A demonstration system and method for automatically acquiring air situation target information at an observation post

Abstract

The present invention provides a demonstration system and a demonstration method for an observation post to automatically obtain air situation target information. The demonstration system includes demonstration hardware, a host, a display and a power supply. The demonstration hardware includes an experimental platform, an aircraft model power system and an observation post unit. The experimental platform is a platform frame with a sky background. The aircraft model power system is used to simulate low-altitude and ultra-low-altitude air targets. The observation post unit is used to automatically obtain video and image information of the target. The host is installed with a software system for collecting video and image information obtained from the observation post unit of the demonstration hardware, displaying and processing in real time, identifying air targets, generating target frames and target models, and solving feature information of air targets. The display device is connected to the host to display a software interface and recognition results. The present invention is used to simulate the slide rail and power system of the air target flight and observe.

Description

Technical Field

[0001] The invention relates to the field of demonstration, in particular to a demonstration technology for obtaining air situation target information. Background Art

[0002] At present, observers need to dispatch a large number of air observation posts when performing air reconnaissance missions. Each observation post still uses old air optical equipment to capture and collect aerial targets, and after manual identification by the observation post, they are reported to the command and decision-making units through radio, Beidou and other communication equipment. The accuracy and efficiency of manual identification mainly depend on the observer's prior knowledge, experience and cognition. Summary of the invention

[0003] The technical problem to be solved by the present invention is how to improve the accuracy of obtaining air situation target information.

[0004] The present invention solves the above-mentioned technical problems through the following technical means: a demonstration system for an observation post to automatically obtain air situation target information, comprising demonstration hardware, a host, a display and a power supply, the demonstration hardware comprising an experimental platform, an aircraft model power system and an observation post unit, the experimental platform is a platform frame with a sky background, the aircraft model power system is used to simulate low-altitude and ultra-low-altitude air targets, the observation post unit is used to automatically obtain video and image information of the target, the host is installed with a software system, which is used to collect video and image information obtained from the observation post unit of the demonstration hardware, display and process in real time, identify air targets, generate target frames and target models, and solve feature information of air targets, the display device is connected to the host, and is used to display a software interface and recognition results, the recognition results include air target models and positions, and the power supply is used to provide the required voltage for the normal operation of the host.

[0005] As an optimized technical solution, the platform frame is a cubic frame, and a sky background is laid on the top of the platform frame.

[0006] As an optimized technical solution, the aircraft model power system includes a slide rail, a slider, a slider power device, a support bracket, and an aircraft model. The slider power device includes a power supply, a variable speed motor, a speed regulation circuit board, and a traction rope. The variable speed motor and the speed regulation circuit board are powered by a power supply. The speed regulation circuit board is connected to the variable speed motor. The variable speed motor is connected to the traction rope. The other end of the traction rope is connected to the slider on the slide rail. The slide rail is fixed to the top of the platform frame with a sky background and is used to support the mounted slider. The slider mounts the aircraft model.

[0007] As an optimized technical solution, stainless steel pipes are used to build the cubic platform frame, and TBR cylindrical optical axis slides are used to fix them in reverse on the top of the frame.

[0008] As an optimized technical solution, the observation post unit includes a camera, a pitch servo, an azimuth servo, a servo controller, an encoder DC reduction motor, an encoder DC reduction motor controller, a host computer, and a power supply. The camera is set on the aircraft model, the pitch servo and the azimuth servo are respectively connected to the camera and the servo controller, the servo controller, the encoder DC reduction motor controller and the camera are connected to the host computer, and the encoder DC reduction motor controller is connected to the encoder DC reduction motor.

[0009] As an optimized technical solution, the camera is used to collect ground-to-air images and video data in the demonstration scene, the pitch servo is used to receive the PWM signal output by the servo controller to adjust the height angle of the camera, the azimuth servo is used to receive the PWM signal output by the servo controller to adjust the horizontal angle of the pitch servo, so as to adjust the horizontal angle of the camera, the servo controller is used to store a predetermined program to control the pitch servo and the azimuth servo to move within a certain range, so as to control the camera to search within a certain range, or receive the serial port signal output by the host computer from the GPIO pin, and control the movement of the pitch servo and the horizontal servo after modulation, and the encoder is directly The DC reduction motor receives the PWM signal output by the encoder DC reduction motor controller, and rotates forward or reverse at a constant speed or a variable speed. The encoder DC reduction motor controller is used to store a predetermined program to control the encoder DC reduction motor to rotate forward or reverse, so as to control the observation post platform to move within a certain range, or receive the serial port signal output by the host computer from the GPIO pin, and control the encoder DC reduction motor to move after modulation. The host computer is used to obtain camera image data, calculate the relative position of the target relative to the center of the camera, and output a signal to the servo controller to correct the servo position, or control the encoder DC reduction motor to rotate forward or reverse toward the target direction.

