Vehicle-mounted anti-unmanned aerial vehicle intelligent photoelectric device and system

Abstract

The invention relates to the field of anti-unmanned aerial vehicle systems, in particular to a vehicle-mounted anti-unmanned aerial vehicle intelligent photoelectric device and system. The technical problems are as follows: a miniaturized light-weight photoelectric system adopts an FPGA + DSP photoelectric system to track and detect a small unmanned aerial vehicle, the tracking and identification capabilities are weak, the tracking of the unmanned aerial vehicle is unstable, and a target cannot be detected and identified; according to the technical scheme, the vehicle-mounted anti-unmanned aerial vehicle intelligent photoelectric device and system comprise a photoelectric load assembly and a servo control assembly; compared with the prior art that a miniaturized and lightweight photoelectric system adopts an FPGA + DSP photoelectric system to track and detect a small unmanned aerial vehicle, the tracking and recognition capability is weak, the tracking of the unmanned aerial vehicle is unstable, and a target cannot be detected and recognized, the system adopts an intelligent target detection and recognition model based on a YOLOv5 algorithm, and the system has a wide application prospect in a domestic Hisilicon processor Hi3559AV100. 30 frames of images can be processed per second, the detection precision is high, the result is accurate, the anti-interference capability is strong, and the all-weather and all-time target identification requirement can be met.

Description

Technical Field

[0001] The present invention relates to the field of anti-UAV systems, and in particular to a vehicle-mounted anti-UAV intelligent optoelectronic device and system. Background Art

[0002] The countermeasure of UAV optoelectronic system refers to the use of different technologies and equipment to defend against the optoelectronic system of UAV; these optoelectronic systems include infrared, laser and visual sensors, etc.; the countermeasure of UAV optoelectronic system aims to detect and interfere with the sensors of UAV to reduce the surveillance and strike capabilities of UAV against targets; in the optoelectronic system against UAV, the most commonly used technologies are infrared jamming and laser jamming; infrared jamming technology uses heat sources to interfere with the infrared sensors of UAV, making it unable to accurately detect targets; this jamming technology can confuse the sensors of UAV by releasing a large amount of heat sources, such as flames and smoke; in addition, infrared curtains can be used to reduce infrared radiation, thereby reducing the detection capability of UAV; laser jamming technology disrupts the visual system of UAV by emitting laser beams to its visual sensors; laser jamming technology can be divided into active jamming and passive jamming; active jamming destroys the sensors of UAV by sending laser beams with high power and high energy density; passive jamming uses laser jammers, such as interferometers and speckle plates, to interfere with UAV the sensors of the drone; in addition, there are other technologies for anti-UAV optoelectronic systems, such as electronic countermeasures, electromagnetic interference and reflective interference; electronic countermeasures technology interferes with the communication and navigation systems of drones by sending electromagnetic signals; electromagnetic interference technology interferes with the radar and communication systems of drones by emitting electromagnetic waves; reflective interference technology uses reflective materials such as metal foil and mirrors to reflect light, thereby reducing the detection capability of drones; in summary, anti-UAV optoelectronic systems interfere with the optoelectronic sensors of drones by using various interference technologies and equipment to reduce the surveillance and strike capabilities of drones; these countermeasures include infrared interference, laser interference, electronic countermeasures, electromagnetic interference and reflective interference; these technologies can effectively protect targets from the threat of drones; when stably tracking and measuring small drones, the miniaturized and lightweight optoelectronic system uses an FPGA+DSP optoelectronic system to track small drones, with weak detection, tracking and identification capabilities, unstable tracking of drones, and failure to detect and identify targets. Summary of the invention

[0003] In order to overcome the problems of weak detection, tracking and identification capabilities, unstable tracking of drones, and failure to detect and identify targets when stably tracking and measuring small drones, a miniaturized and lightweight optoelectronic system uses an FPGA+DSP optoelectronic system to track small drones.

[0004] The technical solution of the present invention is: a vehicle-mounted anti-UAV intelligent optoelectronic device, including a lower shell, an azimuth component, a pitch component, a middle frame, a front cover, a rear cover, an optoelectronic load component and a servo control component. The azimuth component is arranged above the lower shell, the pitch component is arranged above the azimuth component, the middle frame is arranged inside the pitch component, the front cover is arranged on one side of the middle frame, the rear cover is arranged on the other side of the middle frame, the optoelectronic load component is arranged inside the middle frame, the servo component control component is arranged inside the pitch component, the lower shell is made of hard aluminum alloy material, the middle frame is made of hard aluminum alloy material, the front cover is made of hard aluminum alloy material, and the rear cover is made of hard aluminum alloy material.

[0005] Preferably, by setting the lower shell, the azimuth component can be installed and fixed, and by setting the azimuth component, the pitch component can be driven to rotate, that is, the azimuth adjustment can be performed. By setting the middle frame, the front cover and the rear cover, the photoelectric load component can be installed and fixed, and the structural weight reduction and heat dissipation design can be optimized. By setting the pitch component, the middle frame can be driven to rotate, that is, the pitch angle can be adjusted, thereby adjusting the azimuth of the photoelectric load component. By setting the photoelectric load component, uninterrupted monitoring and detection of the target can be achieved day and night, so that the azimuth component and the pitch component can be used to drive the photoelectric load component to adjust the azimuth and pitch angles, thereby achieving all-round and uninterrupted monitoring and detection of the target day and night. By setting the servo control component, the azimuth component and the pitch component can be adjusted according to the data of the photoelectric load component.

