, moving object detection, drone countermeasures, and drone tethering power supply. Figure 1 As shown, the drone tracking method includes the following steps:

[0059] 1) Searching for drones through drone tracking devices. When no drone is found and the drone is in the drone search state, a deep neural network visual template detection mode is used to search for suspicious targets in the field of view through a deep neural network; wherein the drone is used as the search target;

[0060] 2) When the target is found, the recognition result is used to initialize the machine vision detection mode so as to track the target through machine vision; at this time, the deep neural network visual template detection mode exits.

[0061] Specifically, the suspicious target is continuously tracked using machine vision. After the initialization is completed, the deep neural network detection is exited, and the suspicious target is re-detected at a preset interval for calibration.

[0062] 3) When the target is lost, the tracking result is inaccurate, the target moves too fast or the background interference is serious during the tracking process, switch to the mixed mode to cooperate with each other to achieve accurate tracking of the target.

[0063] In the above step 1), in the deep neural network visual template detection mode, the deep neural network uses the improved YOLOv10. Specifically, in this embodiment, when no drone is found and the drone is in the drone search state, the deep neural network visual template detection mode is used to search for suspicious targets in the field of view through the deep neural network, such as Figure 2 As shown, the following steps are included:

[0064] 1.1) Resize the pixels of each frame captured by the drone tracking device proportionally and convert the image color back to RGB color space for subsequent recognition and calculation; and convert the read image data into QImage format and pass it to Label for display, so as to facilitate observation of the original image.

[0065] In this embodiment, image capture is performed by a camera in the drone tracking device.

[0066] Reset the pixel size, specifically: scale the pixel size proportionally to 640*480.

[0067] 1.2) Copy the image data to obtain the same image data, pass any image data to the improved YOLOv10 algorithm, obtain the video stream feedback, target confidence, IOU (intersection over union) size after recognition, and display the recognition result in a new window.

[0068] In this embodiment, the image data is copied into new identical image data, one of which is processed by deep learning vision and the other is processed by machine vision.

[0069] The improved algorithm of YOLOv10 is: using gradient features, NWDloss improvement (NWD (Normalized Gaussian Wasserstein distance) is a new measurement method for target detection, especially suitable for small target detection) and Swin Transformer convolutional layer to improve YOLOv10.

[0070] 1.3) The target confidence and IOU information are stored in an array cache, and the identified target position, identification box size, etc. are extracted from the tensor through feedback and stored in the array to obtain the coordinate information of the target in the image.

[0071] After the YOLOv10 algorithm recognizes and processes the image, it will output the pixel coordinates of all suspicious points (XY coordinates of the upper left corner and lower right corner and the center point coordinates). The data returned after the algorithm successfully recognizes is in the form of a tensor (torch).

[0072] In this embodiment, when the target recognition result is less than the set threshold (for example, 5% of the horizontal and vertical coordinates) (pixel size), the camera can be controlled to zoom, the image can be enlarged to extract the clear appearance features of the object, and the image can be recognized again using the YOLOv10 improved algorithm.

[0073] 1.4) Repeat the above steps to complete the tracking and identification of the target, and then track it through the control bus servo in the drone tracking device.

[0074] In the above step 2), the machine vision detection mode adopts the KCF+HOG (kernelized correlation filter target tracking algorithm+histograph of directional gradients) detection mode.

[0075] Specifically, after the target is searched, the recognition result is used to initialize the machine vision detection mode so that the target can be tracked by machine vision in the subsequent tracking process, such as Figure 3As shown, the following steps are included:

[0076] 2.1) The human-machine system circles the scope of the tracking target in the displayed image (for example, the scope of the tracking target can be circled in the image window by touch or other means), and the machine vision algorithm extracts features of the target range area, captures the original feature information of the target, and describes the background features of the target area.

[0077] 2.2) Mapping the original target features extracted from the first frame image into a high-dimensional space to distinguish different target areas; and using the original target features as training samples to train the relevant filters in the high-dimensional feature space.

[0078] Correlation filtering is used to detect and identify the presence of known patterns or signals in a complex background. By calculating the correlation between the input signal and the known template, the correlation filter can highlight the characteristics of the target signal while suppressing noise and other interference.

[0079] 2.3) Entering the target tracking stage, by comparing adjacent frame images, the features of the required target candidate area are extracted, and the similarity between the candidate area and the target is evaluated through the correlation filter trained with the original features of the target.

[0080] 2.4) Based on the similarity obtained by the evaluation, select the candidate area with the highest similarity in a frame as the position of the target in this frame, and use the position of the target in the previous frame as the tracking target in the next frame; repeat the above steps, continuously train the relevant filters, and evaluate the similarity.

