High-precision laser glare closed-loop control system based on deep learning vision algorithm

Abstract

The application relates to the technical field of photoelectric tracking and confrontation, and discloses a high-precision laser dazzling closed-loop control system based on a deep learning vision algorithm, which comprises a laser dazzling rejection unit and a high-precision servo tracking device, adopts a double-rate asynchronous closed-loop architecture, a deep learning vision processing unit outputs target observation data with absolute time stamps in a slow loop, a high-precision servo controller executes high-frequency deduction based on inertial data in a fast loop, utilizes a historical state buffer queue and a forward reintegration algorithm to perform space-time alignment measurement update, and eliminates vision delay; the system further establishes an adaptive region of interest adjustment mechanism based on interaction of a prediction covariance matrix, and integrates a reverse PWM trigger exposure control, a dynamic geographic fence verification and a power nonlinear modulation function. The application effectively solves the tracking lag problem caused by vision delay under high dynamic conditions, and improves the tracking precision, robustness and operation safety of the system.

Description

Technical Field

[0001] This invention relates to the field of photoelectric tracking and countermeasures technology, specifically a high-precision laser dazzle closed-loop control system based on deep learning vision algorithms. Background Technology

[0002] Laser dazzling systems, as a non-lethal optical deterrent, are widely used in low-altitude security, border control, and key area protection. With increasingly complex application environments, traditional visual tracking algorithms based on brightness thresholds, edge features, or correlation filtering often struggle to maintain stable tracking performance when faced with complex background interference, target occlusion, and drastic changes in lighting. While introducing deep learning target detection algorithms improves the system's ability to identify and resist interference from low-speed, small targets and targets in complex backgrounds, it also presents new challenges to real-time closed-loop control.

[0003] The large number of parameters and computational complexity of deep neural network models result in significant computational time consumption for vision processing units when performing target detection inference. On high-speed moving platforms or when facing highly maneuverable targets, this non-deterministic processing delay leads to a large time difference between image acquisition and control response. Existing servo control systems typically assume that feedback data is real-time or has only a small fixed delay. Directly using lagging visual observation data for position closure introduces severe phase lag, which in turn increases the dynamic tracking error of the servo system and may even cause system oscillations and target loss.

[0004] Furthermore, existing systems typically employ fixed image acquisition modes, making it impossible to dynamically adjust sensor parameters based on the confidence level of the tracking status. In full-frame high-resolution processing mode, the large data throughput limits the visual feedback frequency, making it difficult to meet the high bandwidth requirements of fast control loops; while simply reducing resolution results in the loss of detailed features of distant targets. Simultaneously, during laser dazzle operations, the backscattering of the emitted high-energy laser on airborne particles or optical windows can easily saturate the imaging sensor, causing visual blindness and affecting continuous tracking capabilities. Regarding safety, traditional geofencing or power control strategies often rely on current static location judgments, lacking dynamic prediction of system mechanical delays and target movement trends, posing risks of misfires or delayed protection. Therefore, how to achieve high-bandwidth, high-precision, spatiotemporally consistent tracking while introducing significant latency with deep learning algorithms, and how to collaboratively address adaptive perception and active safety issues, are pressing technical challenges in this field. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a high-precision closed-loop control system for laser glare based on deep learning visual algorithms. This system solves the technical defects of existing laser glare systems, such as tracking lag and system instability caused by computational delays while improving recognition robustness by introducing deep learning algorithms, as well as difficulties in adaptively balancing visual acquisition modes, laser scattering interference imaging, and lack of dynamic prediction in safety protection mechanisms.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solution: a high-precision laser dazzle closed-loop control system based on deep learning vision algorithms. This system aims to solve the tracking accuracy and safety problems of laser dazzle equipment under high dynamic conditions, under the influence of factors such as visual delay, target maneuvering, and its own line-of-sight jitter.

[0007] In terms of hardware architecture, the system consists of a signal transmission unit connecting a laser glare rejection unit and a high-precision servo tracking device. The laser glare rejection unit integrates high-definition video image acquisition, deep learning vision processing, and laser emission functions. It is physically mounted on the high-precision servo tracking device as an end load. The high-precision servo tracking device includes a servo controller, a two-dimensional motion platform, and an inertial measurement unit fixed to the rotating base of the platform.

[0008] In terms of control architecture, the system constructs a dual-rate asynchronous closed-loop control model, which is divided into a slow perception loop and a fast control loop. The deep learning vision processing unit runs in the slow perception loop, which is responsible for target detection and inference on image data and outputs target observation data containing absolute acquisition timestamps. The high-precision servo controller runs in the fast control loop, which reads the three-axis angular velocity data of the inertial measurement unit at fixed control cycles and performs high-frequency kinematic state deduction. To address the time delay caused by vision processing, after receiving the target observation data, the high-precision servo controller calculates the lag time between the current moment and the absolute acquisition timestamp and performs spatiotemporally aligned asynchronous measurement updates to obtain the final state estimate of the two-dimensional motion platform.

[0009] Furthermore, to balance computing resources and tracking performance, the system establishes a state covariance matrix interaction mechanism between the deep learning vision processing unit and the high-precision servo controller. The high-precision servo controller feeds back the prediction covariance matrix generated during the state deduction process to the deep learning vision processing unit in real time. The vision processing unit reads the matrix and extracts the diagonal elements in the azimuth and pitch directions. These values ​​represent the uncertainty of the current tracking state. Based on this, the vision processing unit calculates the region of interest size required for the next frame and writes it to the register interface of the image acquisition unit. This mechanism allows the system to reduce the field of view to increase the frame rate when tracking is stable, and expand the field of view to prevent target loss when uncertainty increases, thus achieving adaptive adjustment of resolution and frame rate.