[0006] Preferably, the azimuth assembly includes a power filter, an azimuth motor, an azimuth encoder and an azimuth base. The power filter is arranged inside the lower shell, the azimuth motor is arranged above the lower shell, the power filter is connected to the azimuth motor, the azimuth encoder is arranged inside the lower shell, the azimuth encoder is connected to the azimuth motor, the azimuth base is arranged above the lower shell, the azimuth base is arranged outside the azimuth motor, and the azimuth base is made of hard aluminum alloy material. When in use, the rotation angle of the azimuth motor can be detected by setting the azimuth encoder so that the azimuth of the photoelectric load assembly can be accurately fed back. By setting the power filter, a stable power supply can be provided for the azimuth motor to prevent the noise in the power supply from affecting the normal operation of the device. By setting the azimuth base, the azimuth motor can be installed and fixed to stabilize the supporting structure of the device. By setting the azimuth motor, the pitch assembly can be driven to rotate.

[0007] Preferably, the pitch assembly includes a U-shaped main frame, a side cover, a pitch motor and a pitch encoder, the U-shaped main frame is arranged above the azimuth base, the U-shaped main frame is fixedly connected to the azimuth base, the side covers are arranged in two groups, the side covers are arranged on opposite sides of the U-shaped main frame, the pitch motor is arranged inside one of the two groups of side covers, and the pitch encoder is arranged inside the other of the two groups of side covers, the U-shaped main frame is made of hard aluminum alloy material, the side cover is made of hard aluminum alloy material, the output end of the azimuth motor is connected to the U-shaped main frame, and the output end of the pitch motor is connected to the middle frame. When in use, the side cover, the pitch motor, the pitch encoder and the middle frame can be installed by setting the U-shaped main frame, the side cover can be set to accommodate and seal the pitch motor and the pitch encoder, the pitch motor can drive the middle frame to rotate by setting the middle frame, and the pitch encoder can identify the rotation angle of the middle frame by setting the pitch encoder.

[0008] Preferably, the optoelectronic payload assembly includes an infrared imager, a visible light camera, an inertial gyroscope, a laser rangefinder and a tracking and control unit. The infrared imager is arranged inside the middle frame, the visible light camera is arranged on one side of the infrared imager, the inertial gyroscope is arranged on one side of the infrared imager, the laser rangefinder is arranged below the visible light camera, and the tracking and control unit is arranged inside the middle frame. The inertial gyroscope adopts a micromechanical gyroscope, and the tracking and control unit adopts the domestic HiSilicon processor Hi3559AV100. When in use, the infrared imager can be used to collect infrared images, the visible light camera can be used to collect high-definition images, the inertial gyroscope can be used to assist the azimuth encoder and the pitch encoder to increase accuracy, the laser rangefinder can be used to detect the distance to the target, and the tracking and control unit can be used to control the infrared imager and the visible light camera according to the detection result of the laser rangefinder.

[0009] Preferably, the servo control assembly includes a power isolation unit and a servo control unit, the power isolation unit is arranged inside the side cover, the servo control unit is arranged inside the side cover, and the servo control unit includes:

[0010] The initialization module is used to initialize the processor's internal data storage, registers, system clock, interrupt system, GPIO, RS422 and SPI communications when the system is powered on and reset;

[0011] The platform self-check module is used to check whether the positions of the inertial gyroscope, azimuth encoder and pitch encoder are correct, and whether the working angle range of the U-shaped main frame and middle frame is correct;

[0012] The data communication processing module is used to send the azimuth and pitch angle information of the middle frame to the main control module according to the azimuth encoder data, the pitch encoder data and the speed data of the inertial gyroscope;

[0013] The servo control calculation module is used for servo control calculation. It mainly uses the corresponding infrared imager, visible light camera or laser rangefinder data to filter and calculate the servo control quantity according to the instructions, outputs the current control instruction according to the calculated control quantity, and sends the platform angle information back to the main control module;

[0014] The driver board communication module is used to realize the communication between the main control module and the azimuth motor and the pitch motor, wherein the information includes: current control instructions, received current feedback signals, azimuth motor movements and pitch motor movements;

[0015] The main control module is used to control the device according to the data of the infrared imager, visible light camera and laser rangefinder.

[0016] A vehicle-mounted anti-UAV intelligent optoelectronic system includes an intelligent target detection and recognition model based on the YOLOv5 algorithm. The intelligent target detection and recognition model includes:

[0017] Image input, used for pre-processing of various images. It is mainly composed of various image pre-processing modules, including tricks such as Mosaic data enhancement and adaptive anchor box calculation;

[0018] The backbone feature extraction network is mainly composed of a feature extraction network composed of different feature extraction modules nested together;

[0019] The neck network adopts the FPN+PAN structure. The combination of FPN layer and PAN layer solves the problem that when using neural network to extract features, the feature semantic information of the low layer is relatively small, but the target position is accurate, while the feature semantic information of the high layer is relatively rich, but the target position is relatively rough.

[0020] The result prediction model,uses the result prediction layer model of GIOU_loss regression position information,and obtains more accurate detection results through multi-scale prediction of the neck network output.