[0081] 2.5) Store the target location in the cache, control the camera zoom ratio to continuously track the target and ensure that the target does not exceed the image range. Adjust the servo position to align the additional device with the target, and determine whether the additional device is working properly based on the returned data.

[0082] In the above step 3), the hybrid mode is a mode in which the deep neural network visual template detection mode and the machine vision detection mode cooperate with each other.

[0083] Specifically, in this embodiment, the mixed modes cooperate with each other to achieve accurate tracking of the target, such as Figure 4 As shown, the following steps are included:

[0084] 3.1) The acquired image is transmitted to the improved YOLOv10 detection algorithm in the deep neural network visual template detection mode for target recognition, and the target location and recognition box size (i.e., range size) are obtained;

[0085] Specifically, similar to the deep neural network visual template detection mode in step 1), the following steps are included:

[0086] 3.1.1) First, scale the pixel size of each frame captured by the camera to 640*480, convert the image color back to RGB color space; convert the read image data into QImage format and copy the image data to obtain new data.

[0087] 3.1.2) The copied original image is passed into the improved YOLOv10 detection algorithm for recognition, and the target position, recognition box size, etc. obtained by recognition are extracted from the tensor through feedback and stored in the cache array to obtain the location and appearance features of the target.

[0088] 3.2) The target location information and recognition box size obtained by the improved YOLOv10 algorithm are passed into the machine vision detection mode, and the KCF+HOG algorithm is used to extract the features of the target area, capture the appearance information of the target, and describe the characteristics of the target area.

[0089] Specifically, the target position of the first frame and the copied image are passed into the KCF+HOG algorithm, which extracts features of the target area, captures the appearance information of the target, describes the features of the target area, and performs target recognition and tracking.

[0090] 3.3) The deep vision algorithm re-detects the target location and recognition box size at set intervals, and compares them with the position tracked by the correlation filter obtained by training the original features of the target in the KCF+HOG algorithm to determine which algorithm is more accurate.

[0091] 3.4) Compare the recognition results of the two algorithms and obtain the confidence level of tracking accuracy respectively; judge the accuracy of the two algorithms by the confidence level and correct the tracking deviation.

[0092] For example, when the confidence level is greater than the set value (the confidence level is different in different scenarios. For example, in a rainy background, the confidence level is set to 0.6) and the recognition difference between the two algorithms (intersection over union (IoU)) is less than the threshold, it can be determined that the target is occluded or tracking is lost during the recognition process, resulting in tracking errors in the machine vision algorithm.

[0093] In this embodiment, when it is determined that the tracking result of a certain algorithm is wrong, another algorithm can be used to correct the error to ensure the accuracy of the tracking result. Specifically, judging the accuracy of the two algorithms by confidence and correcting the tracking deviation includes the following steps:

[0094] 3.4.1) Improve the IoU size in the loss function of the YOLOv10 algorithm according to the coordinate calculation;

[0095] 3.4.2) Use Old Position to store the first or last identified target center position, calculate the moving speed of the tracked target in the horizontal and vertical directions, and when the moving speed and moving distance are higher than the set threshold, it can be judged as an interference object.

[0096] Because there are multiple objects to be detected in a certain frame of image, it is necessary to screen multiple targets based on the recognition results. Considering that the movement is continuous in space, the pixel movement of the same target in the image should be a finite value. Therefore, the judgment formula is designed as:

[0097]

[0098] Where x2 represents the center x-point coordinate of the newly identified target, x1 represents the stored x-point coordinate, and ImgsizeX represents the horizontal pixel size of the image. From this, the horizontal moving speed Xspeed of the tracked target can be obtained; the vertical moving speed can be obtained similarly.

[0099] In this embodiment, considering that the flying speed of the drone is different in different situations, the threshold setting is not fixed, and the threshold value can be automatically adjusted according to the speed range obtained by multiple detections.

[0100] 3.4.3) Combine IoU and confidence to further determine the accuracy of target recognition, calculate the flight speed and IoU value of the drone in the two recognition results, and convert the speed difference into a normalized value Divc to calculate the normalized value flag.

[0101] During the flight of the drone, due to the limitation of motor power, there is no situation where the speed increases dramatically instantly. Therefore, when the speed changes nonlinearly, it is considered that the target recognition is inaccurate, and further judgment can be made by combining IoU and confidence. After the KCF algorithm processes the next frame to identify the location of the object, the target coordinates in the KCF tracking process are compared with the target coordinates obtained by the YOLOv10 algorithm in the Detect part with the introduction of gradient features, and IoU is used to represent the coincidence relationship. The IoU intersection-over-union ratio calculation formula is:

[0102]

[0103] Where IntersectionArea represents the overlapping area of ​​two recognition boxes; UnionArea represents the sum of the areas of the two recognition boxes.