[0010] In terms of the specific implementation of state estimation, the high-precision servo controller is equipped with a historical state buffer queue, which stores the state estimate, covariance matrix and inertial data of historical time according to the control cycle. When the target observation data with time delay arrives, the controller retrieves the historical record with the closest timestamp in the buffer queue, calculates the Kalman gain using the observation data, historical covariance matrix and historical state estimate, and performs a posteriori correction on the historical state. Then, the system takes the corrected historical state as the starting point, extracts all historical inertial data from the acquisition time to the current time, performs forward reintegration using the state transition equation, and recursively pushes the delayed measurement information to the current time, thus eliminating the impact of visual processing delay on the closed-loop bandwidth.

[0011] To address situations where the target is temporarily lost or occluded, the system incorporates a sensor mode switching mechanism. When invalid visual data is detected, the system enters inertial flywheel mode, cutting off the visual measurement update loop and relying solely on the inertial measurement unit to maintain the inertial deduction of the state vector. Simultaneously, the system superimposes blind zone process noise onto the predicted covariance matrix and performs dilation processing to accurately reflect the range of estimation errors accumulated over time. Once the target is recaptured, the system converges rapidly through the expanded search area.

[0012] To improve the robustness of target recognition, the deep learning target detection model includes an independent embedding branch that maps deep semantic features of the image into a unitized feature vector. During the target recapture stage, the cosine similarity of the feature vectors is calculated to verify identity consistency. In addition, the system introduces a kinematic constraint filter that calculates the theoretically reachable region based on the target's maximum maneuverability and eliminates false detection boxes that violate kinematic laws.

[0013] In terms of anti-interference and electromagnetic compatibility design, the system utilizes the pulse width modulation characteristics of the laser emission unit's driving signal. An inverted signal is transmitted through the signal transmission unit as the external trigger shutter signal for the image acquisition unit. Image acquisition only involves exposure integration during the laser pulse off-state interval, and the maximum integration time is limited by the difference between the laser off-state duration and the safety protection interval. This timing control physically eliminates the interference of laser backscattering on the imaging sensor, ensuring all-weather operation capability.

[0014] In terms of safety and performance control, the system generates power modulation commands based on the final state estimate. The power duty cycle is inversely proportional to the pixel area of ​​the target in the image. That is, the power is reduced when the target is close (large pixel area) and increased when the target is far away (small pixel area), balancing glare effect and eye safety. At the same time, the system calculates the aiming error and allows light to be emitted only when the error is less than the effective glare threshold. In addition, the system performs geofence security verification, using the absolute encoder feedback angle combined with the current angular velocity to calculate the dynamic prediction angle including the total cut-off delay of the system. Once the prediction angle touches the no-fire zone, the transmission circuit is cut off to prevent accidental firing due to system delay. Meanwhile, the system corrects the thermal deformation and installation error of the optomechanical structure through temperature sensor data to ensure the axis pointing accuracy under long-term operation. A hardware heartbeat monitoring mechanism is also established to forcibly trigger the underlying safety shutdown when communication is abnormal.

[0015] This invention provides a high-precision closed-loop control system for laser glare based on deep learning vision algorithms. It has the following advantages:

[0016] 1. This invention effectively decouples the frequency limitations of the visual perception loop and the servo control loop by constructing a dual-rate asynchronous closed-loop architecture and a historical state buffer re-integration mechanism. The high-precision servo controller does not directly drive the motor using the lagging visual data, but instead uses observation data with absolute timestamps to backtrack and correct the historical state, and combines it with the high-frequency inertial data stored in the inertial measurement unit to perform forward re-integration, recursively pushing the state to the current moment. This spatiotemporal alignment processing method eliminates the phase lag caused by the time consumption of deep learning algorithms, enabling the system to maintain high-bandwidth stable tracking when facing high-speed maneuvering targets, avoiding the oscillation or loss phenomenon common in traditional visual servo systems.

[0017] 2. This invention utilizes a state covariance matrix interaction mechanism to establish a dynamic correlation between control uncertainty and visual acquisition parameters. The deep learning visual processing unit adjusts the region of interest size and frame rate of image acquisition in real time based on the prediction covariance matrix fed back by the servo controller. When tracking is stable and the error covariance is small, the system automatically reduces the field of view to obtain a higher processing frame rate and improve dynamic performance. When the target maneuver causes the prediction error to diverge, the system automatically expands the field of view to reduce the probability of target loss. This mechanism maximizes the system's efficiency with limited hardware computing power, taking into account both the requirements of search field of view coverage and precise tracking refresh rate.

[0018] 3. In terms of hardware, this invention employs an inverse PWM signal to trigger exposure control, ensuring that image acquisition strictly avoids the laser emission window. This completely eliminates the glare interference from laser backscattering on the visual sensor from the physical optical path, guaranteeing all-weather imaging clarity. In terms of safety, the system integrates nonlinear modulation of laser power based on pixel area, dynamic geofence verification including system delay compensation, and identity consistency detection based on depth features. These measures effectively prevent mis-projection and collateral damage caused by misidentification, system delay, or out-of-range movement, ensuring that laser energy projection is always in a controllable, effective, and safety-compliant state. Attached Figure Description

[0019] Figure 1 This is a block diagram of the overall hardware composition of the system of the present invention;

[0020] Figure 2 This is a flowchart of the control method of the present invention;

[0021] Figure 3 This is a schematic diagram comparing the system tracking performance trajectories of the present invention;

[0022] Figure 4 This is a schematic diagram illustrating the real-time tracking error comparison of the present invention;

[0023] Figure 5 This is a schematic diagram of the adaptive modulation of laser power according to the present invention.