[0021] Preferably, the target detection and detection tasks mainly use the YOLOv5 algorithm. The YOLOv5 algorithm is the mainstream algorithm in the current target detection field. It follows the end-to-end training method of the YOLO series of algorithms, which makes its training and adding algorithm modules more convenient and quicker; compared with FasterR-CNN, the end-to-end training method of the YOLO series of algorithms is more convenient and quicker, and the image processing speed is greatly improved after removing the RPN method. Compared with the SSD algorithm, the YOLO series of algorithms has higher detection accuracy and faster processing speed. Under the hardware conditions of HiSilicon HI3559A, YOLOV5 can process 30 frames of images per second, which can meet the real-time detection requirements in the demand. On this basis, the present invention customizes different enhancement modules and training methods for each situation based on YOLOv5 according to different deployment locations and implementation functions.

[0022] Preferably, when pre-processing various images at the image input end, the following method is included:

[0023] A11: For different data sets, we set anchor frames of different lengths and widths, and output the predicted frame based on the initial anchor frame during image pre-processing. We then compare it with the groundtruth frame, calculate the difference between the two, and then reversely update and iterate the network parameters to provide several possible target frame sizes.

[0024] A12: The improved distance-based clustering algorithm K-means++ is used, and the distance is used as the similarity evaluation index to obtain the prior frame. The rules of the improved distance-based clustering algorithm K-means++ include: in the process of calculating the distance with K-means++, a maximum number of iterations is added in each round, such as selecting n key points once by roulette, and then selecting the cluster center points that can minimize the sum of squared distances in turn, and the principle of selecting n key points is: 2 + the number of cluster center points;

[0025] A13: Adaptively add the least amount of black borders to the original image, uniformly scale images of different sizes, and scale the original images of the same type to a standard size, reducing the computational complexity of the backbone feature network and increasing the speed of object detection.

[0026] Preferably, the backbone feature extraction module includes: a feature extraction module and an attention structure, wherein the feature extraction module includes a CBL joint feature extractor and a residual unit module; wherein the CBL joint feature extractor adopts a convolution layer using three technologies: Padding technology, Striding technology and batch normalization technology, and in the Striding technology, the leaky_ReLU function is used as the activation function, and the expression of the leaky_ReLU function is:

[0027]

[0028] The principle formula of batch normalization technology is:

[0029]

[0030] In addition, two parameters γ and β are added to the batch normalization technology to optimize the robustness of the principle formula of the batch normalization technology, and the batch normalization formula is obtained:

[0031]

[0032] The residual unit module takes the original input of the two CBL joint feature extractors as a layer of input and superimposes it with the normal output of the CBL module, and saves the original input results of multiple residual blocks. After the multi-layer calculation is completed, it is combined with the output results into a larger-scale feature map, which better ensures the preservation of various types of information in the feature map.

[0033] Preferably, the convolutional layer using the three technologies of Padding technology, Striding technology and batch normalization technology has the following advantages:

[0034] 1. Speed ​​up the training process, which can increase the learning rate. Even a small learning rate can achieve a fast learning rate.

[0035] 2. Don’t worry about the parameter selection of dropout and L2 regularization in fitting. The BN algorithm can save these two items or only require a small L2 regularization constraint. The reason is that after the BN algorithm, the parameters are normalized, and the distribution of neurons that were not greatly affected by the activation function becomes obvious. After an activation function, the neurons will automatically weaken or remove some neurons, so there is no need to dropout them. In addition, L2 regularization is normalized every time training, so it is rare that the parameters change too much due to different data distribution, resulting in a continuous increase in parameters.

[0036] 3. Shuffle the training data set to prevent training deviation.

[0037] Preferably, on top of these two basic structures is a CSP structure built with reference to CSP-DenseNet. The gradients from the residual unit module are integrated separately; on the other hand, the feature maps that do not pass through the residual unit module are also integrated separately; increase the gradient path: the number of gradient paths can be doubled through the split-merge strategy; due to the adoption of a cross-stage strategy, the disadvantages of using explicit feature map replication for connection can be alleviated, and the gradient information will not be reused because the gradient flow is truncated; by using a split and merge strategy across stages, the possibility of duplication in the information integration process can be effectively reduced; if the duplication of gradient information can be effectively reduced, the learning ability of the network will be greatly improved.

[0038] As a preferred method, the attention structure is used to slice the image before it enters the backbone. The specific operation is to get a value for every other pixel in an image, similar to neighboring downsampling, to obtain four images. The four images are complementary and have similar patterns, but no information is lost. Thus, the W and H information are concentrated in the channel space, and the input channel is expanded by 4 times, that is, the fused image becomes 12 channels relative to the original RGB three-channel mode. Finally, the new image is subjected to a convolution operation to obtain a double-downsampled feature map without information loss.

[0039] As a preference, when the neck network extracts the features of the image, the FPN layer conveys strong semantic features from top to bottom, while the PAN layer conveys strong positioning features from bottom to top, and features of different detection layers are aggregated from different backbone layers; when the result prediction model obtains a more accurate detection result by multi-scale prediction of the neck network output, CIOU_Loss is used as the loss function of the prediction layer for the frame position regression, where the formula of CIOU_Loss is:

[0040]

[0041] Where v is a parameter to measure the consistency of aspect ratio, which can also be defined as:

[0042]

[0043] Preferably, the target detection algorithm needs to go through the processes of feature design, data set establishment, model establishment, model training, and forward deployment from establishment to deployment; in order to achieve all-day (including day and night) and all-weather (including rainy, snowy, windy / foggy and other extreme weather environments), typical target data in visible, infrared, rainy, snowy, foggy, enhanced and other conditions are collected when establishing a data database, a special database is established, and the model is trained in a targeted manner to meet the target recognition needs all-weather and all-day.