[0010] As an optimized technical solution, the observation post unit also includes wheels and a chassis. The wheels are installed at the bottom of the chassis of the observation post unit. The chassis is used to carry all components of the observation post unit. The encoder DC reduction motor is connected to the wheels.

[0011] As an optimized technical solution, the software system includes a model file and an observation post air target automatic recognition software. The model file is used to save a .pt file trained using a variety of air target videos and pictures. The observation post air target automatic recognition software is used to detect and recognize air target information in the acquired video according to a deep learning operation process. The observation post air target automatic recognition software is an improved aircraft target recognition algorithm based on GELAN-C, and is trained using a large amount of air target image data to obtain a .pt model file.

[0012] As an optimized technical solution, the improvement of the aircraft target recognition algorithm based on GELAN-C includes the following four aspects:

[0013] (1) Adding a preprocessing step for enhancing low-light images to the GELAN-C network, using a self-calibrated illumination network for preprocessing;

[0014] (2) Use HWD_ADown downsampling module;

[0015] (3) Using the DySample upsampling operator;

[0016] (4) Introduce the SGE attention mechanism.

[0017] The present invention also provides a method for performing a demonstration using any of the above-mentioned observation post automatic acquisition of air situation target information demonstration systems, comprising the following steps:

[0018] S1. Low and ultra-low altitude simulated flight of aerial targets

[0019] According to the flight parameters of the aerial targets attacking at low altitude or ultra-low altitude, the aerial target model can move at the set altitude, speed and attitude;

[0020] S2. Observation post automatically tracks and captures aerial targets

[0021] According to the tasks undertaken by the artificial air defense observation post unit in actual use scenarios, the task of automatically tracking and capturing air targets by the observation post unit is disassembled and implemented step by step;

[0022] S3. Identify aerial targets and display results

[0023] According to the need for the manual air defense observation post unit to identify the model of the aerial target during the process, the automatic observation post unit obtains the image and video data of the aerial target, the software system obtains the location and model information of the aerial target, and summarizes the identification results.

[0024] The advantages of the present invention are as follows: the present invention designs a demonstration system for automatically acquiring air situation target information at an observation post, which is used to simulate the slide rail and power system of an air target flight and observe it, so as to accurately acquire air situation target information, further identify the air target model through an identification method, and display the identification area and identification result statistical information. The present invention provides a reduced-size air target support environment for the training and learning of automatic identification of air targets at observation posts.

[0025] At present, the improved Faster R-CNN is used to detect and identify small sample targets of known drone models. It can identify close-range civil drone targets with known signals, but has certain limitations. The improved YOLOv3 is used to process the infrared and visible light images of drone targets. This method does not detect and identify the drone category. The RetinaNet deep neural network is used to detect multi-rotor drone targets. The detected drones are only for close-range multi-rotor drones, with low detection accuracy and poor applicability. The drone target detection algorithm based on YOLOv5-AntiUAV has a high detection efficiency for small targets, but no further research has been done on drone target recognition, and it has certain limitations in intercepting and capturing urban drones. The present invention adopts an improved drone target recognition algorithm based on GELAN-C. The algorithm first adds a SCINet illumination enhancement network to the GELAN-C network structure, and then uses the downsampling operator HWD_ADown operator to replace the Adown downsampling operator in the original GELAN-C network structure, and retains the image information in the encoder of the deep neural network as much as possible to improve the detection performance. The DySample upsampling operator replaces the standard sampling operator Upsample in the original GELAN-C network structure, further optimizes the model structure, reduces the number of parameters, and improves the detection and recognition efficiency of drone targets; finally, the SGE attention module is added to the entire GELAN-C network structure to enhance the focus on important features of the image and suppress the influence of noise. Through experiments, it is found that the GELAN-C model with an improved network structure model has greatly improved the model accuracy compared with the original GELAN-C model, while slightly increasing the number of parameters, the amount of calculation, and the size of the model weight. Through experiments, it is also found that the accuracy of the network model is significantly improved compared with other YOLOv9 models, and the number of parameters, the amount of calculation, and the size of the model weight are significantly reduced. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a block diagram of a demonstration system for automatically acquiring air situation target information of an observation post according to an embodiment of the present invention;

