High-precision aircraft automatic identification and tracking system based on deep network

Abstract

The application relates to a high-precision aircraft automatic identification and tracking system based on a deep network, which adopts at least one imaging system of a landing area on one side of a runway safety interval to record an aircraft target, carries a high-definition visible light imaging module and a continuous zoom long-focus lens, and completes the visible light imaging and collection functions of a covered area; after the aircraft target is imaged in a recognizable size in a field of view, the system enters an automatic identification and tracking state, outputs a real-time tracking miss distance of the aircraft target relative to the center of the field of view, and keeps the target in the image center through closed-loop control of an optoelectronic turntable servo system. The application realizes the accurate positioning of the aircraft target by adopting the cooperative strategy of automatic identification and tracking, improves the identification and tracking precision of the aircraft target, meets the challenges of interference factors such as target size transformation, target rapid movement and complex background when the aircraft lands, and has good aircraft target identification and tracking capability.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of target recognition and tracking, in particular to a high-precision aircraft automatic recognition and tracking system based on a deep network. BACKGROUND

[0002] Most of the existing tracking systems for airport runways use multiple photoelectric turntables to alternately record the working mode. This scheme has lower requirements for the imaging system and the turntable, but is limited by the system performance, and it is difficult to effectively observe the aircraft target far away from the runway before landing and moving at high speed. Moreover, the alternation of multiple imaging systems will cause the landing process to be discontinuous, thereby affecting the tracking accuracy of the subsequent target. In the automatic tracking problem, the complex scene and the highly variable appearance of the target make it difficult to build a robust target appearance model. Traditional centroid and correlation tracking algorithms are simple to calculate and have better tracking effects in a single background, but they are less robust in complex tracking scenes such as cloud and fog obstruction and poor target contrast. In recent years, tracking algorithms based on deep learning have been widely used in the field of target tracking due to their strong feature learning ability. At present, the research on tracking algorithms based on deep learning mainly focuses on bounding-box networks. However, the bounding-box representation model only considers the spatial range within the bounding-box, ignoring the important local region position of the target shape, pose and semantics, which is easily affected by the background or the foreground region containing a small amount of semantic information, resulting in inaccurate prediction of the tracking model. In this application, a target recognition and tracking collaborative strategy is proposed for the tracking problem of a specified aircraft target. By integrating an end-to-end YOLOv8 recognition network and an accurate contour-based target tracking algorithm, the target is accurately located, and the recognition and tracking accuracy of the target is improved. SUMMARY

[0003] To solve the problem that the existing tracking equipment for airport runways is difficult to track the entire process of the aircraft from far to near landing, taxiing and stopping due to the long length of the runway and the high speed of the tracked target, and the low target tracking accuracy, the present application proposes a high-precision aircraft automatic recognition and tracking system based on a deep network. The present application first controls the photoelectric turntable to point to the aircraft target based on external guidance information, and uses a visible light imaging system to complete clear imaging of the aircraft during the entire landing process. At the same time, the automatic recognition and tracking collaborative strategy based on the deep network model realizes the capture, recognition and tracking of the specified aircraft target.

[0004] To solve the above problems, the present application adopts the following technical scheme:

[0005] A high-precision aircraft automatic identification and tracking system based on a deep network, comprising at least one front-end device arranged on one side of a runway safety landing interval and a back-end device remotely controlling the front-end device;

[0006] The front-end device comprises an imaging system mounted on an optoelectronic turntable for real-time high-definition imaging of an aircraft target, the imaging system being equipped with a high-definition visible light SDI imaging module and a continuous zoom long-focus lens, and further comprising a gimbal control system for guiding the optoelectronic turntable to perform high-precision pointing motion in a horizontal direction of 0-360° and a vertical direction of 0-90°, and a transmission conversion system for transmitting the SDI images collected by the imaging system to the back-end device, transmitting the serial port control signals of the imaging system sent by the back-end device to the imaging system, and issuing the tracking control commands output by the back-end device to the gimbal control system.