[0044] Beneficial effects of the present invention:

[0045] 1. Compared with the existing technology, when tracking and measuring small drones stably, the miniaturized and lightweight optoelectronic system uses an FPGA+DSP optoelectronic system to track small drones, but the detection, tracking and identification capabilities are weak, the tracking of drones is unstable, and the target cannot be detected and identified. This system uses an intelligent target detection and recognition model based on the YOLOv5 algorithm. On the domestic HiSilicon processor Hi3559AV100, it can process 30 frames of images per second, meeting the real-time detection requirements during use, and has high detection accuracy, accurate results, and strong anti-interference ability, which can meet the target recognition requirements all day and all day;

[0046] 2. This solution uses HiSilicon's latest SOC chip, namely the domestic HiSilicon processor Hi3559AV100, which can not only complete the tracking algorithm, but also use the powerful computing power of the chip to achieve secondary recognition, capture and tracking of lost targets;

[0047] 3. The structural design of this scheme is of high precision, the main measuring devices are matched with the system, and the whole machine has a high angle measurement capability; at the same time, the servo and tracking system of this system is mature and reliable, which can realize the accurate tracking and measurement positioning of small targets at a long distance;

[0048] 4. This system adopts a miniaturized, lightweight, integrated and low-cost design. On the basis of a total weight of 30kg, it can achieve stable tracking and measurement of small UAV targets with a visible range of 2.5km and an infrared range of 1.5km. At the same time, it adopts an embedded installation form, which can sink the optoelectronic system into the vehicle body, reducing the risk of self-exposure;

[0049] 5. By adopting the convolutional layer using three technologies, namely Padding technology, Striding technology and batch normalization technology, the training speed can be accelerated and the learning rate can be increased. Even a small learning rate can have a fast learning rate. There is no need to pay attention to the parameter selection of dropout and L2 regularization terms in fitting. The BN algorithm can save these two items or only require a small L2 regularization constraint. The reason is that after the BN algorithm, the parameters are normalized, and the distribution of neurons that were not greatly affected by the activation function becomes obvious. After an activation function, the neurons will automatically weaken or remove some neurons, so there is no need to dropout them. In addition, L2 regularization, since normalization is performed every time training, it is rare that the parameters change too much due to different data distribution, resulting in a continuous increase in parameters; it has the advantage of disrupting the training data set to prevent training deviation. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Shown is a schematic diagram of the three-dimensional structure of the vehicle-mounted anti-UAV intelligent optoelectronic device of the present invention;

[0051] Figure 2 Shown is a schematic diagram of the internal three-dimensional structure of the vehicle-mounted anti-UAV intelligent optoelectronic device of the present invention;

[0052] Figure 3 Shown is a schematic diagram of the structure of the optoelectronic load component in the vehicle-mounted anti-UAV intelligent optoelectronic device of the present invention;

[0053] Figure 4 Shown is a schematic diagram of a control link of a servo control unit in a vehicle-mounted anti-UAV intelligent optoelectronic device of the present invention;

[0054] Figure 5 Shown is a schematic diagram of the control link of the vehicle-mounted anti-UAV intelligent optoelectronic device of the present invention;

[0055] Figure 6 What is shown is a schematic diagram of the network structure of the result prediction model in the vehicle-mounted anti-UAV intelligent optoelectronic system of the present invention;

[0056] Explanation of the reference numerals: 1. lower shell; 2. azimuth assembly; 3. pitch assembly; 4. middle frame; 5. front cover; 6. rear cover; 7. optoelectronic load assembly; 8. servo control assembly; 201. power filter; 202. azimuth motor; 203. azimuth encoder; 204. azimuth base; 301. U-shaped main frame; 302. side cover; 303. pitch motor; 304. pitch encoder; 701. infrared imager; 702. visible light camera; 703. inertial gyroscope; 704. laser rangefinder; 705. tracking and control unit; 801. power isolation unit; 802. servo control unit. DETAILED DESCRIPTION

[0057] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0058] See also Figure 1-2 The present invention provides an embodiment: a vehicle-mounted anti-UAV intelligent optoelectronic device, comprising a lower shell 1, an azimuth component 2, a pitch component 3, a middle frame 4, a front cover 5, a rear cover 6, an optoelectronic load component 7 and a servo control component 8, the azimuth component 2 is arranged above the lower shell 1, the pitch component 3 is arranged above the azimuth component 2, the middle frame 4 is arranged inside the pitch component 3, the front cover 5 is arranged on one side of the middle frame 4, the rear cover 6 is arranged on the other side of the middle frame 4, the optoelectronic load component 7 is arranged inside the middle frame 4, the servo component control component is arranged inside the pitch component 3, the lower shell 1 is made of hard aluminum alloy material, the middle frame 4 is made of hard aluminum alloy material, the front cover 5 is made of hard aluminum alloy material, and the rear cover 6 is made of hard aluminum alloy material.