[0027] Figure 2 It is a flow chart of a demonstration system for an observation post to automatically acquire air situation target information according to the present invention. DETAILED DESCRIPTION

[0028] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0029] See also Figure 1 As shown, the demonstration system for automatically acquiring air situation target information of the observation post of the present invention includes demonstration hardware, a host, a display and a power supply. The demonstration hardware is used to simulate low-altitude and ultra-low-altitude aerial targets and automatically acquire video and image information of the target. The host is installed with a software system for collecting video and image information acquired from the multi-channel cameras of the demonstration hardware, displaying and processing in real time, identifying aerial targets, generating target frames and target models, solving feature information of aerial targets, and uploading result information. The display device is connected to the host and is used to display the software interface and recognition results, and the recognition results include the model and position of the aerial target. The power supply is used to provide the required voltage for the normal operation of the host.

[0030] The demonstration hardware includes an experimental platform, an aircraft model power system, and an observation post unit.

[0031] The experimental platform is a platform frame with a sky background. As a specific example, the platform frame is a cubic frame, and a sky background is laid on the top of the platform frame. The platform frame with a sky background is used to provide a framework support for the demonstration scene, and to provide support for the slide rail layout and the aircraft model hanger. The size of the platform frame can be set to 1m*1m*1m or 2m*2m*2m. The specific size can be freely adjusted according to the needs of the demonstration.

[0032] The aircraft model power system includes a slide rail, a slider, a slider power device, a support bracket, and an aircraft model. The slider power device includes a 3.7V power supply, a variable speed motor, a speed regulating circuit board, and a traction rope. The variable speed motor and the speed regulating circuit board are powered by a 3.7V power supply. The speed regulating circuit board is connected to the variable speed motor. The variable speed motor is connected to the traction rope. The other end of the traction rope is connected to the slider on the slide rail.

[0033] The slide rail is fixed on the top of the platform frame with a sky background, and is used to support the mounted slider to run at a constant speed or variable speed along the slide rail. The slider mounts the aircraft model through the support bracket, and the support bracket is used to fix the aircraft model in various flight postures. The aircraft model is used to simulate a real air target during the demonstration. The slider power device pulls the slider to move at a constant speed or variable speed along the slide rail through a traction rope. As a specific example, the size of the slide rail corresponding to the platform frame adopts two sizes: 1m or 2m.

[0034] The observation post unit includes a camera, a pitch servo, an azimuth servo, a servo controller, an encoder DC reduction motor, an encoder DC reduction motor controller, a host computer, wheels, a chassis, and a 7.4V power supply. The camera is set on the aircraft model and is used to collect ground-to-air images and video data in the demonstration scene. The pitch servo and the azimuth servo are connected to the camera and the servo controller respectively. The pitch servo is used to receive the PWM signal output by the servo controller to adjust the height angle of the camera. The azimuth servo is used to receive the PWM signal output by the servo controller to adjust the horizontal angle of the pitch servo, so as to adjust the horizontal angle of the camera. The servo controller, the encoder DC reduction motor controller and the camera are connected to the host computer. The servo controller is used to store a predetermined program to control the pitch servo and the azimuth servo to move within a certain range, so as to control the camera to search within a certain range, or receive the serial port signal output from the GPIO pin of the host computer, and control the movement of the pitch servo and the horizontal servo after modulation. The encoder DC reduction motor controller is connected to the encoder DC reduction motor and the wheels of the observation post unit, receives the PWM signal output by the encoder DC reduction motor controller, and rotates forward or reverse at a fixed speed or variable speed. The encoder DC reduction motor controller is used to store a predetermined program to control the encoder DC reduction motor to rotate forward or reverse, so as to control the observation post platform to move within a certain range, or receive the serial port signal output from the host computer from the GPIO pin, and modulate and control the encoder DC reduction motor to move. The host computer is used to obtain camera image data, calculate the relative position of the target relative to the center of the camera in combination with the moving target capture and tracking algorithm in opencv, and output the serial port signal to the servo controller through the GPIO pin to correct the servo position, or control the encoder DC reduction motor to rotate forward or reverse to approach the target direction. The wheel is installed at the bottom of the observation post unit to support the forward and backward movement of the entire observation post unit. The trolley chassis is used to carry all the components of the above-mentioned observation post units. The 7.4V power supply is used to power the host computer, the encoder DC reduction motor driver and the servo driver.