[0024] Among them, 100 is a laser glare rejection unit; 110 is a high-definition video image acquisition unit; 120 is an intelligent vision processing unit; 130 is a deep learning vision processing unit; 200 is a signal transmission unit; 300 is a high-precision servo tracking device; 310 is a high-precision servo controller; 320 is a high-precision two-dimensional motion platform; 321 is an azimuth motor; 322 is a pitch motor; 323 is an azimuth encoder; 324 is a pitch encoder; and 330 is an inertial measurement unit. Detailed Implementation

[0025] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see the appendix Figure 1 The present invention provides a high-precision laser glare closed-loop control system based on deep learning vision algorithms, including a laser glare rejection unit 100, a signal transmission unit 200 and a high-precision servo tracking device 300.

[0027] The laser glare rejection unit 100 is physically installed at the end load position of the high-precision servo tracking device 300. The laser glare rejection unit 100 integrates a high-definition video image acquisition unit 110, an intelligent vision processing unit 120, a deep learning vision processing unit 130, a system control unit 140, and a laser emission unit 150.

[0028] The high-definition video image acquisition unit 110 adopts a CMOS industrial image sensor that supports windowed reading mode. The image sensor is equipped with a register interface and can receive external commands to change the starting coordinates and length and width of the effective pixel readout area. The optical axis of the high-definition video image acquisition unit 110 and the emission optical axis of the laser emission unit 150 are kept parallel through mechanical calibration.

[0029] The intelligent vision processing unit 120 is communicatively connected to the high-definition video image acquisition unit 110, and is used to receive the raw video data stream and perform noise reduction and image enhancement preprocessing. The deep learning vision processing unit 130 is connected to the intelligent vision processing unit 120, and internally stores and runs the YOLOv8 deep learning object detection model. The deep learning vision processing unit 130 has the function of calculating the parameters of the region of interest based on the state covariance matrix described later, and can send instructions to the high-definition video image acquisition unit 110 to adjust the resolution and frame rate of image acquisition.

[0030] The system control unit 140 is connected to the deep learning vision processing unit 130 and the laser emission unit 150 respectively, and is responsible for protocol encapsulation of the miss data and controlling the emission state of the laser beam.

[0031] The signal transmission unit 200 uses an aviation plug assembly 210 with an electromagnetic shielding layer to connect the laser glare rejection unit 100 and the high-precision servo tracking device 300. The aviation plug assembly 210 is used to transmit differential signals and power supply current conforming to the RS422 standard between the two units.

[0032] The high-precision servo tracking device 300 includes a high-precision servo controller 310, a high-precision two-dimensional motion platform 320, and an inertial measurement unit 330. The high-precision servo controller 310 uses an embedded real-time microprocessor and internally runs a high-frequency kinematic state inference algorithm and a Kalman filter algorithm.

[0033] The high-precision two-dimensional motion platform 320 includes an azimuth motor 321, a pitch motor 322, an azimuth encoder 323, and a pitch encoder 324. The azimuth motor 321 and the pitch motor 322 are used to drive the platform to rotate in the horizontal and vertical directions, respectively. The azimuth encoder 323 and the pitch encoder 324 are used to collect the rotation angle of the platform in real time and feed the angle signal back to the high-precision servo controller 310.

[0034] The inertial measurement unit 330 is rigidly fixed on the rotating seat of the high-precision two-dimensional motion platform 320 and remains relatively stationary with the laser glare rejection unit 100. The inertial measurement unit 330 is used to collect the three-axis angular velocity data of the platform in inertial space and send the data to the high-precision servo controller 310 at a high-frequency sampling rate.

[0035] Please see the appendix Figure 2 Based on the aforementioned hardware components, this system 10 is logically constructed as a dual-rate asynchronous closed-loop control architecture. This architecture includes a slow sensing loop and a fast control loop.

[0036] The slow-speed perception loop is mainly composed of a high-definition video image acquisition unit 110 and a deep learning vision processing unit 130. The operating cycle of this loop is not fixed and depends on the current region of interest size setting of the high-definition video image acquisition unit 110. This loop is responsible for outputting target visual observation values ​​with absolute timestamps.

[0037] The fast control loop mainly consists of a high-precision servo controller 310, a high-precision two-dimensional motion platform 320, and an inertial measurement unit 330. The loop operates with a fixed high-frequency control cycle. The high-precision servo controller 310 uses the data from the inertial measurement unit 330 to perform kinematic state deduction and drives the azimuth motor 321 and the pitch motor 322 according to the state estimate.

[0038] The high-precision servo controller 310 establishes data interaction with the deep learning vision processing unit 130 through the state covariance matrix. The state covariance matrix represents the uncertainty of the current system's estimation of the target position. The deep learning vision processing unit 130 reads the matrix data and determines the size of the acquisition area that the high-definition video image acquisition unit 110 should use in the next frame through reverse calculation, thereby realizing the dynamic coupling between the perception strategy and the control state.