[0104] IoU can well describe the quality of an object detection model. IoU has the advantages of scale invariance, non-negative results and normalized value range.

[0105] In this embodiment, the calculation formulas for Divc, flag and confidence are shown in formula 3-5, where OldSpeed ​​and NewSpeed ​​respectively represent the image movement speed between two recognitions, speed threshold represents the speed threshold, and the setting parameter Confidence represents the credibility of this recognition.

[0106]

[0107] flag=(Divc×parameter+IoU×(1-parameter)) (4)

[0108]

[0109] The parameter in formula 4 represents the ratio of the two parameters. Different parameters are set in different scenarios to achieve the best tracking effect.

[0110] 3.4.4) The position deviation between the current recognition and the previous recognition result is determined by the flag value. When the position deviation is greater than the threshold and the intersection-over-union ratio is less than the set threshold, the tracking deviation is corrected;

[0111] Specifically, correcting tracking deviations includes the following steps:

[0112] 3.4.4.1) When the fluctuation of the flag value is normal, there is no need to correct the tracking target.

[0113] 3.4.4.2) When the flag value fluctuates beyond the set range, interference is considered to occur during the tracking process. At this time, the flag and the improved YOLOv10 algorithm are used to obtain the confidence conf, and the parameter Confidence is calculated based on this. Confidence is used to determine whether the algorithm recognition result is accurate. Based on the tracking results, it is comprehensively determined whether to re-transmit the new tracking target coordinates to the KCF algorithm for correction.

[0114] 3.4.43) When interference or multiple targets gather, Divc and flag as conditions cannot accurately determine the correct target. The introduction of the parameter Confidence can assist Divc and flag in making judgments. When the KCF tracking result needs to be corrected, the recognition result of the improved algorithm is passed to the KCF algorithm. When it is judged that the tracking result of a certain algorithm is wrong, another algorithm can be used to correct it to ensure the accuracy of the tracking result.

[0115] In summary, when the target identified in the image is larger or smaller than the set threshold, resulting in the inability to obtain clear and complete appearance features, the camera can be controlled to zoom to ensure that the target size is appropriate and to ensure clear and complete appearance features of the target in the image.

[0116] The present invention uses three different tracking and positioning methods: (1) Manually select the tracking target and use the visual algorithm to realize drone tracking. (2) Use the improved YOLOv10 algorithm to automatically lock the location of the drone and track it. (3) Combine the improved YOLOv10 algorithm with the KCF+HOG algorithm. After the YOLOv10 improved algorithm detects the location of the drone, the location information is passed to the KCF target tracking algorithm, and the target tracking algorithm continues to track it to save computing power.

[0117] Embodiment 2, in this embodiment of the present invention, a drone tracking device is provided, which is used to implement the drone tracking method in the above embodiment 1, and is a pan-tilt structure. In this embodiment, the present invention can track multi-purpose and multi-field drones, including but not limited to drone identification and tracking, moving object detection, drone countermeasures, and drone tethering power supply. Specifically, as Figures 5 to 7 As shown, the UAV tracking device includes: a servo bracket 1 at the bottom of the gimbal, a servo connecting bracket 2, a servo bracket 3, a single-axis bus servo servo 4, a bus servo servo 5, a connecting plate 6, a camera and additional equipment 7 and an adapter plate 8.

[0118] The bottom steering gear bracket 1 is used to support the components of the entire drone tracking device, and a control panel is arranged on the bottom steering gear bracket 1, and a program for executing the drone tracking method in the above-mentioned embodiment 1 is arranged in the control panel;

[0119] The single-axis bus servo steering engine 4 is arranged below the bottom steering engine bracket 1 and connected to the bottom steering engine bracket 1 through bolts and nuts, and the output end of the single-axis bus servo steering engine 4 passes through the top of the bottom steering engine bracket 1;

[0120] The steering gear connecting bracket 2 is connected to the output end of the single-axis servo bus steering gear 4 through the connecting plate 6, and the single-axis servo bus steering gear 4 drives the steering gear connecting bracket 2 to rotate 360°;

[0121] The top of the steering gear bracket 3 is fixed to the bottom of the adapter plate 8, and one end of the bottom is movably connected to the steering gear connecting bracket 2;

[0122] The bus servo servo 5 is arranged on the servo connecting bracket 2 by means of bolts and nuts. The output end of the bus servo servo 5 is connected to the other end of the bottom of the servo bracket 3 through the connecting plate 6. The connecting plate 6 and the servo bracket 3 are fixedly connected by screws. The bus servo servo 5 drives the servo bracket 3 and the adapter plate 8 to pitch and rotate.