[0007] The back-end device comprises an image display enhancement system for performing image enhancement processing on the images transmitted by the transmission conversion system using an image enhancement algorithm, and an image intelligent processing system for automatically identifying and tracking the aircraft target based on the enhanced images, the image intelligent processing system adopting a cooperative strategy of target identification and tracking to obtain a real-time tracking miss distance of the aircraft target, and generating the tracking control commands based on the real-time tracking miss distance or received external guidance information, and then sending the tracking control commands to the transmission conversion system to realize servo closed-loop control of the optoelectronic turntable, wherein the cooperative strategy of target identification and tracking comprises the following steps:

[0008] An enhanced image after identification preprocessing is obtained, and the enhanced image after preprocessing is input into a YOLOv8 deep network model for operation, a target identification result is output at a time interval of multiple frames of images, the target identification result comprising a target position, a category probability and a confidence level;

[0009] A target sub-region template of a current frame of image is obtained based on the target identification result, and the target sub-region template is quickly matched with a globally enhanced image input in real time to obtain a real-time tracking miss distance of the aircraft target, and the target sub-region template is adjusted based on the real-time tracking miss distance, the adjusted target sub-region template being used for aircraft target tracking calculation of the next frame;

[0010] A target tracking algorithm based on an accurate contour is used for fast template matching, comprising the following steps:

[0011] The target sub-region template and a frame of globally enhanced image of a search target position are respectively input into a twin convolution network, and the twin convolution network outputs respective feature maps;

[0012] generate a feature map window response based on convolution calculation;

[0013] input the feature map window response into three different convolution network branches respectively to realize prediction of binary segmentation, bounding box coordinates and similarity score respectively;

[0014] take the bounding box corresponding to the position with the highest similarity score to update the tracking bounding box of the aircraft target to realize target tracking, and finally output the miss distance of the aircraft target relative to the center of the field of view as the real-time tracking miss distance of the aircraft target.

[0015] Compared with the prior art, the present application has the following beneficial effects:

[0016] The present application proposes a high-precision aircraft automatic identification and tracking system based on a deep network, which uses at least one photoelectric turntable arranged on one side of the runway safety landing area to record the aircraft target, and through the high-definition visible light imaging system and the continuous zoom long-focus lens, completes the visible light imaging and collection function of the coverage area. Then, when the aircraft target is imaged in the field of view and keeps the recognizable size, the system enters the automatic identification and tracking state, outputs the real-time tracking miss distance of the aircraft target relative to the center of the field of view, and through the closed-loop control of the photoelectric turntable servo system, keeps the target always located at the center of the image. The present application proposes a collaborative strategy for automatic identification and tracking of aircraft targets based on a deep network model, so as to realize accurate positioning of the aircraft target and improve the identification and tracking accuracy of the aircraft target. Through experiments on high-resolution image sequences of aircraft landing, the experimental results show that the algorithm used by the present system can completely meet the challenges of interference factors such as target size transformation, target rapid movement and complex background faced by the aircraft during landing, and verifies that the present system has good aircraft target identification and tracking capability and has broad application prospects. BRIEF DESCRIPTION OF DRAWINGS

[0017] Figure 1 is a structural block diagram of the high-precision aircraft automatic identification and tracking system based on a deep network described in the present application;

[0018] Figure 2 is a control data flow diagram of the high-precision aircraft automatic identification and tracking system based on a deep network;

[0019] Figure 3 is a schematic diagram of the shortest distance between the photoelectric turntable and the runway;

[0020] Figure 4 is a runway environment layout diagram;

[0021] Figure 5 is a schematic diagram of horizontal direction projection length calculation;

[0022] Figure 6 A schematic diagram for calculating the length of the vertical direction projection;

[0023] Figure 7 A flowchart of a cooperative strategy for target recognition and tracking;

[0024] Figure 8 A structural block diagram of a YOLOv8 deep network model;

[0025] Figure 9 A test result graph;

[0026] Figure 10 A target tracking network structure diagram based on an accurate contour;

[0027] Figure 11 An experimental result graph of a cooperative strategy for target recognition and tracking. DETAILED DESCRIPTION

[0028] The technical solutions of the present application will be described in detail below in combination with the preferred embodiments and the accompanying drawings.