[0035] The host includes a motherboard, GPU, CPU, voltage-stabilized power supply, water-cooling system, memory, hard disk, etc. that can load the Pytorch deep learning framework, and is mainly used to support the software system to use deep learning methods to solve the acquired videos and images.

[0036] The software system includes a model file and an observation post air target automatic recognition software. The model file is used to save a .pt file trained using a variety of air target videos and pictures, and the observation post air target automatic recognition software is used to detect and recognize air target information in the acquired video according to the deep learning operation process. The software not only supports real-time recognition but also supports offline recognition of recorded video image data.

[0037] According to the above technical scheme, based on the various flight parameters of the aerial targets attacking at low or ultra-low altitudes, a platform frame is built to support the flight altitude of the aerial targets, a slide rail is constructed to fix the flight trajectory of the aerial targets, a power device is set to tow the aerial targets to move at a constant speed or variable speed, and a fixed support frame is set to simulate the different flight postures of the attacking aerial targets, thereby realizing the movement of the aerial target model at the set altitude, speed and posture.

[0038] See also Figure 2 As shown, the method for demonstrating the automatic acquisition of air situation target information demonstration system by the above observation post includes the following steps:

[0039] S1. Low and ultra-low altitude simulated flight of aerial targets

[0040] According to the flight parameters of the aerial targets attacking at low altitude or ultra-low altitude, the flight altitude of the aerial targets is supported by building a platform frame, the flight trajectory of the aerial targets is fixed by a slide rail, a power device is set to tow the aerial targets to move at a fixed speed or variable speed, and a fixed support frame is set to simulate the different flight postures of the attacking aerial targets, so as to realize the movement of the aerial target model at the set altitude, speed and posture;

[0041] S2. Observation post automatically tracks and captures aerial targets

[0042] According to the actual use scenario of the artificial air defense observation post unit, the task of automatically tracking and capturing air targets is disassembled and implemented step by step, so that it can "act like a real image" during the experimental demonstration and have all the functions of the air defense observation post unit;

[0043] S3. Identify aerial targets and display results

[0044] According to the characteristics that manual air defense observation post units need to identify the models of aerial targets in actual use scenarios to assist commanders in making decisions, the automatic observation post unit obtains aerial target images and video data, identifies the location and model information of the aerial targets based on the observation post automatic recognition software, and summarizes the recognition results.

[0045] As an example, the above observation post automatic acquisition of air situation target information demonstration system can adopt the following settings:

[0046] According to the height information of low-altitude and ultra-low-altitude attack targets, an aircraft model with a scale less than 1:100 is used to simulate air targets below 200m. The model needs to be placed within a 2m height range. Use 2m*1.5cm stainless steel pipes to build a 2m*2m*2m cubic platform frame, lay a sky background on the top of the platform frame, and use TBR cylindrical optical axis slide rails to fix it in reverse on the top of the frame. The highest target can be set at 2m.

[0047] According to the speed information of low-altitude and ultra-low-altitude attack targets, a variable speed motor is used with a speed range of 40rpm to 236rpm. Two types of pulleys with diameters of 3.5cm or 5cm are installed on the variable speed motor. A traction rope is fixed on the pulley and controlled by a speed regulating circuit board. Combined with the scale of the aircraft model and the height of the platform frame, it can simulate an aircraft model flying at a speed of 0.67*2*11m / s to 3.93*2*11m / s, that is, an aerial target flying at a speed of 14.74m / s to 86.46m / s; it can also simulate an aerial target flying at a speed of 0.67*2*15.7m / s to 3.93*2*15.7m / s, that is, an aerial target flying at a speed of 21.04m / s to 123.4m / s.

[0048] According to the attitude information of the low-altitude and ultra-low-altitude attacking targets, the aircraft model is reversely welded to the support bracket and hung upside down on the slider to fix the aerial target. During the welding process, the target attitude of the aircraft model can be appropriately adjusted so that the aerial target has a certain roll angle and pitch angle during the simulation process, simulating the actual flight attitude of the aerial target.

[0049] According to the characteristics of the artificial air defense observation post unit that it can move quickly in actual application scenarios, 4 sets of encoder DC reduction motors are used as independent power devices for the 4 wheels. Since the DC reduction motor has a large torque, it can convert the high-speed and low-torque output of the motor into a low-speed and high-torque output. The selection of this motor can further increase the movement power of the observation post unit. In addition, the encoder can provide real-time feedback on the running status of the motor, so that the encoder DC motor controller can accurately adjust the speed and position of the encoder DC reduction motor, and the exact position of the observation post unit can be obtained more accurately. According to the output speed of the encoder DC reduction motor is 450±10rpm, the tire diameter is 4.5mm, and the observation post unit can move at a maximum speed of 1.06m / s under full speed conditions. According to the requirements of the demonstration system, the movement state and speed of the artificial air defense observation post unit can be fully imitated under the conditions of scaling.