[0039] After the system is powered on, the initialization program is executed. The high-precision servo controller 310 sets the internal state vector and state covariance matrix to the initial values ​​and sends a reset command to the laser glare rejection unit 100. The high-definition video image acquisition unit 110 uses the image reading mode for the full-resolution panoramic search mode and sets the initial frame rate and maximum field of view.

[0040] The panoramic target capture and tracking system is established. The high-definition video image acquisition unit 110 acquires full-resolution image data and transmits it to the deep learning vision processing unit 130. The deep learning vision processing unit 130 runs the YOLOv8 model to infer the image data, identify the target object, and extract the center position coordinates of the target in the image coordinate system. These coordinates are transmitted to the high-precision servo controller 310 to complete the state initialization of the Kalman filter. The system then enters the dual-rate asynchronous closed-loop control stage.

[0041] The high-precision servo controller 310 performs kinematic state deduction with a fixed high-frequency control cycle. In each control cycle, the high-precision servo controller 310 reads the three-axis angular velocity data output by the inertial measurement unit 330, combines it with the state estimate value of the previous moment, and uses the kinematic equation to calculate the target state prediction value and the state covariance matrix prediction value at the current moment. During this process, the high-precision servo controller 310 detects the state of the sensor mode switching flag. If the flag is valid, it determines that the system is in the blind zone switching stage. The high-precision servo controller 310 automatically enters the inertial flywheel mode, maintains the inertial deduction of the state vector only based on the data of the inertial measurement unit 330, and suspends the execution of measurement update operations based on visual data. At the same time, blind zone process noise is superimposed on the state covariance matrix.

[0042] To implement a hierarchical hysteresis perception strategy based on predictive uncertainty, the deep learning vision processing unit 130 reads the current state covariance matrix of the high-precision servo controller 310 through the communication interface before preparing to acquire the next frame of image. Based on the diagonal elements of the matrix, it calculates the ideal region of interest size required for the next frame of image. The deep learning vision processing unit 130 compares the ideal size with the currently set image acquisition level. If the ideal size exceeds the hysteresis threshold range corresponding to the current level, it generates a sensor mode switching command, sets the sensor mode switching flag, and controls the high-definition video image acquisition unit 110 to modify the register configuration to switch to the new region of interest level. If the ideal size does not exceed the hysteresis threshold range, it keeps the current image acquisition level unchanged.

[0043] Asynchronous measurement updates with spatiotemporal alignment are performed. When the high-definition video image acquisition unit 110 completes image acquisition and outputs target observation data with absolute acquisition timestamps via the deep learning vision processing unit 130, the high-precision servo controller 310 receives the observation data and calculates the lag time between the current time and the acquisition timestamp. Based on the lag time, the high-precision servo controller 310 retrieves the state prediction value and state covariance matrix of the corresponding time from the historical state buffer queue, calculates the Kalman gain using the observation data and corrects the historical state, and then performs forward reintegration based on the corrected historical state and the historical data of the inertial measurement unit 330 to obtain the final state estimate value at the current time.

[0044] The closed-loop servo drive is executed. The high-precision servo controller 310 converts the azimuth and pitch components in the final state estimate obtained in step S500 into pulse width modulation signals, which drive the azimuth motor 321 and the pitch motor 322 to rotate the high-precision two-dimensional motion platform 320, so that the emission optical axis of the laser emission unit 150 is aligned with the target object in real time. The system continues to execute the above steps in a loop until a stop command is received or the target is lost.

[0045] This embodiment mainly includes a laser glare rejection unit 100, a signal transmission unit 200, and a high-precision servo tracking device 300.

[0046] The laser glare denial unit 100, serving as both the sensing front-end and execution end of the system, integrates photoelectric imaging, edge computing, and laser emission functions. The high-definition video image acquisition unit 110 employs an image sensor with addressable pixel array readout capabilities. This sensor allows setting parameters of the region of interest (ROI) through configuration of internal registers. In this embodiment, utilizing the physical characteristics of the sensor readout mechanism—that is, the frame rate is inversely proportional to the number of vertical readout rows—the frame rate of the image sensor increases accordingly when the vertical size of the ROI decreases, thereby achieving high-frequency window acquisition. Simultaneously, the automatic exposure statistics area of ​​the image sensor is used to dynamically follow the ROI, i.e., exposure parameters are calculated solely based on the pixel brightness distribution within the ROI. This ensures that when the field of view switches from a wide-angle background to a local target, even if the background brightness differs significantly from the target brightness (e.g., a dark drone against a bright sky background), the target subject maintains optimal exposure, preventing feature loss due to insufficient dynamic range. Furthermore, to prevent the image sensor from being saturated by the backscattered spot generated by the high-energy laser beam emitted by the laser glare denial unit in the near-field atmosphere, the system employs a microsecond-level laser imaging timing synchronization mechanism.

[0047] The signal transmission unit 200 inverts the PWM modulation signal of the laser and uses it as the external trigger shutter signal for the image sensor. This forces the sensor to perform exposure integration only during the laser pulse's off-state interval, thereby completely shielding the laser from visual perception at the physical level and ensuring that target features are clearly visible during continuous glare. Specifically, the automatic exposure control loop of the image sensor incorporates saturation clamping logic based on the laser modulation parameters, setting the maximum allowable integration time. ;

[0048] in, This is the current laser pulse period. Duty cycle, A time-safe protection interval (e.g., 50μs) is reserved; when insufficient light causes the theoretical exposure requirement to exceed... At that time, the system forcibly truncates the exposure time to... Simultaneously, the analog gain is increased, sacrificing some signal-to-noise ratio to ensure that the imaging timing never overlaps with the laser emission window.