[0123] The camera and additional equipment 7 are arranged at the center of the adapter plate 8, and are used to obtain the target image information of the UAV and transmit it to the control panel.

[0124] The servo bracket 3 is fixed to the adapter plate 8 by bolts and nuts, and the adapter plate 8 is connected to the camera and the additional equipment 7 by bolts and nuts. The power supply and information transmission between the single-axis bus servo 4 and the bus servo 5 are connected to the control board by a single bus.

[0125] In this embodiment, the camera and additional equipment 7 includes a camera module, a power module, an information transmission module, an external expansion module, an information display and human-computer interaction module, etc. The signal line of the camera and additional equipment 7 is directly connected to the control board through a wiring harness, and the power supply is provided by an external power module.

[0126] In this embodiment, in order to facilitate the removal of batteries or installation of power supplies, the power supply device is externally connected, and the power supply line is fixed to the servo bracket 1 at the bottom of the gimbal. Due to the weight of the control board and the battery, the overall center of gravity of the device moves downward, ensuring the stability of the device during tracking.

[0127] In the above embodiment, the programs of the drone tracking method in the control panel are: deep neural network visual target detection mode, machine vision detection mode, and deep vision and machine vision detection mixed mode.

[0128] In the above embodiment, the bus steering gear 5 adopts a bus servo steering gear, and its control modes include: single-turn angle mode, multi-turn angle mode and rotation mode.

[0129] Single-turn angle control mode: refers to the rotation angle of the bus servo 5 being limited to the range of 0 to 360 degrees (or -180 to +180 degrees). This mode has the advantages of a small range of motion and rapid response of the servo, and can achieve rapid rotation of the PTZ. This control mode is suitable for tracking targets at a long distance. Since the target at a long distance has a small range in the image, the target appearance features can be clear and complete after zooming the camera, but the target tracking process does not require a large rotation, only a small and high-precision rotation.

[0130] Multi-turn angle control mode: This mode allows the bus servo 5 to rotate more than 360 degrees, suitable for applications that require infinite rotation or a specific number of turns. It has continuous rotation without resetting, extremely precise position control and is suitable for motion scenarios that require complex path planning and precise repetition. This mode is suitable for close distances or when the target needs to pass overhead. It can track the target without the servo turning and without tracking dead zone.

[0131] Wheel control mode: This mode can keep the bus servo 5 rotating, without being limited to a fixed angle range like traditional servos, and can accurately specify the number of turns of the servo. And the rotation direction and speed can be adjusted. This mode can be used when searching without a target or when extremely fast rotation speed positioning is required. For example, if there is a suspicious target in the air, it can be set to a rotation mode. Through the rotation control mode, the camera shooting range can sweep across every area in the air to search for a specific suspicious target.

[0132] In the above embodiment, the bus servo 5 tracks the target and the camera zooms, such as Figure 8 As shown, the following steps are included:

[0133] (1) The position of the target in the image is obtained by three methods of drone tracking methods (deep neural network visual target detection mode, machine vision detection mode, and deep vision and machine vision detection mixed mode). After calculating the motion space of the gimbal (i.e., the adapter plate 8 and the servo connecting bracket 2) and the gimbal motion planning and target tracking, the angular speed and other information required for each servo to move are obtained.

[0134] The calculation of motion space and the motion planning of the gimbal are as follows: Assuming that the gimbal is installed at a fixed position, take this position as the origin of the coordinate system, and represent the motion space of the gimbal as a spherical area determined by the pitch angle and the yaw angle. The pitch angle range is 0° to 270°, indicating the up and down swing of the gimbal. The yaw angle range is different for each control mode. For example, in the single-turn angle mode, it is set to -180° to 180°. According to the different angles (pitch angle and yaw angle) calculated by the offset and the servo steering angle obtained by the query, the target angle is calculated; the target angle of the servo is decomposed into the pitch angle, and the optimal control route is determined according to the different pitch angles, that is, how to drive to track the target at the fastest speed.

[0135] (2) The control board sends the angle information required by each servo through the serial port, and the servo is controlled by PID to perform continuous tracking.

[0136] (3) By evaluating the size of the target frame during image tracking, the ratio parameter (the ratio of the target frame pixel size to the overall image pixel size) is obtained. When the ratio parameter is lower or higher than the set threshold, the camera zoom ratio is increased or decreased so that the target in the image becomes larger or smaller accordingly, ensuring the size of the target in the image and the integrity of the background information.