[0029] The high-precision aircraft automatic identification and tracking system based on a deep network provided by the present application is mainly used for tracking the whole process of the aircraft from far to near landing, taxiing and stopping, automatically identifying and tracking targets with different flight speeds, recording and saving clear flight and landing taxiing process image data, and can be transmitted to a designated place in real time. The system is divided into a front-end device and a back-end device. The front-end device is mainly composed of an imaging system, a pan-tilt control system and a transmission conversion system, and is placed near the runway. The back-end device is mainly composed of an image display enhancement system, an image intelligent processing system and other supporting devices, and the back-end device adopts the design idea of high-performance processing computers to remotely control the front-end device. The composition of the high-precision aircraft automatic identification and tracking system is shown in Figure 1 , and the system control data flow diagram is shown in Figure 2 .

[0030] The imaging system is installed on an optoelectronic turntable arranged on one side of the runway safety landing interval. The imaging system is equipped with a high-definition visible light SDI imaging module and a continuous zoom long-focus lens, which realizes 1080p resolution 25 frame real-time high-definition imaging of the aircraft target. At the same time, the imaging system can automatically focus and zoom according to the target size and observation distance, so that the size of the aircraft target in the field of view is moderate and can be distinguished. In addition, in order to ensure that clear imaging can be realized in backlight and frontlight conditions, one front-end device is arranged on each side of the runway safety landing interval, and the working mode of two optoelectronic turntables is adopted for simultaneous recording, and each front-end device is in remote communication with the respective back-end device.

[0031] The front-end device is arranged near the runway. Its arrangement position should not only consider that the target imaging size of the aircraft is large enough to be detected by the device 10 kilometers away, and the distance from the runway should not be too far, but also ensure that the imaging field of view completely contains the aircraft target, and the motion ability of the photoelectric turntable is considered. Therefore, the distance from the runway should not be too close. According to the task requirements, it is known that the maximum speed of the aircraft during landing is about v max = 390 km / h, and the maximum angular velocity of the photoelectric turntable is w max = 60° / s. According to formula (1), the shortest distance between the photoelectric turntable and the runway should be greater than 103.45 meters, as shown in Figure 3 .

[0032]

[0033] Considering the complex environment near the runway, in order to avoid the on-site device and building from blocking the imaging, after obtaining the safe landing plan interval of the aircraft, which is 500 meters from the southwest end (i.e. the starting end) to 3000 meters from the southwest end, the front-end device is arranged as follows: two front-end devices are arranged on both sides of the safe landing interval of the runway, and the shortest distance between each photoelectric turntable and the runway is 110 m, as shown in Figure 4 .

[0034] After the front-end device is arranged, in order to realize automatic identification of the specified aircraft target (5 meters x 4 meters x 3 meters) at the farthest 10 kilometers, the imaging size of the target needs to meet the optimal state. Therefore, the focal length range of the imaging system is analyzed as follows.

[0035] The horizontal imaging size P x and the vertical imaging size P y of the aircraft in the imaging system can be represented by the following optical imaging formula:

[0036]

[0037] Where l x and l y represent the horizontal and vertical projection lengths of the aircraft in the imaging system, respectively; f represents the focal length of the optical system; L represents the distance between the aircraft and the photoelectric turntable; u represents the camera pixel size, and the pixel size u = 3.5 um in the imaging system.

[0038] When calculating the projection length of the aircraft, the aircraft is regarded as a cuboid with a size of 5 meters x 4 meters x 3 meters. The horizontal projection length l x is related to the length and width of the aircraft, as shown in Figure 5 ; and the vertical projection length l y is only related to the height of the aircraft, as shown in Figure 6The horizontal projection length of the aircraft at a distance of 10 km from the photoelectric turntable is calculated to be 30 meters, and the vertical projection length is calculated to be 3 meters. y

[0039] The recognition method based on the deep network adopted by the system has certain requirements for the imaging size of the target. If the imaging size of the target is too small, the detailed features are not obvious, and the recognition effect will be very poor. Therefore, the horizontal and vertical imaging sizes of the target need to be ensured to be more than 30 pixels, which can ensure good recognition and tracking capability. That is, the horizontal imaging size P x and the vertical imaging size P y of the aircraft in the imaging system need to meet the following conditions:

[0040]

[0041] Under the 330mm focal length of the visible light imaging system, the imaging size of the aircraft target at a distance of 10 km from the photoelectric turntable can be calculated according to the above formula (2) to be about 60x28 pixels, which meets the imaging size requirement of the target. Therefore, when the out-of-guide or tracking miss distance is directed to the aircraft at a distance of 10 km, the imaging system can automatically recognize and stably track the target, so as to guide the photoelectric turntable to track the whole process of the aircraft from far to near, landing, taxiing and stopping.