[0049] The intelligent vision processing unit 120 is connected to the back end of the acquisition unit and is responsible for preprocessing the raw data and outputting a digital video stream. The deep learning vision processing unit 130 serves as the edge computing core. During this process, the intelligent vision processing unit 120 executes a digital image stabilization algorithm based on sparse optical flow or feature point matching. This algorithm calculates the global affine transformation matrix between adjacent frames and performs inverse geometric distortion correction on the image to digitally eliminate residual high-frequency micro-vibrations that exceed the mechanical bandwidth limit of the servo mechanism. This ensures that no matter how severe the carrier vibration is, the image input to the deep learning model always maintains clear texture and stable background, maximizing the signal-to-noise ratio of the target features. A deep learning target detection algorithm model is deployed. This unit includes a preprocessing scaling module that normalizes and scales the cropped ROI image of arbitrary resolution to the fixed input tensor size of the model. The final output includes the target category probability and the center coordinates of the bounding box. and width and height dimensions Tensor data;

[0050] To ensure that the vision loop meets high-frequency real-time requirements (e.g., inference latency less than 10ms), the final output includes the target class probability and the bounding box center coordinates. Width and height dimensions The model also includes tensor data of a high-dimensional appearance feature description vector. Specifically, the deep learning model employs a lightweight backbone network based on CSPDarknet or ShuffleNet, and constructs a decoupled prediction head after the Neck layer. In addition to the conventional regression branch for predicting bounding box coordinates and the classification branch for predicting class probabilities, an independent embedding branch is designed in parallel. This embedding branch contains convolutional layers and global average pooling layers to map deep semantic features into unitized feature vectors of length 128 or 256. During model training, a joint loss function of detection loss and metric learning loss is used for end-to-end optimization, thereby ensuring that target localization and feature extraction are completed simultaneously in a single forward inference, without the need to run an additional recognition network serially.

[0051] The joint loss function is specifically defined as follows:

[0052] ;

[0053] in, To account for the bounding box regression loss that takes into account the overlapping area, center point distance, and aspect ratio, Classification loss based on confidence level; The triplet loss is used to constrain the feature space of the embedding branch, and its calculation formula is as follows: Anchor point samples (A) of the same target are forced to have a distance in feature space from positive samples (P) that is less than the distance from negative samples (N) by at least one interval. This gives the model the ability to distinguish different appearance targets at the feature level;

[0054] This feature vector is used to verify identity consistency by calculating cosine similarity after target loss and recapture, preventing erroneous target switching after multi-target interference or temporary occlusion. The deep learning target detection algorithm model undergoes INT8 or FP16 precision model quantization before deployment to prevent erroneous target switching after multi-target interference or temporary occlusion. Simultaneously, a kinematic constraint filter based on target type is introduced. This filter calculates the theoretically reachable region of the detection box in the current frame based on the target state at the previous moment and the preset maximum target maneuverability (e.g., the maximum acceleration G value of a drone). The specific mathematical definition of this region is a circular area with the predicted target position as the center and the maximum physical displacement as the radius. The criterion inequality is as follows:

[0055] ;

[0056] in, The detection coordinates are the output of the neural network in the current frame. These are the prior predicted coordinates of the Kalman filter. The target velocity modulus is currently estimated. This is the inter-frame time interval. The preset target maximum acceleration physical limit (e.g., 5g). To account for the redundancy tolerance of measurement noise;

[0057] If the center of the detection box output by the deep learning model exceeds the physically reachable area, the system will determine it as an outlier and remove it, even if the confidence level is high. The deep learning object detection algorithm model undergoes model quantization processing with INT8 or FP16 precision before deployment and runs on a high-performance inference engine dedicated to the neural processing unit, outputting the computation results deterministically with low latency.

[0058] The target solution logic executed by the deep learning vision processing unit 130 is as follows:

[0059] Obtain the current frame image data and its corresponding absolute timestamp of hardware acquisition. The timestamp is embedded as metadata in the image frame header, ensuring the physical uniqueness of the time reference. To ensure the uniformity of this time reference across heterogeneous hardware, the signal transmission unit 200 includes a dedicated hardware synchronization trigger line (Sync / Trigger). The high-precision servo tracking device 300 periodically (e.g., every second) sends nanosecond-level pulse signals to calibrate the system clock of the laser glare rejection unit 100, thereby eliminating clock drift between the two independent computing platforms and ensuring the timestamp... Strictly aligned with the controller's clock domain.

[0060] Image data is input into a neural network for inference, and the output is the pixel center coordinates of the target in the current local image coordinate system.

[0061] The pixel coordinates are converted to angular deviations. Based on the pinhole camera model and the pre-calibrated intrinsic parameter matrix, the pixel coordinates are converted into normalized angular deviations relative to the optical axis using geometric projection relationships. In addition, before outputting the final angular deviation, a pre-calibrated line-of-sight optical axis installation error compensation amount is added to correct the angular deviation calculated based on the camera coordinate system to the laser emission coordinate system, eliminating the aiming axis misalignment error caused by mechanical installation tolerances.