[0059] Preferably, by setting the lower shell 1, the orientation component 2 can be installed and fixed, and by setting the orientation component 2, the pitch component 3 can be driven to rotate, that is, the azimuth can be adjusted. By setting the middle frame 4, the front cover 5 and the rear cover 6, the photoelectric load component 7 can be installed and fixed, and the structural weight reduction and heat dissipation design can be optimized. By setting the pitch component 3, the middle frame 4 can be driven to rotate, that is, the pitch angle can be adjusted, so that the orientation of the photoelectric load component 7 can be adjusted. By setting the photoelectric load component 7, uninterrupted monitoring and detection of the target can be achieved day and night, so that the orientation component 2 and the pitch component 3 can be used to drive the photoelectric load component 7 to adjust the azimuth and pitch angles, so as to achieve all-round uninterrupted monitoring and detection of the target day and night. By setting the servo control component 8, the orientation component 2 and the pitch component 3 can be adjusted according to the data of the photoelectric load component 7.

[0060] See also Figure 3-5 In this embodiment, the azimuth component 2 includes a power filter 201, an azimuth motor 202, an azimuth encoder 203 and an azimuth base 204. The power filter 201 is arranged inside the lower shell 1, the azimuth motor 202 is arranged above the lower shell 1, the power filter 201 is connected to the azimuth motor 202, the azimuth encoder 203 is arranged inside the lower shell 1, the azimuth encoder 203 is connected to the azimuth motor 202, the azimuth base 204 is arranged above the lower shell 1, the azimuth base 204 is arranged outside the azimuth motor 202, and the azimuth The base 204 is made of hard aluminum alloy material. When in use, the azimuth encoder 203 can be set to detect the rotation angle of the azimuth motor 202 so that the azimuth of the photoelectric load assembly 7 can be accurately fed back. The power supply filter 201 can be set to provide a stable power supply for the azimuth motor 202 to prevent the noise in the power supply from affecting the normal operation of the device. The azimuth base 204 can be set to install and fix the azimuth motor 202 to stabilize the supporting structure of the device. The azimuth motor 202 can be set to drive the pitch assembly 3 to rotate.

[0061] Preferably, the pitch assembly 3 includes a U-shaped main frame 301, a side cover 302, a pitch motor 303 and a pitch encoder 304. The U-shaped main frame 301 is arranged above the azimuth base 204, and the U-shaped main frame 301 is fixedly connected to the azimuth base 204. The side covers 302 are arranged in two groups, and the side covers 302 are arranged on opposite sides of the U-shaped main frame 301. The pitch motor 303 is arranged inside one of the two groups of side covers 302, and the pitch encoder 304 is arranged inside the other group of the two groups of side covers 302. The U-shaped main frame 301 is made of hard aluminum alloy material. The side cover 302 is made of hard aluminum alloy material, the output end of the azimuth motor 202 is connected to the U-shaped main frame 301, and the output end of the pitch motor 303 is connected to the middle frame 4. When in use, the side cover 302, the pitch motor 303, the pitch encoder 304 and the middle frame 4 can be installed by setting the U-shaped main frame 301, the pitch motor 303 and the pitch encoder 304 can be accommodated and sealed by setting the side cover 302, the pitch motor 303 can be set to drive the middle frame 4 to rotate, and the rotation angle of the middle frame 4 can be identified by setting the pitch encoder 304.

[0062] Preferably, the optoelectronic payload assembly 7 includes an infrared imager 701, a visible light camera 702, an inertial gyroscope 703, a laser rangefinder 704 and a tracking and control unit 705. The infrared imager 701 is arranged inside the middle frame 4, the visible light camera 702 is arranged on one side of the infrared imager 701, the inertial gyroscope 703 is arranged on one side of the infrared imager 701, the laser rangefinder 704 is arranged below the visible light camera 702, and the tracking and control unit 705 is arranged inside the middle frame 4. The inertial gyroscope 703 adopts a micromechanical gyroscope. The tracking and control unit 70 5 adopts the domestic HiSilicon processor Hi3559AV100. When in use, the infrared imager 701 can be set to collect infrared images, the visible light camera 702 can be set to collect high-definition images, the inertial gyroscope 703 can be set to assist the azimuth encoder 203 and the pitch encoder 304 to increase accuracy, the laser rangefinder 704 can be set to detect the distance of the target, and the tracking and control unit 705 can be set to control the infrared imager 701 and the visible light camera 702 according to the detection result of the laser rangefinder 704.

[0063] Preferably, the servo control assembly 8 includes a power isolation unit 801 and a servo control unit 802. The power isolation unit 801 is arranged inside the side cover 302, and the servo control unit 802 is arranged inside the side cover 302. The servo control unit 802 includes:

[0064] The initialization module is used to initialize the processor's internal data storage, registers, system clock, interrupt system, GPIO, RS422 and SPI communications when the system is powered on and reset;

[0065] The platform self-check module is used to check whether the positions of the inertial gyroscope 703, the azimuth encoder 203 and the pitch encoder 304 are correct, and whether the working angle ranges of the U-shaped main frame 301 and the middle frame 4 are correct;

[0066] The data communication processing module is used to send the azimuth and pitch angle information of the middle frame 4 to the main control module according to the data of the azimuth encoder 203, the data of the pitch encoder 304 and the speed data of the inertial gyroscope 703;

[0067] The servo control operation module is used for servo control operation. It mainly uses the corresponding infrared imager 701, visible light camera 702 or laser rangefinder 704 data to filter and calculate the servo control amount according to the instructions, outputs the current control instruction according to the calculated control amount, and sends the platform angle information back to the main control module;

[0068] The drive board communication module is used to realize the communication between the main control module and the azimuth motor 202 and the pitch motor 303, wherein the information includes: current control instructions, received current feedback signals, the movement of the azimuth motor 202 and the movement of the pitch motor 303;

[0069] The main control module is used to control the device according to the data of the infrared imager 701 , the visible light camera 702 and the laser rangefinder 704 .