[0042] The pan-tilt control system includes a pan-tilt mechanism and a servo control system, which is used to guide the movement of the photoelectric turntable through the out-of-guide or tracking miss distance, and realize the high-precision pointing function in the horizontal direction of 0-360° and the vertical direction of 0-90°.

[0043] The transmission conversion system includes a communication control conversion module connected with the pan-tilt control system and the imaging system, a near-end optical transmitter for receiving the SDI image collected by the imaging system, and a far-end optical transmitter connected with the near-end optical transmitter, wherein the far-end optical transmitter is connected with the backend device. The transmission conversion system is used to convert the SDI image collected by the imaging system into an optical signal and transmit it to the backend device through an optical fiber, and is also used to convert the serial port control signal of the imaging system and the tracking control command (azimuth, elevation) of the photoelectric turntable sent by the backend device into the corresponding imaging system and pan-tilt control system.

[0044] The backend device mainly includes an image display enhancement system, an image intelligent processing system and other supporting devices, wherein the image display enhancement system has the function of enhancing the image transmitted by the transmission conversion system, can adopt corresponding image enhancement algorithm for image enhancement processing according to different scenes and different weather conditions, and improve the contrast and clarity of the image target; at the same time, the real-time recording function of the image can be realized.

[0045] ​​The image intelligent processing system has automatic identification and tracking functions, can receive external guidance information, adopts a cooperative strategy of target identification and tracking to obtain a real-time tracking miss distance of the aircraft target, integrates and generates a tracking control command according to the real-time tracking miss distance or the received external guidance information, and then sends the tracking control command to the gimbal control system through a transmission conversion system to realize servo closed-loop control of the photoelectric rotary table.

[0046] Other supporting devices in the back-end device mainly include a power supply system, which is mainly composed of a diesel engine, an UPS and supporting tools.

[0047] Further, as shown in Figure 2 , the back-end device further includes a display for high-definition display of visible light images, a display for display of control software of the image intelligent processing system, and a single rod for motion control of the photoelectric rotary table.

[0048] The present application realizes reliable identification of an aircraft target at a distance of 10km, and at the same time, completes real-time miss distance calculation of the aircraft target to control the servo system to realize stable tracking of the target. A cooperative strategy of target identification and tracking is adopted to calculate the real-time tracking miss distance of the aircraft target. The flow of the cooperative strategy of target identification and tracking is shown in Figure 7 , which is divided into two operation flows of target identification and target tracking, wherein:

[0049] Target identification flow: target identification is quasi-real-time work, that is, a recognition operation result is output at a time interval of multiple images to give sufficient operation time of a deep network for accurate identification of the target. The specific steps of the target identification flow include: obtaining an enhanced image after identification preprocessing, improving the contrast and definition of the image target, then transmitting the enhanced image after preprocessing to a YOLOv8 deep network model for operation, realizing automatic identification of the target, and outputting a target identification result including target position, category probability and confidence information, to provide target position indication correction for the target tracking flow.

[0050] Target tracking flow: target tracking is real-time work, and an operation result is output for each frame of image to give sufficient data rate of stable target follow-up tracking of the servo system. The specific steps of the target tracking flow include: obtaining a current frame enhanced image after tracking preprocessing, improving the contrast and definition of the image target, then obtaining a target sub-region template of the current frame image according to the target identification result obtained by the target identification flow, and performing fast template matching with the real-time input tracking preprocessed global enhanced image based on the template features to obtain a real-time tracking miss distance of the aircraft target, and then adjusting the target sub-region template based on the tracking result, and the adjusted target sub-region template is used for aircraft target tracking calculation of the next frame.