[0062] Simultaneously, the system reads data from the temperature sensor array deployed on key nodes of the optomechanical structure in real time. Based on a pre-established polynomial model of temperature axis deviation through high and low temperature environment experiments, it calculates the thermal deformation correction under the current temperature field and adds it to the angular deviation. Specifically, this polynomial model uses data related to temperature changes. The third-order fitting formula:

[0063] ;

[0064] in, The current temperature measured by the sensor. The reference temperature for calibration. to The system uses the least squares method to fit the axis thermal sensitivity coefficients. The system maintains separate sets of coefficients for the azimuth and pitch axes to dynamically compensate for minute thermal drift of the optical axis caused by laser self-heating or environmental temperature changes, ensuring line-of-sight consistency during long-term high-power operation. This conversion process calculates the azimuth angle deviation separately. Deviation from pitch angle If the current ROI clipping mode is in operation, the offset of the top left corner of the ROI relative to the full-resolution origin must be introduced before calculation to restore the coordinates.

[0065] This will include the target's presence markers and its azimuth angle deviation. Pitch angle deviation and absolute timestamp of hardware acquisition time The data packets are sent to the external high-precision servo tracking device 300 through the signal transmission unit 200.

[0066] The high-precision servo tracking device 300 includes a high-precision servo controller 310 and a high-precision two-dimensional motion platform 320. An inertial measurement unit 330 is rigidly mounted on the load surface at the end of the platform, parallel to the laser emission optical axis, and is used to measure the triaxial angular velocity of the load optical axis in inertial space in real time. As a key feature of this embodiment, inertial data is not only used for the stabilization of the underlying velocity loop, but also used as a feedforward quantity to participate in the state estimation of the upper layer. Before this, the original IMU data needs to be low-pass filtered to remove high-frequency vibration noise of the body and deduct the real-time estimated or pre-calibrated gyroscope zero bias to prevent the optical axis from deviating rapidly from the target due to the accumulation of angular velocity integral error during the pure inertial state inference process.

[0067] This embodiment constructs a dual-rate asynchronous closed-loop control architecture, which is logically divided into a fast control loop running in the high-precision servo controller 310 and a slow sensing loop running in the laser glare rejection unit 100.

[0068] The two loops achieve asynchronous data exchange by sharing state vectors and covariance matrices, and quickly control the loop to calculate the predicted covariance matrix. The diagonal element components are fed back to the slow sensing loop in real time; when the variance value is large (high uncertainty), the sensing system automatically configures a larger ROI to prevent target loss; when the variance value is small (tracking convergence), the system automatically shrinks the ROI to obtain a higher frame rate.

[0069] The high-frequency kinematic state deduction module runs in the high-precision servo controller 310.

[0070] Discrete-time state transition equations are constructed. For typical non-cooperative targets, their angular motion relative to the system is approximated as following a constant angular velocity (CV) model within an extremely short control period, and a triaxial angular velocity is introduced. The corresponding components in the equation are used as the control inputs of the system, and the following state transition equation is constructed:

[0071] ;

[0072] Wherein, the state vector Includes azimuth angle deviation relative to the optical axis Pitch angle deviation And the corresponding relative angular velocity components; Here is the state transition matrix. To control the input matrix, The process noise follows a Gaussian distribution and has a process noise covariance matrix. Furthermore, the noise covariance matrix of this process An adaptive adjustment strategy is adopted. Furthermore, an extended state observer or disturbance observer is introduced into the state extrapolation loop to treat dynamic factors not reflected in the CV model, such as wind load torque and shaft nonlinear friction torque, as lumped disturbances for real-time estimation. This disturbance estimate is then superimposed as a feedforward compensation term into the state transition equation, thereby endowing the controller with strong robustness against external gust interference and internal mechanical friction. The controller monitors the magnitude of the filter information vector in real time; when the deviation between the observed and predicted values ​​continuously exceeds the statistical significance boundary (indicating that the target is maneuvering), the magnitude is increased online. The values ​​of the main diagonal elements of the matrix are used to temporarily reduce the dependence on the constant velocity motion model, increase the weight of real-time visual measurements, and reduce maneuver delay.

[0073] Configure the system matrix, where the matrix It is used to embody the inertial feedforward mechanism, that is, when the carrier rotates in a certain direction, the relative position of the target in the field of view moves in the opposite direction, so that the system can offset the influence of the base disturbance on the direction of the line of sight in real time based solely on inertial data before visual feedback arrives.

[0074] The prediction update of execution state and covariance is derived using the posterior estimate from the previous time step, and the prediction covariance matrix is ​​calculated. And send it to the sensing unit.

[0075] The inertial endurance (flywheel) strategy module is used to maintain control continuity of the system during sensor mode switching blind spots.

[0076] Detect sensor mode switching status and maintain a flag bit. .

[0077] Performing a pure inertial state deduction, when in the switching blind zone ( The controller disconnects the measurement update loop, retaining only the prediction loop. At this time, the system continues to utilize the real-time acquired triaxial angular velocities. Substituting the values ​​into the state transition equation for recursion; simultaneously, a safety watchdog timer is activated internally in the controller. At this time, the system continuously utilizes the real-time acquired triaxial angular velocities. Substituting into the state transition equation for recursion; during this process, even if the target within the ROI is lost, the system still uses the background global motion vector output by the intelligent vision processing unit 120 to perform visual odometry calculation, calculate the camera's attitude change, and input the visual attitude change as the observation value into the Kalman filter to correct the random walk drift of the gyroscope online, thereby realizing vision-assisted inertial navigation (VIO) and significantly extending the system's high-precision maintenance time after the target is lost.