[0050] According to the fixed position of the artificial air defense observation post unit in the actual application scenario, the airspace within the surrounding 360-degree range can be searched and reconnaissance. The automatic observation post unit uses two types of servo units, namely, the directional servo and the pitch servo, which can rotate the rocker to place the camera imaging plane in any direction within the 360-degree range around the fixed position. The directional servo uses a 20kg metal digital servo with a rotation angle of 270 degrees and a no-load speed of 0.2sec / 60°; the pitch servo uses a 25kg metal digital servo with a rotation angle of 180 degrees and a no-load speed of 0.17sec / 60°. Both types of servos can work in a 6V-7.4V voltage environment, with a pulse width range of 500us-2500us corresponding to 0°-180°, and both use PWM pulse width control. In a fixed position, rotating the rocker toward any place within the surrounding 360-degree range can be completed within 1s, and the angular velocity can fully meet the speed of simulating the search and aiming of the artificial observation post.

[0051] In order to realize the search and tracking of moving targets in the airspace by the artificial observation post unit, a Raspberry Pi 4B host computer is used with a Picamera, and a moving target search and tracking algorithm based on opencv is written in Python and implanted in the Raspberry Pi hardware. According to the offset and offset speed between the moving target in the field of view and the center of the photo frame, the GPIO interface is used to output PWM signals to control the movement of two servos, and the rocker is rotated to achieve the camera plane aiming at the moving target in the field of view. The GPIO interface can be used to output PWM signals, and the encoder DC reduction motor controller is used to simultaneously control the movement of four groups of motors, so that the entire observation post unit moves to the most suitable position for observing aerial targets.

[0052] Based on the fact that the human eye can capture the image information of aerial targets through observation equipment in the actual application scenario, the automatic observation post unit has installed a DJI OSMO Action 4 sports camera on the rocker arm of the pitch servo. The camera has 9.98 million pixels and supports wide-angle shooting. In terms of image acquisition capabilities, it can completely simulate the image acquisition capabilities of the human eye. The image information acquired by Action 4 is transmitted to the host using a USB3.0 high-definition data cable, and the observation post air target automatic recognition software is run to obtain 4 video information streams, which can simulate the process of the human air defense observation post acquiring the image information of aerial targets.

[0053] According to the characteristic that the artificial air defense observation post unit can use a large amount of aircraft target recognition training in the early stage in the actual application scenario to quickly identify the aircraft model after observing the aerial target, the observation post aerial target automatic recognition software is based on the aircraft target recognition algorithm improved by GELAN-C, and uses a large amount of aerial target image data for training to obtain the .pt model file. The software interface for the observation post to automatically identify the aerial target is written using PySide, PyQt5 and Python. Before using the software to process the 4-channel video information stream, it is necessary to load the trained .pt model file first. Through the improved UAV target recognition algorithm based on GELAN-C, the 4-channel video signals can be processed in real time, the detected aerial targets can be framed, and the models of the aerial targets can be marked in the video. In addition, the video signal that detects the aerial targets can be displayed on the software system interface in real time. If multiple video signals can detect the aerial targets, the result display interface of the software system will switch to display the detection results of each video signal, which can simulate the process of identifying the aerial targets after the artificial observation post observes them.

[0054] The demonstration method of the demonstration system for automatically acquiring air situation target information by an observation post using the above specific embodiment specifically includes:

[0055] S1. Low and ultra-low altitude simulated flight of aerial targets

[0056] Build a 2m*2m*2m platform frame to support the aerial target demonstration platform;

[0057] Hang the three m rails upside down on the top of the platform frame;

[0058] The 3.7V power supply drives the variable speed motor to mount the traction rope to drag the aircraft model welded on the slider at a fixed speed or variable speed to make it move at a certain speed, attitude and direction;

[0059] S2. Observation post automatically tracks and captures aerial targets

[0060] The OSMO Action4 camera on the rocker arm mounts acquires video image information within the field of view;

[0061] Receive the GPIO signal from the host computer, and the servo driver drives the pitch servo and the direction servo to drive the rocker arm to move;

[0062] Receive the GPIO signal of the host computer, and the encoded DC motor driver drives 4 motors to drive the four-wheel movement;