[0137] Specifically, in this embodiment, by presetting the K value (control scale factor) for different scenes in advance, the PID control algorithm is used to control the rotation direction and angle of the servo. After the target is successfully tracked using the above three methods to obtain the position coordinates, the coordinates are calculated, and the normalized offset is used in the process of calculating the image center point and identifying the target position. The offset is normalized and expressed as a decimal in the range of [-1,1]. The normalized offset is calculated by formula 6 (the image pixels are scaled to 640*480):

[0138] offsetx=2.0*(0.5-x / imgheight)

[0139] offsety=2.0*(0.5-y / imgheight) (6)

[0140] Wherein, offsety represents the normalized offset in the x direction; offsety represents the normalized offset in the y direction; imgwidth represents the width of the image pixel horizontal coordinate (640px); imgheight represents the height of the image pixel vertical coordinate (480px); (x, y) represents the horizontal and vertical coordinates of the target, respectively.

[0141] After the normalized offset is calculated, the angle angle required for the single-axis bus servo 4 and the bus servo 5 to achieve tracking is calculated using formula 7. After the required rotation angle is calculated, it is determined whether it exceeds the preset maximum servo rotation angle.

[0142] angle=k*offset (7)

[0143] Where k represents the control proportional coefficient; offset represents the normalized offset of the target.

[0144] The rotation angle is passed into the servo control function of the control board. The function calls the preset servo control parameters to realize the servo rotation. The servo control parameters are: servo number, servo speed (degrees / second), uniform acceleration time (ms), uniform deceleration time (ms). The flexible impact generated by the uniform acceleration and deceleration motion law can ensure the stability of the tracking state.

[0145] When the target identified in the image is larger or smaller than the threshold, resulting in the inability to obtain clear and complete appearance features, the control board can be used to send instructions to control the camera to zoom to ensure that the target size is appropriate and the target in the image is clear and complete.

[0146] In summary, the present invention has the following characteristics: 1. The improved YOLOv10 algorithm and the KCF+HOG algorithm are combined to solve the problems of target occlusion and target loss in traditional visual algorithms, with good recognition and tracking accuracy, and can quickly identify targets in more complex environments. 2. The edge computing development board is used for control, which has the characteristics of high computing power and fast response speed. 3. It is composed of a control board, a servo steering engine part, a power management part, a camera, and an additional device; it has the characteristics of easy maintenance and strong scalability, and different equipment can be installed or replaced according to different needs to meet the requirements. 4. A variable-focus optical camera is used to adjust the focal length when the object is too far away, to ensure the size of the object in the image and to ensure that the details of the object are not lost. Through the variable-focus camera, the enlarged image can be obtained without losing details when the object is too small to be accurately recognized, thereby reducing the failure rate of recognition and tracking. 5. The bus servo steering gear is adopted, which integrates DC motor, drive control, reduction gearbox, sensor and communication protocol. It has the advantages of high integration, high control accuracy, fast response, simple wiring, small hardware resource occupation, smooth flow and accurate data feedback. The servo bus steering gear has an angle range of 0-360 degrees and a theoretical control accuracy of 0.087890625 degrees. It can achieve fast response speed to the target, stable tracking and high tracking accuracy while ensuring a simple appearance, ensuring that the target will not be lost due to mechanical reasons; and the bus steering gear can adapt to complex environments and ensure the normal operation of the system in slightly harsh environments. 6. The 360-degree bus servo steering gear is used to form a two-degree-of-freedom pan-tilt. Theoretically, there is a dead point position. For example, after the steering gear rotates to 359 degrees, it cannot continue to rotate to 390 degrees in the same direction (rotate 30 degrees in the same direction after rotating to 0 degrees), which can achieve no-dead-angle tracking. Thereby ensuring the effect of target tracking, avoiding the loss of the target during the tracking process, and improving the tracking accuracy. 7. The recognition of different types of targets can be achieved by changing the data set and deep learning model. The system is not limited to drone identification and tracking, but can also identify and track civil airliners or detect other moving objects. 8. Cameras with different zoom ratios can be replaced according to the different heights of the objects to be identified. After modifying the camera parameters in the algorithm, accurate identification and tracking can be guaranteed.

[0147] In one embodiment of the present invention, a drone tracking system is provided, comprising:

[0148] The deep neural network search module searches for drones through drone tracking devices. When no drone is found and the drone is in the drone search state, the deep neural network visual template detection mode is used to search for suspicious targets in the field of view through the deep neural network; wherein the drone is used as the search target;

[0149] The machine vision detection module, when searching for a target, uses the recognition result to initialize the machine vision detection mode, so as to track the target through machine vision, and the deep neural network visual template detection mode exits;

[0150] The hybrid detection module switches to the hybrid mode to cooperate with each other to achieve accurate tracking of the target when the target is lost, the tracking result is inaccurate, the tracked target moves too fast, or the background interference is serious during the tracking process.