[0078] Simultaneously, a safety watchdog timer is activated internally by the controller. If no new valid image data is received within the preset maximum blind zone tolerance time (e.g., 500ms), or if the predicted covariance matrix... If the track exceeds the safety threshold, the tracking mode will be forcibly exited and the sensor will be reset to the maximum field of view search state to prevent filter divergence or gimbal malfunction due to lack of observation correction for a long time. At this time, the motion platform does not wait still, but performs active re-acquisition scanning based on Archimedes spiral or sector grating scanning strategy. The scanning center is set to the last predicted position before the target is lost, and the scanning range expands linearly with time to maximize the probability of re-acquiring the moving target in the uncertain area.

[0079] To reflect the accumulation of uncertainty caused by the lack of observation corrections within the blind zone, the covariance prediction equation is modified by performing blind zone expansion on the covariance matrix.

[0080] ;

[0081] Among them, the predicted covariance matrix An additional blind zone noise matrix is ​​superimposed during the blind zone period. The elements of this matrix were used to represent values ​​significantly larger than the process noise covariance matrix. This processing allows the Kalman gain to be significantly increased in the first filtering calculation after visual recovery, thereby achieving rapid convergence from the blind zone to the tracking state.

[0082] Bypass measurement updates, during the blind zone, directly set the posterior estimate to equal the prior estimate, maintaining the filter's iterative loop.

[0083] The hierarchical hysteresis perception module driven by prediction uncertainty runs on the deep learning vision processing unit 130.

[0084] Analyze the state variance and read the prediction covariance matrix of the feedback. Extract the variance value corresponding to the azimuth angle deviation. The variance value corresponding to the pitch angle deviation .

[0085] The field of view required for computational theory is based on a Gaussian distribution. The rule is to calculate the minimum physical field of view required to cover the possible distribution range of the target:

[0086] ;

[0087] ;

[0088] in, and These are the required field of view angles for azimuth and elevation, respectively; Confidence coefficient; This is a safety redundancy angle.

[0089] The hierarchical discretization mapping and hysteresis comparison logic are executed to compare the calculated required field of view with the threshold of the preset ROI level. The level switching is triggered only when the value exceeds the hysteresis range to avoid frequent ping-pong effects.

[0090] A configuration command is generated. When a switch is triggered, the upper left corner coordinates of the new ROI are calculated based on the target's current position, written to the sensor register, and the controller is notified to enter the blind zone inertial endurance mode. The state prediction function of the high-precision servo controller is used to calculate the predicted pixel center of the target on the image plane at the time of the next frame image acquisition, and this predicted center is used as the inverse solution reference for the geometric center of the new ROI. The upper left corner coordinates calculated in this way can compensate for the target's motion displacement during sensor configuration to the greatest extent, ensuring that the target is always located in the center of the reduced field of view.

[0091] The spatiotemporal alignment and asynchronous measurement fusion module runs on the high-precision servo controller 310 and is used to eliminate system errors caused by visual processing lag.

[0092] The controller maintains a historical state buffer queue. It allocates a circular buffer in memory to store historical state tuples according to the control cycle. Each tuple contains at least the historical time, historical state estimate, historical covariance matrix, and historical inertia data.

[0093] Calculate and retrieve the lag time in response to receiving the absolute timestamp of the hardware acquisition time. The system calculates the lag time from the observed data and retrieves the closest historical state record from the buffer. At the same time, it calculates the Mahalanobis distance between the observed data and the predicted historical state value. If the distance exceeds the preset chi-square distribution threshold, the current visual observation is determined to be an outlier or a false detection by the algorithm. The data of that frame is discarded directly without performing subsequent backtracking updates, thereby ensuring the robustness of the system under complex background interference.

[0094] The backtracking measurement update is performed. Based on the retrieved historical state, the measurement update step of Kalman filtering is performed using the residual between the observed data and the predicted historical state value to correct the historical state vector and covariance matrix, thereby compressing the uncertainty of historical moments. In this measurement update step, the measurement noise covariance matrix R is not a fixed constant, but is dynamically scaled in reverse according to the target class probability (confidence) output by the deep learning model.

[0095] Specifically, a mapping function is established that automatically increases the values ​​of the main diagonal elements of matrix R when the target detection confidence decreases (indicating potential local occlusion, motion blur, or background interference). This mapping function is set as follows:

[0096] ;

[0097] in, For the reference measurement noise covariance at high confidence levels, Normalized confidence scores for the output of deep learning models. This is the penalty gain coefficient (e.g., set to 100). The sensitivity index is set to 2 (e.g., 2). This formula causes the R value to increase slowly when the confidence level drops slightly, and to explode exponentially when the confidence level drops drastically. This almost completely cuts off the influence of the visual measurement in the frame in the filter, smooths the estimation jump caused by detection noise, and prevents the system from generating violent control oscillations in the low confidence state.

[0098] Perform the first double integral, using the corrected historical state as the starting point, and sequentially extract the absolute timestamps from the buffer at the time of hardware acquisition. All inertial data up to the current moment are recursively derived back to the current moment through the state transition equation.

[0099] The final output state, obtained by repeated integration, is the current optimal estimate incorporating the lag observation information, used for closed-loop control. Specifically, the controller generates torque commands for the servo motor based on the optimal estimated state to drive the optical axis to align with the target, and calculates the aiming error in real time. Furthermore, the laser emission enable signal contains a power modulation command; the controller generates a torque command for the servo motor based on the optimal estimated state; in this process, the original torque command needs to pass through a series of adaptive notch filters, which identifies the amplitude-frequency response of the mechanical system online, tracks and attenuates specific resonant frequency components of the mechanical structure in real time, and prevents image jitter caused by structural resonance induced by high dynamic tracking.