[0063] The host computer obtains the moving target information in the field of view through PiCamera, and sends GPIO signals to the motor driver and servo driver through solving;

[0064] The multi-layered trolley chassis supports the various components of the entire automatic observation post unit;

[0065] S3. Identify aerial targets and display results

[0066] Real-time summary display of each frame recognition result, and display of recognition quantity, recognition model and recognition speed;

[0067] The target recognition software written based on the improved aircraft target recognition algorithm of GELAN-C is used to detect the air target information from the 4-channel video signals in real time, and the recognition result is marked in the video, and the air target model information is displayed next to the frame mark;

[0068] Use a high-definition USB 3.0 cable to get video information from 4 OSMO Action4 cameras.

[0069] The aircraft target recognition algorithm based on GELAN-C is improved by introducing the GELAN-C lightweight model in the YOLOv9 model as the basic network model, and improving the GELAN-C lightweight model according to the characteristics of the aircraft model to be identified. The main improvements include the following four aspects:

[0070] 1. SCINet Illumination Self-Enhanced Network

[0071] In the process of aircraft model detection and recognition, recognition usually fails due to its small size and weak ability to reflect light. If the input video image is preprocessed similar to the traditional target detection algorithm before detection and recognition, the detection and recognition effect of such video images will be better. Based on this, adding enhancement processing steps for low-light images to the GELAN-C network can better improve the algorithm's detection and recognition accuracy for targets. SCINet (Self-Calibrated Illumination Network) processes images through a cascaded illumination learning process and a weight sharing mechanism, improving the model's ability to adapt to different low-light conditions and a variety of scenes, further improving the model's generalization ability and practical application value; the network also introduces a self-correction module to reduce the computational burden while ensuring the stability of the results, and uses an unsupervised training loss function to increase the model's adaptability to different scenes.

[0072] 2. HWD-ADown downsampling operator

[0073] In the YOLO series of algorithms, the role of downsampling operation is to change the resolution of the input image, aggregate its local features, expand the receptive field of the neural network and reduce the computational burden. The GELAN-C network uses the ADown downsampling module, which first reduces the feature map size by performing a 2D average pooling operation on the input channel image. Secondly, it divides it into two channels for convolution and pooling operations. Channel 1 only performs convolution operations, and channel 2 first performs a 2D maximum pooling operation and then a convolution operation. The third step is to connect and output the feature maps of the two channels. Through analysis, it is found that although the ADown downsampling module can dynamically adjust the downsampling ratio according to the input feature map, improve the detection performance while maintaining the resolution of the feature map, and reuse feature information of different resolutions to improve the detection effect of small targets, its design increases the computational complexity and parameter amount of the model, and the steps of channel division, average pooling, maximum pooling and channel merging may lead to the loss of feature information. The core idea of ​​the HWD_ADown (Haar wavelet downsampling) downsampling module is to use Haar wavelet transform to reduce the spatial resolution of the feature map without losing information. During the calculation process, the HWD_ADown module uses Haar wavelet transform to increase the number of channels of the feature map and reduce its resolution, and uses convolution operation to perform representative feature tilt filtering redundant information. The HWD_ADown module can retain image information in the encoder of the deep neural network as much as possible, so that it is easier to obtain more image discriminant features, thereby improving detection performance.

[0074] 3. DySample upsampling operator

[0075] In the GELAN-C network structure, the Upsample upsampling operator is used, which can process operations such as image size adjustment and feature map amplification at low cost, but lacks the ability to process complex scenes and images with rich details, and cannot effectively process structural images with complex backgrounds and retain their important feature information. The present invention selects a lighter upsampling operator with smaller parameters. The FADE (Full Resolution Adaptive Estimation) upsampling operator can adaptively sample feature maps with more contextual information, so as to better restore image details, but the high resolution of this operator will introduce a lot of computational complexity; the SAPA (Semantic-Aware Pixel Aggregation) upsampling operator can upsample feature maps in semantic segmentation tasks, so as to improve the accuracy of semantic segmentation, but the computational complexity of this operator is high, and the operation process requires additional cuda code to run, which is inconvenient to use; the CARAFE (Content-Aware ReAssembly of Features) upsampling operator has a larger receptive field, stronger content perception ability, and higher computational efficiency than upsampling operators such as FADE and SAPA, but it also needs to use an additional sub-network to generate dynamic kernels, which requires a larger amount of computation. The DySample (Dynamic upSample) upsampling operator is an ultra-light and efficient dynamic upsampling operator. This operator is implemented from the perspective of point sampling through PyTorch standard built-in functions, does not require the use of additional CUDA packages, can effectively utilize resources, and reduce parameters and computational complexity. Therefore, the conventional Upsample upsampling operator is replaced by the DySample upsampling operator in the GELAN-C network model to further optimize the model structure, reduce the number of parameters, and improve the detection and recognition efficiency of UAV targets.