[0070] See also Figure 6 The present invention provides an embodiment: a vehicle-mounted anti-UAV intelligent optoelectronic system, including an intelligent target detection and recognition model based on the YOLOv5 algorithm, the intelligent target detection and recognition model includes:

[0071] Image input, used for pre-processing of various images. It is mainly composed of various image pre-processing modules, including tricks such as Mosaic data enhancement and adaptive anchor box calculation;

[0072] The backbone feature extraction network is mainly composed of a feature extraction network composed of different feature extraction modules nested together;

[0073] The neck network adopts the FPN+PAN structure. The combination of FPN layer and PAN layer solves the problem that when using neural network to extract features, the feature semantic information of the low layer is relatively small, but the target position is accurate, while the feature semantic information of the high layer is relatively rich, but the target position is relatively rough.

[0074] The result prediction model,uses the result prediction layer model of GIOU_loss regression position information,and obtains more accurate detection results through multi-scale prediction of the neck network output.

[0075] Preferably, the target detection and detection tasks mainly use the YOLOv5 algorithm. The YOLOv5 algorithm is the mainstream algorithm in the current target detection field. It follows the end-to-end training method of the YOLO series of algorithms, which makes its training and adding algorithm modules more convenient and quicker; compared with FasterR-CNN, the end-to-end training method of the YOLO series of algorithms is more convenient and quicker, and the image processing speed is greatly improved after removing the RPN method. Compared with the SSD algorithm, the YOLO series of algorithms has higher detection accuracy and faster processing speed. Under the hardware conditions of HiSilicon HI3559A, YOLOV5 can process 30 frames of images per second, which can meet the real-time detection requirements in the demand. On this basis, the present invention customizes different enhancement modules and training methods for each situation based on YOLOv5 according to different deployment locations and implementation functions.

[0076] Preferably, when pre-processing various images at the image input end, the following method is included:

[0077] A11: For different data sets, we set anchor frames of different lengths and widths, and output the predicted frame based on the initial anchor frame during image pre-processing. We then compare it with the groundtruth frame, calculate the difference between the two, and then reversely update and iterate the network parameters to provide several possible target frame sizes.

[0078] A12: The improved distance-based clustering algorithm K-means++ is used, and the distance is used as the similarity evaluation index to obtain the prior frame. The rules of the improved distance-based clustering algorithm K-means++ include: in the process of calculating the distance with K-means++, a maximum number of iterations is added in each round, such as selecting n key points once by roulette, and then selecting the cluster center points that can minimize the sum of squared distances in turn, and the principle of selecting n key points is: 2 + the number of cluster center points;

[0079] A13: Adaptively add the least amount of black borders to the original image, uniformly scale images of different sizes, and scale the original images of the same type to a standard size, reducing the computational complexity of the backbone feature network and increasing the speed of object detection.

[0080] Preferably, the backbone feature extraction module includes: a feature extraction module and an attention structure, wherein the feature extraction module includes a CBL joint feature extractor and a residual unit module; wherein the CBL joint feature extractor adopts a convolution layer using three technologies: Padding technology, Striding technology and batch normalization technology, and in the Striding technology, the leaky_ReLU function is used as the activation function, and the expression of the leaky_ReLU function is:

[0081]

[0082] The principle formula of batch normalization technology is:

[0083]

[0084] In addition, two parameters γ and β are added to the batch normalization technology to optimize the robustness of the principle formula of the batch normalization technology, and the batch normalization formula is obtained:

[0085]

[0086] The residual unit module takes the original input of the two CBL joint feature extractors as a layer of input and superimposes it with the normal output of the CBL module, and saves the original input results of multiple residual blocks. After the multi-layer calculation is completed, it is combined with the output results into a larger-scale feature map, which better ensures the preservation of various types of information in the feature map.

[0087] Preferably, the convolutional layer using the three technologies of Padding technology, Striding technology and batch normalization technology has the following advantages:

[0088] 1. Speed ​​up the training process, which can increase the learning rate. Even a small learning rate can achieve a fast learning rate.

[0089] 2. Don’t worry about the parameter selection of dropout and L2 regularization in fitting. The BN algorithm can save these two items or only require a small L2 regularization constraint. The reason is that after the BN algorithm, the parameters are normalized, and the distribution of neurons that were not greatly affected by the activation function becomes obvious. After an activation function, the neurons will automatically weaken or remove some neurons, so there is no need to dropout them. In addition, L2 regularization is normalized every time training, so it is rare that the parameters change too much due to different data distribution, resulting in a continuous increase in parameters.

[0090] 3. Shuffle the training data set to prevent training deviation.

[0091] Preferably, on top of these two basic structures is a CSP structure built with reference to CSP-DenseNet. The gradients from the residual unit module are integrated separately; on the other hand, the feature maps that do not pass through the residual unit module are also integrated separately; increase the gradient path: the number of gradient paths can be doubled through the split-merge strategy; due to the adoption of a cross-stage strategy, the disadvantages of using explicit feature map replication for connection can be alleviated, and the gradient information will not be reused because the gradient flow is truncated; by using a split and merge strategy across stages, the possibility of duplication in the information integration process can be effectively reduced; if the duplication of gradient information can be effectively reduced, the learning ability of the network will be greatly improved.