[0151] In the above embodiment, when no drone is found and the drone is in the drone search state, a deep neural network visual template detection mode is adopted to search for suspicious targets in the field of view through a deep neural network, including:

[0152] The deep neural network uses improved YOLOv10;

[0153] Resize the pixels of each frame captured by the drone tracking device proportionally, convert the image color back to RGB color space, and convert the read image data into QImage format and pass it to Label for display;

[0154] Copy the image data to get the same image data, pass any image data to the improved YOLOv10 algorithm, obtain the video stream feedback, target confidence and IOU size after recognition, and display the recognition result in a new window;

[0155] The target confidence and IOU information are stored in an array for caching, and the identified target position, identification box size, etc. are extracted from the tensor through feedback and stored in the array to obtain the coordinate information of the target in the image;

[0156] Repeat the above steps to complete the target tracking and identification.

[0157] In the above embodiment, after the target is searched, the recognition result is used to initialize the machine vision detection mode, so as to track the target by machine vision in the subsequent tracking process, including: the machine vision detection mode adopts KCF+HOG detection mode;

[0158] The human-machine system delineates the range of the target to be tracked in the displayed image, and the machine vision algorithm extracts features from the target range area, captures the original feature information of the target, and describes the background features of the target area.

[0159] Mapping the original target features extracted from the first frame image into a high-dimensional space to distinguish different target areas; and using the original target features as training samples to train the relevant filters in the high-dimensional feature space;

[0160] By comparing each frame of the image, the features of the required target candidate area are extracted, and the similarity between the candidate area and the target is evaluated through the correlation filter trained with the original features of the target;

[0161] According to the similarity degree obtained by evaluation, the candidate area with the highest similarity in a frame of image is selected as the position of the target in this frame, and the position of the target in the previous frame is used as the tracking target in the next frame; repeat the above steps, continuously train the relevant filter, and evaluate the similarity degree;

[0162] The target location is stored in the cache, the target is continuously tracked and ensured not to exceed the image range.

[0163] In the above embodiment, the hybrid mode is: a mode in which the deep neural network visual template detection mode and the machine vision detection mode cooperate with each other;

[0164] Hybrid modes work together to achieve accurate tracking of targets, including:

[0165] The acquired image is transmitted to the improved YOLOv10 detection algorithm in the deep neural network visual template detection mode for target recognition, and the target location and recognition box size are obtained;

[0166] The target location and the size of the recognition frame are passed into the machine vision detection mode, and the KCF+HOG algorithm is used to extract the features of the target area, capture the appearance information of the target, and describe the features of the target area;

[0167] The deep vision algorithm re-detects the target location and recognition box size at set intervals, and compares them with the position tracked by the correlation filter obtained by training the original features of the target in the KCF+HOG algorithm;

[0168] The recognition results of the two algorithms are compared to obtain the confidence levels of tracking accuracy. The accuracy of the two algorithms is judged by the confidence levels to correct tracking deviations.

[0169] In this embodiment, the accuracy of the two algorithms is judged by confidence and the tracking deviation is corrected, including:

[0170] Improve the IoU size in the loss function of the YOLOv10 algorithm according to the coordinate calculation;

[0171] Use Old Position to store the first or last recognized target center position, calculate the moving speed of the tracked target in the horizontal and vertical directions, and judge it as an interference object when the moving speed and moving distance are higher than the set threshold;

[0172] Combine IoU and confidence to further determine the accuracy of the target recognition, calculate the flight speed and IoU value of the drone in the two recognition results, and convert the speed difference into a normalized value Divc to calculate the normalized value flag;

[0173] The flag value is used to determine the position deviation between this recognition and the previous recognition result. When the position deviation is greater than the threshold and the intersection-union ratio is less than the set threshold, the tracking deviation is corrected.

[0174] Correcting tracking deviations includes:

[0175] When the fluctuation of the flag value is normal, there is no need to correct the tracking target;

[0176] When the flag value fluctuates beyond the set range, interference is considered to occur during the tracking process. At this time, the flag and the improved YOLOv10 algorithm are used to obtain the confidence conf, and the parameter Confidence is calculated based on this. Confidence is used to determine whether the algorithm recognition result is accurate. Based on the tracking result, it is comprehensively determined whether to re-transmit the new tracking target coordinates to the KCF algorithm for correction;

[0177] When interference or multiple targets gather, Divc and flag as conditions cannot accurately determine the correct target. The parameter Confidence is introduced to assist Divc and flag in making judgments. When the KCF tracking result needs to be corrected, the recognition result of the improved algorithm is passed into the KCF algorithm. When it is judged that the tracking result of a certain algorithm is wrong, another algorithm is used to correct it to ensure the accuracy of the tracking result.