[0051] The selected recognition network takes the YOLOv8 network structure framework as a prototype. The network is based on the end-to-end idea, uses one network to complete the classification and object frame prediction at the same time, and realizes the accurate recognition and detection of the aircraft target. The structure diagram of the YOLOv8 deep network model is as shown in Figure 8 Firstly, the resized image is input into the CSPDarknet backbone extraction network to extract three feature maps of different branches and different scales. The CSPDarknet backbone extraction network is a deep convolutional neural network containing multiple convolutional layers, which can be used to extract the low-level to high-level feature representation of the input image, extract feature maps of different depths and different granularities, and ensure that the detection network has multi-scale target detection capability. The convolutional layers are continuously alternately used with 3x3 and 1x1 two different sizes of convolution kernels. The 3x3 convolution kernel makes the number of obtained feature maps double, which is convenient for the network to extract more target features; the 1x1 convolution kernel compresses the feature map dimension, reduces the model parameters, and reduces the complexity of network calculation. In order to ensure that the model converges quickly during training and achieves the purpose of stable training, a batch normalization layer BN is used after each convolutional layer to normalize the input data to have a mean of 0 and a variance of 1. In addition, an activation function is used after the batch normalization layer for nonlinear operation, so that the network can be applied in a nonlinear model. The convolutional layer, the batch normalization layer and the activation function layer constitute the basic structure of the convolutional block. Since the deep convolutional neural network is prone to gradient disappearance during training, which leads to saturation of the detection accuracy during the training process, a residual jump connection structure is introduced to deepen the network layer while solving the network degradation problem. In the feature pyramid structure, 2 times of up-sampling and down-sampling operations are adopted, and the same spatial resolution feature maps are fused through cascading to strengthen the feature representation of three different scales. The strengthened feature maps are input into three detection heads respectively, and finally the prediction of different scale targets is completed. In the detection head part, two prediction branches are designed to realize the decoupling of classification and regression tasks, and the output results include the predicted frame position coordinates, class probability and target confidence.

[0052] Further, for the aircraft image in a complex background, a rotation and cropping transformation method is adopted for data enhancement during model training. The original data is rotated and transformed at multiple angles, which can effectively enhance the robustness of the training model to angle changes and ensure that the model still has good recognition performance when the input aircraft image changes by a small angle. The experimental test results are as shown in Figure 9 The results show that the YOLOv8 deep network model used in the application can still accurately identify the target when the input image changes.

[0053] In actual video tracking problems, the complex scene and the appearance of the target with great changes make it difficult to build a robust target appearance model. The current representation method can be classified into two kinds: the widely used bounding-box representation method and the emerging detailed representation method. The bounding-box representation method inevitably contains the corresponding background information. Considering the foreground and background information comprehensively can effectively solve the tracking problems such as light, occlusion and the like, but the effect is not ideal when dealing with tracking problems caused by the structure or representation change of the target itself, and the background information contained will also interfere with the tracking. The emerging detailed representation method can provide a good solution to the representation problem caused by the structure change of the target itself.

[0054] Since the bounding-box representation only considers the spatial range within the bounding-box, the shape, posture and semantically important local region position of the object are not considered, and it is easily affected by the background content or the foreground region containing a small amount of semantic information; when tracking is performed by using the bounding-box, it is usually divided into regular matrix grid points for feature extraction, but since the grid points may not accurately fall within the meaningful object region, this also leads to inaccuracy of the tracking model; when the object rotates, the simple bounding-box representation usually causes great loss, and the reason is that the representation method itself has defects, so this representation method not only leads to a target model with low effectiveness, but also reduces the tracking performance of the algorithm.

[0055] The tracking algorithm is to track the target object itself rather than the corresponding bounding-box, so only accurate representation of the target can reach the upper bound of tracking accuracy. Therefore, unlike the existing object representation method which depends on low reliability, the application innovatively uses a target representation method based on accurate contour to essentially improve the accuracy and robustness of target tracking.