[0076] 4. SGE Attention Mechanism

[0077] During the detection and recognition process of aircraft models, different video image contrasts will be generated due to changes in lighting, background and other conditions, resulting in large differences in image tones. In addition, due to the flexibility of the aircraft model, its contour features, key points, lines and surface features become blurred as it moves. In addition, the sky background makes it difficult to separate the drone target from the background information.

[0078] With the widespread use of deep learning in the field of vision, the attention mechanism has also been widely used in the field of computer vision. In the process of image processing, the attention mechanism can be used as a computing resource allocation scheme to concentrate limited computing resources on processing more important information. It can focus on processing a certain part of the input information, thereby improving the efficiency of the network model. SGE is better than CBAM and A 2-Nets and SE attention mechanisms not only have a lighter structural design, but also do not require additional training supervision, and can improve the detection ability of small objects and have enhanced representation capabilities. Therefore, the SGE attention mechanism is introduced in the GELAN-C model.

[0079] In order to verify the improved aircraft target recognition algorithm based on GELAN-C, ablation experiments were carried out on multiple combinations of added and replaced modules. The experimental results on the test set are shown in Table 1. Using GELAN-C as the pre-trained model and adding the illumination enhancement network SCINet, mAP_0.5 increased by 0.99% and mAP_0.5-0.95 increased by 1.09%; after adding the SCINet network and replacing the downsampling operator HWD_ADown, mAP_0.5 increased by 11.33% and mAP_0.5-0.95 increased by 4.92%; after adding the SCINet network and replacing the upsampling operator DySample, mAP_0.5 increased by 9.22% and mAP_0.5-0.95 increased by 5.14%; after adding the SCINet network and using the SGE attention mechanism, mAP_0.5 increased by 11.4% and mAP_0.5-0.95 increased by 5.83%; after adding SCINet at the same time and replacing the downsampling operator HWD_ADown and the upsampling operator DySample, mAP_0.5 increased by 8.3% and mAP_0.5- After adding SCINet and replacing the downsampling operator HWD_ADown and the SGE attention mechanism, mAP_0.5 increased by 8.85%, and mAP_0.5-0.95 increased by 4.82%; after adding SCINet and replacing the upsampling operator DySample and the SGE attention mechanism, mAP_0.5 increased by 10.49%, and mAP_0.5-0.95 increased by 6.5%; finally, after adding the SCINet network and the SGE attention mechanism, and replacing the sampling operator HWD_ADown and the upsampling operator DySample, the improved model increased mAP by 11.09% and mAP_0.5-0.95 by 7.92% compared with the GLEAN-C model alone, and the number of parameters, calculation amount and model size increased by 6.45%, 5.07% and 6.21% respectively, further meeting the high-precision requirements for detection and recognition of drone targets.

[0080] Table 1 Ablation experiment results

[0081]

[0082]

[0083] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. 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 demonstration system for automatically acquiring air situation target information at an observation post, characterized in that: The device comprises demonstration hardware, a host, a display and a power supply. The demonstration hardware comprises an experimental platform, an aircraft model power system and an observation post unit. The experimental platform is a platform frame with a sky background. The aircraft model power system is used to simulate low-altitude and ultra-low-altitude aerial targets. The observation post unit is used to automatically obtain video and image information of the target. A software system is installed on the host for collecting, displaying and processing the video and image information obtained from the observation post unit of the demonstration hardware in real time, identifying aerial targets, generating target frames and target models, and solving feature information of aerial targets. The display device is connected to the host for displaying the software interface and recognition results. The recognition results include the model and position of the aerial target. The power supply is used to provide the required voltage for the normal operation of the host.

2. A demonstration system for automatically acquiring air situation target information at an observation post according to claim 1, characterized in that: The platform frame is a cubic frame, and a sky background is laid on the top of the platform frame.

3. The automatic acquisition of air situation target information demonstration system for an observation post according to claim 2 is characterized in that: The aircraft model power system includes a slide rail, a slider, a slider power device, a support bracket, and an aircraft model. The slider power device includes a power supply, a variable speed motor, a speed regulating circuit board, and a traction rope. The variable speed motor and the speed regulating circuit board are powered by the power supply. The speed regulating circuit board is connected to the variable speed motor. The variable speed motor is connected to the traction rope. The other end of the traction rope is connected to the slider on the slide rail. The slide rail is fixed to the top of the platform frame with a sky background and is used to support the mounted slider. The slider mounts the aircraft model.