[0178] The system provided in this embodiment is used to execute the above-mentioned method embodiments. Please refer to the above-mentioned embodiments for specific processes and detailed contents, which will not be repeated here.

[0179] In one embodiment of the present invention, a computing device is provided, which may be a terminal, and may include: a processor, a communication interface, a memory, a display screen, and an input device. Among them, the processor, the communication interface, and the memory communicate with each other through a communication bus. The processor is used to provide computing and control capabilities. The memory includes a non-volatile storage medium and an internal memory, and the non-volatile storage medium stores an operating system and a computer program. When the computer program is executed by the processor, the drone tracking method in each of the above embodiments is implemented; the internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a management network, NFC (near field communication) or other technologies. The display screen may be a liquid crystal display screen or an electronic ink display screen, and the input device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computing device, or an external keyboard, touchpad or mouse, etc. The processor may call the logic instructions in the memory.

[0180] In addition, the logic instructions in the above-mentioned memory can be implemented in the form of software functional units and can be stored in a computer-readable storage medium when sold or used as an independent product. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0181] In one embodiment of the present invention, a computer program product is provided, wherein the computer program product includes a computer program stored on a non-transitory computer-readable storage medium, wherein the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the methods provided by the above-mentioned method embodiments.

[0182] In one embodiment of the present invention, a non-transitory computer-readable storage medium is provided, wherein the non-transitory computer-readable storage medium stores server instructions, and the computer instructions enable a computer to execute the methods provided in the above embodiments.

[0183] The above embodiment provides a computer-readable storage medium, whose implementation principle and technical effect are similar to those of the above method embodiment, and will not be repeated here.

[0184] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0185] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0186] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0187] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A drone tracking method, characterized in that: include: Searching for drones through drone tracking devices. When no drone is found and the drone is in the drone search state, a deep neural network visual template detection mode is used to search for suspicious targets in the field of view through a deep neural network; wherein the drone is used as the search target; When the target is found, the recognition result is used to initialize the machine vision detection mode to track the target through machine vision, and the deep neural network visual template detection mode is exited; When the target is lost, the tracking result is inaccurate, the target moves too fast or the background interference is serious during the tracking process, switch to the mixed mode to cooperate with each other to achieve accurate tracking of the target.

2. The drone tracking method according to claim 1, characterized in that: When no drone is found and the drone is in the drone search state, the deep neural network visual template detection mode is used to search for suspicious targets in the field of view through the deep neural network, including: The deep neural network uses improved YOLOv10; Resize the pixels of each frame captured by the drone tracking device proportionally, convert the image color back to RGB color space, and convert the read image data into QImage format and pass it to Label for display; Copy the image data to get the same image data, pass any image data to the improved YOLOv10 algorithm, obtain the video stream feedback, target confidence and IOU size after recognition, and display the recognition result in a new window; The target confidence and IOU information are stored in an array for caching, and the identified target position, identification box size, etc. are extracted from the tensor through feedback and stored in the array to obtain the coordinate information of the target in the image; Repeat the above steps to complete the target tracking and identification.

3. The drone tracking method according to claim 1, characterized in that: When the target is searched, the recognition result is used to initialize the machine vision detection mode, so as to track the target through machine vision in the subsequent tracking process, including: the machine vision detection mode adopts KCF+HOG detection mode; The human-machine system delineates the range of the target to be tracked in the displayed image, and the machine vision algorithm extracts features from the target range area, captures the original feature information of the target, and describes the background features of the target area. Mapping the original target features extracted from the first frame image into a high-dimensional space to distinguish different target areas; and using the original target features as training samples to train the relevant filters in the high-dimensional feature space; By comparing adjacent frame images, the features of the target candidate area are extracted, and the similarity between the candidate area and the target is evaluated through the correlation filter trained with the original features of the target; According to the similarity degree obtained by evaluation, the candidate area with the highest similarity in a frame of image is selected as the position of the target in this frame, and the position of the target in the previous frame is used as the tracking target in the next frame; repeat the above steps, continuously train the relevant filter, and evaluate the similarity degree; The target location is stored in the cache, the target is continuously tracked and ensured not to exceed the image range.

4. The drone tracking method according to claim 1, characterized in that: The hybrid mode is a mode in which the deep neural network visual template detection mode and the machine vision detection mode cooperate with each other; Hybrid modes work together to achieve accurate tracking of targets, including: The acquired image is transmitted to the improved YOLOv10 detection algorithm in the deep neural network visual template detection mode for target recognition, and the target location and recognition box size are obtained; The target location and the size of the recognition frame are passed into the machine vision detection mode, and the KCF+HOG algorithm is used to extract the features of the target area, capture the appearance information of the target, and describe the features of the target area; The deep vision algorithm re-detects the target location and recognition box size at set intervals, and compares them with the position tracked by the correlation filter obtained by training the original features of the target in the KCF+HOG algorithm; The recognition results of the two algorithms are compared to obtain the confidence levels of tracking accuracy. The accuracy of the two algorithms is judged by the confidence levels to correct tracking deviations.