[0056] Based on the importance of the accurate contour of the target, the application trains the twin network on three tasks at the same time based on the twin network framework, and each task corresponds to a different strategy. Task one is to learn the similarity of the target object and the candidate area in the form of a sliding window, and the output is a similarity score score, which only indicates the similarity of the target, and does not provide any spatial range information. In order to optimize this information, the application uses an algorithm to simultaneously learn two additional tasks: box regression using a convolutional neural network and binary segmentation. The binary segmentation branch can encode the pixel-level mask information, which labels the accurate target contour. The use of additional branches and loss functions extends the existing twin tracking network, improving the accuracy of the network tracking boundary box prediction. The target tracking network structure based on accurate contour is as followsFigure 10 As shown in the figure, the image on the top left side is a target sub-region template output by the target recognition process, and the image on the bottom left side is a frame of global search image for searching the target position.

[0057] When the target tracking algorithm based on the accurate contour is used for fast template matching, the following steps are specifically included: the target sub-region template and a frame of global enhanced image for searching the target position are respectively input into a twin convolution network f θ , the twin convolution network f θ outputs respective feature maps; a feature map window response RoW is generated based on convolution calculation; next, based on the feature map window response RoW, the target tracking network based on the accurate contour utilizes three different convolution network branches b σ , to respectively realize the prediction of binary segmentation, boundary box coordinates box and similarity score score; the model updates the tracking boundary box of the target by using the boundary box corresponding to the position with the highest similarity score score, realizes target tracking, and finally outputs the miss distance of the aircraft target relative to the center of the field of view as the real-time tracking miss distance of the aircraft target.

[0058] The cooperative strategy of target recognition and tracking is tested on an actual video sequence of airplane landing, and the experimental results are shown in Figure 11 The experimental results show that the cooperative strategy of target recognition and tracking adopted in the present application can completely meet the challenges of interference factors such as target scale transformation, target rapid movement and complex background when the airplane lands, and finally realizes stable and accurate tracking.

[0059] In order to realize automatic identification of a designated aircraft target (scale: 5m*4m*3m) at a distance of 10km farthest and automatic tracking in the whole landing process, the application provides a high-precision aircraft automatic identification and tracking system based on a deep network, which adopts at least one photoelectric turntable arranged on one side of a runway safety landing interval to record and shoot the aircraft target, and through a high-definition visible light imaging system and a continuous zoom long-focus lens, completes the visible light imaging and collection functions of the covered area; then, when the aircraft target is imaged in the field of view and keeps the identifiable size, the system enters the automatic identification and tracking state, outputs the real-time tracking miss distance of the aircraft target relative to the field center, and through the closed-loop control of the photoelectric turntable servo system, keeps the target in the image center all the time. The application proposes a cooperative strategy for automatic identification and tracking of the aircraft target based on a deep network model, so as to realize accurate positioning of the aircraft target and improve the identification and tracking accuracy of the aircraft target. Through experiments on the high-resolution image sequence of the aircraft landing, the experimental results show that the algorithm used by the system can completely meet the challenges of the interference factors such as target size transformation, target rapid movement and complex background faced by the aircraft landing, and verifies that the system has good aircraft target identification and tracking capability and has wide application prospect.

[0060] The technical features of the above-described embodiments can be combined in any manner, and in order to make the description concise, all possible combinations of the technical features in the above-described embodiments are not described, however, as long as the combinations of the technical features do not exist contradictions, they should be considered as the scope of the present application.

[0061] The above-described embodiments only express several implementation manners of the application, and the description is relatively specific and detailed, but it should not be understood as a limitation on the scope of the patent. It should be pointed out that for ordinary skilled in the art, without departing from the concept of the application, some modifications and improvements can be made, which are all within the protection scope of the application. Therefore, the protection scope of the patent of the application should be subject to the appended claims.