[0092] As a preferred method, the attention structure is used to slice the image before it enters the backbone. The specific operation is to get a value for every other pixel in an image, similar to neighboring downsampling, to obtain four images. The four images are complementary and have similar patterns, but no information is lost. Thus, the W and H information are concentrated in the channel space, and the input channel is expanded by 4 times, that is, the fused image becomes 12 channels relative to the original RGB three-channel mode. Finally, the new image is subjected to a convolution operation to obtain a double-downsampled feature map without information loss.

[0093] As a preference, when the neck network extracts the features of the image, the FPN layer conveys strong semantic features from top to bottom, while the PAN layer conveys strong positioning features from bottom to top, and features of different detection layers are aggregated from different backbone layers; when the result prediction model obtains a more accurate detection result by multi-scale prediction of the neck network output, CIOU_Loss is used as the loss function of the prediction layer for the frame position regression, where the formula of CIOU_Loss is:

[0094]

[0095] Where v is a parameter to measure the consistency of aspect ratio, which can also be defined as:

[0096]

[0097] Preferably, the target detection algorithm needs to go through the processes of feature design, data set establishment, model establishment, model training, and forward deployment from establishment to deployment; in order to achieve all-day (including day and night) and all-weather (including rainy, snowy, windy / foggy and other extreme weather environments), typical target data in visible, infrared, rainy, snowy, foggy, enhanced and other conditions are collected when establishing a data database, a special database is established, and the model is trained in a targeted manner to meet the target recognition needs all-weather and all-day.

[0098] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments, and various changes can be made within the knowledge scope of those skilled in the art without departing from the spirit of the present invention.

Claims

1. A vehicle-mounted anti-UAV intelligent optoelectronic device, comprising a lower housing (1); characterized in that: The invention also comprises an azimuth component (2), a pitch component (3), a middle frame (4), a front cover (5), a rear cover (6), an optoelectronic load component (7) and a servo control component (8); the azimuth component (2) is arranged above the lower shell (1); the pitch component (3) is arranged above the azimuth component (2); the middle frame (4) is arranged inside the pitch component (3); the front cover (5) is arranged on one side of the middle frame (4); the rear cover (6) is arranged on the other side of the middle frame (4); the optoelectronic load component (7) is arranged inside the middle frame (4); the servo control component is arranged inside the pitch component (3); the lower shell (1) is made of a hard aluminum alloy material; the middle frame (4) is made of a hard aluminum alloy material; the front cover (5) is made of a hard aluminum alloy material; and the rear cover (6) is made of a hard aluminum alloy material.

2. The vehicle-mounted anti-UAV intelligent optoelectronic device according to claim 1 is characterized in that: The azimuth assembly (2) comprises a power filter (201), an azimuth motor (202), an azimuth encoder (203) and an azimuth base (204); the power filter (201) is arranged inside the lower shell (1); the azimuth motor (202) is arranged above the lower shell (1); the power filter (201) is connected to the azimuth motor (202); the azimuth encoder (203) is arranged inside the lower shell (1); the azimuth encoder (203) is connected to the azimuth motor (202); the azimuth base (204) is arranged above the lower shell (1); the azimuth base (204) is arranged outside the azimuth motor (202); and the azimuth base (204) is made of a hard aluminum alloy material.

3. The vehicle-mounted anti-UAV intelligent optoelectronic device according to claim 2 is characterized in that: The pitch assembly (3) comprises a U-shaped main frame (301), a side cover (302), a pitch motor (303) and a pitch encoder (304). The U-shaped main frame (301) is arranged above the azimuth base (204). The U-shaped main frame (301) is fixedly connected to the azimuth base (204). The side covers (302) are arranged in two groups. The side covers (302) are arranged on two opposite sides of the U-shaped main frame (301). The pitch motor (303) is arranged inside one of the two groups of side covers (302). The pitch encoder (304) is arranged inside the other of the two groups of side covers (302). The U-shaped main frame (301) is made of hard aluminum alloy material. The side covers (302) are made of hard aluminum alloy material. The output end of the azimuth motor (202) is connected to the U-shaped main frame (301). The output end of the pitch motor (303) is connected to the middle frame (4).

4. The vehicle-mounted anti-UAV intelligent optoelectronic device according to claim 3 is characterized by: The photoelectric load component (7) comprises an infrared imager (701), a visible light camera (702), an inertial gyroscope (703), a laser rangefinder (704) and a tracking and control unit (705). The infrared imager (701) is arranged inside the middle frame (4), the visible light camera (702) is arranged on one side of the infrared imager (701), the inertial gyroscope (703) is arranged on one side of the infrared imager (701), the laser rangefinder (704) is arranged below the visible light camera (702), the tracking and control unit (705) is arranged inside the middle frame (4), the inertial gyroscope (703) adopts a micromechanical gyroscope, and the tracking and control unit (705) adopts a domestic HiSilicon processor Hi3559AV100.