4. A demonstration system for automatically acquiring air situation target information at an observation post as claimed in claim 3, characterized in that: Use stainless steel pipes to build the cube platform frame, and use TBR cylindrical optical axis slide rails to fix it in reverse on the top of the frame.

5. The automatic acquisition of air situation target information demonstration system for an observation post according to claim 3 is characterized in that: The observation post unit includes a camera, a pitch servo, an azimuth servo, a servo controller, an encoder DC reduction motor, an encoder DC reduction motor controller, a host computer, and a power supply. The camera is arranged on the aircraft model, the pitch servo and the azimuth servo are respectively connected to the camera and the servo controller, the servo controller, the encoder DC reduction motor controller, and the camera are connected to the host computer, and the encoder DC reduction motor controller is connected to the encoder DC reduction motor.

6. A demonstration system for automatically acquiring air situation target information at an observation post as claimed in claim 5, characterized in that: The camera is used to collect ground-to-air images and video data in the demonstration scene, the pitch servo is used to receive the PWM signal output by the servo controller to adjust the height angle of the camera, the azimuth servo is used to receive the PWM signal output by the servo controller to adjust the horizontal angle of the pitch servo, so as to adjust the horizontal angle of the camera, the servo controller is used to store a predetermined program to control the pitch servo and the azimuth servo to move within a certain range, so as to control the camera to search within a certain range, or receive the serial port signal output by the host computer from the GPIO pin, and control the movement of the pitch servo and the horizontal servo after modulation, the encoder DC reduction motor Receive the PWM signal output by the encoder DC reduction motor controller, and rotate forward or reverse at a constant speed or a variable speed. The encoder DC reduction motor controller is used to store a predetermined program to control the encoder DC reduction motor to rotate forward or reverse, so as to control the observation post platform to move within a certain range, or receive the serial port signal output from the GPIO pin of the host computer, and control the movement of the encoder DC reduction motor after modulation. The host computer is used to obtain camera image data, calculate the relative position of the target relative to the center of the camera, and output a signal to the servo controller to correct the servo position, or control the encoder DC reduction motor to rotate forward or reverse toward the target direction.

7. The automatic acquisition of air situation target information demonstration system for an observation post according to claim 5 is characterized in that: The observation post unit also includes wheels and a chassis. The wheels are installed at the bottom of the chassis of the observation post unit. The chassis is used to carry all the components of the observation post unit. The encoder DC reduction motor is connected to the wheels.

8. The automatic acquisition of air situation target information demonstration system for an observation post according to claim 1 is characterized in that: The software system includes a model file and observation post air target automatic recognition software, wherein the model file is used to save a .pt file trained using a variety of air target videos and pictures, and the observation post air target automatic recognition software is used to detect and recognize air target information in the acquired video according to a deep learning operation process, and the observation post air target automatic recognition software is an aircraft target recognition algorithm improved based on GELAN-C, and is trained using a large amount of air target image data to obtain a .pt model file.

9. A demonstration system for automatically acquiring air situation target information at an observation post as claimed in claim 8, characterized in that: The improvements of the aircraft target recognition algorithm based on GELAN-C include the following four aspects: (1) Adding a preprocessing step for enhancing low-light images to the GELAN-C network, using a self-calibrated illumination network for preprocessing; (2) Use HWD_ADown downsampling module; (3) Using the DySample upsampling operator; (4) Introduce the SGE attention mechanism.

10. A method for demonstrating a demonstration system for automatically acquiring air situation target information at an observation post according to any one of claims 1 to 9, characterized in that: The steps include: S1. Low and ultra-low altitude simulated flight of aerial targets According to the flight parameters of the aerial targets attacking at low altitude or ultra-low altitude, the aerial target model can move at the set altitude, speed and attitude; S2. Observation post automatically tracks and captures aerial targets According to the tasks undertaken by the artificial air defense observation post unit in actual use scenarios, the task of automatically tracking and capturing air targets by the observation post unit is disassembled and implemented step by step; S3. Identify aerial targets and display results According to the need for the manual air defense observation post unit to identify the model of the aerial target during the process, the automatic observation post unit obtains the aerial target image and video data, the software system obtains the location and model information of the aerial target, and summarizes the identification results.

Images