Claims

1. A high-precision aircraft automatic identification and tracking system based on deep network, characterized by: The high-precision aircraft automatic identification and tracking system includes at least one front-end device arranged on one side of the runway safe landing zone and a back-end device for remotely controlling the front-end device; The front-end device includes an imaging system mounted on an optoelectronic turntable for real-time high-definition imaging of an aircraft target. The imaging system is equipped with a high-definition visible light SDI imaging module and a continuously variable focal length telephoto lens. The front-end device also includes a pan / tilt control system for guiding the optoelectronic turntable to perform high-precision pointing movements within the range of 0 to 360 degrees in the horizontal direction and 0 to 90 degrees in the vertical direction, and a transmission conversion system. The transmission conversion system is used to transmit the SDI images collected by the imaging system to the back-end device, transmit the imaging system serial port control signals sent by the back-end device to the imaging system, and send the tracking control commands output by the back-end device to the pan / tilt control system. The back-end device includes an image display enhancement system for performing image enhancement processing on the image transmitted by the transmission conversion system using an image enhancement algorithm, and an image intelligent processing system for automatically identifying and tracking the enhanced image. The image intelligent processing system uses a coordinated strategy of target identification and tracking to obtain the real-time tracking miss distance of the aircraft target, and generates the tracking control command based on the real-time tracking miss distance or the received external guidance information, and then sends the tracking control command to the transmission conversion system to realize servo closed-loop control of the optoelectronic turntable, wherein the coordinated strategy of target identification and tracking includes the following steps: Obtain the enhanced image after recognition preprocessing, input the preprocessed enhanced image into the YOLOv8 deep network model for operation, and output a target recognition result at intervals of multiple frames. The target recognition result includes the target location, category probability and confidence level; A target subregion template of the current frame image is obtained according to the target recognition result, and the target subregion template is quickly template matched with the global enhanced image after tracking preprocessing input in real time to obtain a real-time tracking miss distance of the aircraft target, and the target subregion template is adjusted based on the real-time tracking miss distance. The adjusted target subregion template is used for aircraft target tracking calculation in the next frame; The target tracking algorithm based on accurate contour is used for fast template matching, which includes the following steps: The target sub-region template and a frame of global enhanced image at the search target position are respectively input into the twin convolutional network, and the twin convolutional network outputs their respective feature maps; Generate feature map window response based on convolution calculation; The feature map window response is fed into three different convolutional network branches to achieve prediction of binary segmentation, bounding box coordinates, and similarity score respectively; The bounding box prediction corresponding to the position with the highest similarity score is used to update the tracking bounding box of the aircraft target to achieve target tracking. Finally, the miss distance of the aircraft target relative to the center of the field of view is output as the real-time tracking miss distance of the aircraft target.

2. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: In the YOLOv8 deep network model, the CSPDarknet backbone extraction network includes a basic structure of multiple convolutional blocks consisting of convolutional layers, batch normalization layers, and activation function layers, and improves the depth of the backbone extraction network based on the residual skip connection structure. The convolutional layer continuously alternates between two convolution kernels of different sizes, 3×3 and 1×1, to extract feature maps of three different branches and different scales of the input preprocessed image. In the feature pyramid structure, 2x upsampling and downsampling operations are used, and feature maps of the same spatial resolution are fused through cascading to enhance the feature representation of three different scales. The enhanced feature maps are input into three detection heads respectively, and the detection heads output the prediction results of targets of different scales, including the predicted box position coordinates, category probability, and target confidence.

3. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 2 is characterized in that: When training the YOLOv8 deep network model, rotation and cropping transformation methods are used to complete data enhancement.

4. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: The shortest distance L between the front-end device and the runway is calculated according to the following formula min : Among them, v max is the maximum speed of the aircraft during landing, w max is the maximum angular velocity of the photoelectric turntable.

5. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 4 is characterized in that: The shortest distance L between the front-end equipment and the runway min It is 110m.

6. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: A front-end device is arranged on both sides of the safe landing zone of the runway, and each front-end device remotely communicates with its own back-end device.

7. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: The safe landing range of the runway is 500 meters to 3000 meters from the start of the runway.

8. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: The power supply system of the high-precision aircraft automatic identification and tracking system consists of a diesel engine, a UPS and its supporting tools.

9. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1 is characterized in that: The back-end equipment also includes a display for high-definition display of visible light images, a display for displaying the control interface of the image intelligent processing system, and a single rod for controlling the motion of the photoelectric turntable through the pan-tilt control system.

10. The high-precision aircraft automatic identification and tracking system based on deep network according to claim 1, characterized in that: The transmission conversion system includes a communication control conversion module connected to the pan-tilt control system and the imaging system respectively, a near-end optical terminal for receiving the SDI image collected by the imaging system, and a far-end optical terminal connected to the near-end optical terminal, and the far-end optical terminal is connected to the back-end equipment.

Images