5. The vehicle-mounted anti-UAV intelligent optoelectronic device according to claim 4 is characterized in that: The servo control assembly (8) includes a power isolation unit (801) and a servo control unit (802). The power isolation unit (801) is arranged inside the side cover (302). The servo control unit (802) is arranged inside the side cover (302). The servo control unit (802) includes: An initialization module is used to initialize the processor's internal data storage, registers, system clock, interrupt system, GPIO, RS (422) and SPI communication when the system is powered on and reset; A platform self-check module, used to detect whether the positions of the inertial gyroscope (703), the azimuth encoder (203) and the pitch encoder (304) are correct, and whether the working angle ranges of the U-shaped main frame (301) and the middle frame (4) are correct; A data communication processing module is used to send the azimuth and pitch angle information of the middle frame (4) to the main control module according to the azimuth encoder (203) data, the pitch encoder (304) data and the speed data of the inertial gyroscope (703); A servo control operation module is used for servo control operation. It mainly performs filtering and servo control quantity operation based on instructions using corresponding infrared imager (701), visible light camera (702) or laser rangefinder (704) data, outputs current control instructions based on the calculated control quantity, and sends back platform angle information to the main control module; A drive board communication module, used to enable the main control module to communicate with the azimuth motor (202) and the pitch motor (303), wherein the information includes: current control instructions, received current feedback signals, actions of the azimuth motor (202) and the pitch motor (303); The main control module is used to control the device according to the data of the infrared imager (701), the visible light camera (702) and the laser rangefinder (704).

6. A vehicle-mounted anti-UAV intelligent optoelectronic system, characterized by: Intelligent target detection and recognition model based on YOLOv5 algorithm, which includes: The image input end is used to pre-process various images. It is mainly composed of various image pre-processing modules, including tricks such as Mosaic data enhancement and adaptive anchor frame calculation. The backbone feature extraction network is mainly composed of a feature extraction network composed of different feature extraction modules. The neck network adopts the FPN+PAN structure. The combination of FPN layer and PAN layer solves the problem that when using neural network to extract features, the feature semantic information of the low layer is relatively small, but the target position is accurate, while the feature semantic information of the high layer is relatively rich, but the target position is relatively rough. The result prediction model,uses the result prediction layer model of GIOU_loss regression position information,and obtains more accurate detection results through multi-scale prediction of the neck network output.

7. The vehicle-mounted anti-UAV intelligent optoelectronic system according to claim 6 is characterized by: When pre-processing various images at the image input end, the following methods are included: A11: For different data sets, we set anchor frames of different lengths and widths, and output the predicted frame based on the initial anchor frame during image pre-processing. We then compare it with the groundtruth frame, calculate the difference between the two, and then reversely update and iterate the network parameters to provide several possible target frame sizes. A12: The improved distance-based clustering algorithm K-means++ is used, and the distance is used as the similarity evaluation index to obtain the prior frame. The rules of the improved distance-based clustering algorithm K-means++ include: in the process of calculating the distance with K-means++, a maximum number of iterations is added in each round, such as selecting n key points once by roulette, and then selecting the cluster center points that can minimize the sum of squared distances in turn, and the principle of selecting n key points is: 2 + the number of cluster center points; A13: Adaptively add the least amount of black borders to the original image, uniformly scale images of different sizes, and scale the original images of the same type to a standard size, reducing the computational complexity of the backbone feature network and increasing the speed of object detection.

8. The vehicle-mounted anti-UAV intelligent optoelectronic system according to claim 7 is characterized by: The backbone feature extraction module includes: a feature extraction module and an attention structure, wherein the feature extraction module includes a CBL joint feature extractor and a residual unit module; wherein the CBL joint feature extractor uses a convolutional layer using three technologies: Padding technology, Striding technology, and batch normalization technology, and in the Striding technology, the leaky_ReLU function is used as the activation function, and the expression of the leaky_ReLU function is: The principle formula of batch normalization technology is: In addition, two parameters γ and β are added to the batch normalization technology to optimize the robustness of the principle formula of the batch normalization technology, and the batch normalization formula is obtained: The residual unit module takes the original input of the two CBL joint feature extractors as a layer of input and superimposes it with the normal output of the CBL module, and saves the original input results of multiple residual blocks. After the multi-layer calculation is completed, it is combined with the output results into a larger-scale feature map, which better ensures the preservation of various types of information in the feature map.

9. The vehicle-mounted anti-UAV intelligent optoelectronic system according to claim 8 is characterized by: The attention structure is used to slice the image before it enters the backbone. The specific operation is to get a value for every other pixel in an image, similar to neighboring downsampling, to obtain four images. The four images are complementary and have similar patterns, but no information is lost. Thus, the W and H information are concentrated in the channel space, and the input channel is expanded by 4 times, that is, the fused image becomes 12 channels relative to the original RGB three-channel mode. Finally, the new image is subjected to a convolution operation to obtain a double-downsampled feature map without information loss.

10. The vehicle-mounted anti-UAV intelligent optoelectronic system according to claim 9 is characterized in that: When the neck network extracts the features of the image, the FPN layer conveys strong semantic features from top to bottom, while the PAN layer conveys strong positioning features from bottom to top, and aggregates features from different backbone layers to different detection layers; when the result prediction model obtains more accurate detection results through multi-scale prediction of the neck network output, CIOU_Loss is used as the loss function of the prediction layer for the frame position regression, where the formula of CIOU_Loss is: Where v is a parameter to measure the consistency of aspect ratio, which can also be defined as:

Images