5. The drone tracking method according to claim 4, characterized in that: The accuracy of the two algorithms is judged by confidence level and tracking deviations are corrected, including: Improve the IoU size in the loss function of the YOLOv10 algorithm according to the coordinate calculation; Use Old Position to store the first or last recognized target center position, calculate the moving speed of the tracked target in the horizontal and vertical directions, and judge it as an interference object when the moving speed and moving distance are higher than the set threshold; Combine IoU and confidence to further determine the accuracy of the target recognition, calculate the flight speed and IoU value of the drone in the two recognition results, and convert the speed difference into a normalized value Divc to calculate the normalized value flag; The flag value is used to determine the position deviation between this recognition and the previous recognition result. When the position deviation is greater than the threshold and the intersection-union ratio is less than the set threshold, the tracking deviation is corrected.

6. The drone tracking method according to claim 5, characterized in that: Correcting tracking discrepancies includes: When the fluctuation of the flag value is normal, there is no need to correct the tracking target; When the flag value fluctuates beyond the set range, interference is considered to occur during the tracking process. At this time, the flag and the improved YOLOv10 algorithm are used to obtain the confidence conf, and the parameter Confidence is calculated based on this. Confidence is used to determine whether the algorithm recognition result is accurate. Based on the tracking result, it is comprehensively determined whether to re-transmit the new tracking target coordinates to the KCF algorithm for correction; When interference or multiple targets gather, Divc and flag as conditions cannot accurately determine the correct target. The parameter Confidence is introduced to assist Divc and flag in making judgments. When the KCF tracking result needs to be corrected, the recognition result of the improved algorithm is passed into the KCF algorithm. When it is judged that the tracking result of a certain algorithm is wrong, another algorithm is used to correct it to ensure the accuracy of the tracking result.

7. A drone tracking device, used to implement the drone tracking method according to any one of claims 1 to 6, characterized in that: include: A bottom steering gear bracket, on which a control panel is arranged, and the control panel is provided with a method for executing the drone tracking method; The single-axis bus servo steering gear is arranged below the bottom steering gear bracket and connected to the bottom steering gear bracket through bolts and nuts, and the output end of the single-axis bus servo steering gear passes through the top of the bottom steering gear bracket; The servo connection bracket is connected to the output end of the single-axis servo bus servo through the connection plate, and the single-axis servo bus servo drives the servo connection bracket to rotate 360°; A steering gear bracket, the top of which is fixed to the bottom of the adapter plate, and one end of the bottom is movably connected to the steering gear connecting bracket; The bus servo servo is arranged on the servo connecting bracket by means of bolts and nuts. The output end of the bus servo servo is connected to the other end of the bottom of the servo bracket through a connecting plate. The bus servo servo drives the servo bracket and the adapter plate to pitch and rotate. The camera and additional equipment are set at the center of the adapter board to obtain the target image information of the drone and transmit it to the control board.

8. The drone tracking device as claimed in claim 7, characterized in that: The bus servo steering gear adopts bus servo steering gear, and its control modes include: single-turn angle mode, multi-turn angle mode and rotation mode; Single-turn angle control mode: means that the rotation angle of the bus servo is limited to the range of 0 to 360 degrees or -180 to +180 degrees; Multi-turn angle control mode: allows the bus servo to rotate more than 360 degrees; Wheel control mode: can keep the bus servo rotating without being restricted to a fixed angle range, can accurately specify the number of turns of the servo, and can adjust the rotation direction and speed.

9. A drone tracking system, characterized in that: include: The deep neural network search module searches for drones through drone tracking devices. When no drone is found and the drone is in the drone search state, the deep neural network visual template detection mode is used to search for suspicious targets in the field of view through the deep neural network; wherein the drone is used as the search target; The machine vision detection module, when searching for a target, uses the recognition result to initialize the machine vision detection mode, so as to track the target through machine vision, and the deep neural network visual template detection mode exits; The hybrid detection module switches to the hybrid mode to cooperate with each other to achieve accurate tracking of the target when the target is lost, the tracking result is inaccurate, the tracked target moves too fast, or the background interference is serious during the tracking process.

10. A computer-readable storage medium storing one or more programs, characterized in that: The one or more programs include instructions, which, when executed by a computing device, cause the computing device to perform any one of the methods of claims 1 to 6.

Images