[0100] The online identification process specifically employs a frequency estimator based on the least mean square algorithm to monitor the energy concentration frequency points in the velocity loop error signal in real time; for each notch filter, the center frequency update formula is set as follows:

[0101] ;

[0102] in, Step size factor The current filtered residual is... The output is the gradient of the partial derivative with respect to the center frequency. Through this gradient descent logic, the center frequency of the filter can be automatically locked and follow the mechanical resonance peak of temperature or attitude changes, thus achieving resonance suppression under all operating conditions.

[0103] The filtered torque command is used to drive the optical axis to align with the target and calculates the aiming error in real time; the controller uses the pixel area of ​​the target bounding box... The approximate range of the target is inverted. Low-power / low-duty-cycle commands are output to close-range or large-sized targets to avoid permanent damage, while high-power / high-duty-cycle commands are output to long-range or small-sized targets to ensure sufficient dazzling suppression energy, thus achieving graded denial.

[0104] The duty cycle of the power modulation command Based on target pixel area Perform nonlinear mapping:

[0105] ;

[0106] in, As a reference area, The power gain coefficient, and These are the minimum and maximum duty cycle boundaries under system hardware safety constraints; this formula utilizes the optical principle that the square root of the target imaging area is approximately inversely proportional to the distance to automatically match the laser energy density that decreases with distance.

[0107] Only when the aiming error The controller sends a laser emission enable signal to the laser glare rejection unit 100 only when the glare threshold is less than the preset effective threshold (e.g., 0.5 milliradians) and the target presence marker is confirmed to be true. This achieves a high-precision, targeted rejection effect, preventing accidental damage to surrounding areas.

[0108] Simultaneously, the controller performs geofence security verification in parallel, reading the absolute encoder feedback angle of the high-precision two-dimensional motion platform 320 in real time and comparing it with the preset no-shoot angle range (such as densely populated areas on the ground or fixed airway areas in the air). Regardless of the visual tracking status, it compares it with the preset no-shoot angle range. This comparison process is not a static threshold comparison, but adopts a dynamic look-ahead strategy based on the current angular velocity. That is, it calculates the predicted angle in real time by adding the current angle and the angular velocity to the total system cutoff delay, thereby constructing a dynamic safety boundary. When the predicted angle touches the no-shoot zone, the cutoff action is triggered in advance to offset the physical delay of hardware signal transmission and relay action, ensuring that the laser is completely extinguished before the optical axis crosses the physical safety line. Regardless of the visual tracking status, once the optical axis points into the no-shoot zone, the controller immediately cuts off the laser emission circuit at the physical level, as the final safety defense line to prevent accidental damage to the system.

[0109] In addition, a hardware heartbeat monitoring mechanism is established between the high-precision servo controller (MCU domain) and the deep learning vision processing unit (NPU domain) in the signal transmission unit 200. The MCU, as the master control terminal, detects the pulse signal from the NPU at a fixed frequency (such as 100Hz). Once no valid heartbeat is detected within the set time window (indicating that the vision computing unit has experienced software deadlock, bus suspension, or overheating and crash), the MCU will directly trigger the underlying safety shutdown interrupt without relying on any upper-level instructions. Within microseconds, the laser enable pin will be forcibly pulled low to prevent uncontrolled irradiation due to the failure of the computing unit.

[0110] Specific application example: Defense system against low-altitude unauthorized drone flights in key areas;

[0111] In this embodiment, the above system is applied to a low-altitude security scenario in a large open-air stadium. The system is codenamed Sky Shield-X1.

[0112] Scene setup and hardware parameters:

[0113] Target of defense: Common consumer-grade quadcopter drones (approximately 350mm diagonal), with a flight speed of 0-20m / s and a distance of 200m to 800m.

[0114] Laser parameters: wavelength 532nm (green light), maximum power 4W, beam divergence angle 0.5mrad.

[0115] Sensor configuration: CMOS sensor with a resolution of 1920×1080, a pixel size of 2.9μm, and a lens focal length of 300mm.

[0116] Workflow demonstration and diagram analysis:

[0117] Phase 1: Wide-area search and capture:

[0118] The system is in panoramic search mode, with the image sensor running at 30fps and full resolution (1920×1080). When the deep learning vision processing unit 130 detects an intruding drone (confidence > 0.8), it immediately calculates its pixel coordinates and drives the high-precision two-dimensional motion platform 320 to perform coarse alignment.

[0119] Phase Two: High-Frequency Locking and Spatiotemporal Alignment (See Appendix) Figure 3 and appendix Figure 4 ):

[0120] Once the target enters the center of the field of view, the system automatically switches to ROI tracking mode, cropping the resolution to 256×256, and the frame rate instantly increases to 400fps. At this point, the spatiotemporal alignment and asynchronous measurement fusion module begins to work.

[0121] Tracking performance (corresponding appendix) Figure 3):

[0122] As attached Figure 3 (Comparison of system tracking performance trajectories) As shown in the figure, the horizontal axis represents the duration of the experiment (seconds), the vertical axis represents the azimuth angle of the target relative to the system (degrees), and the thick black solid line in the figure represents the actual trajectory of the UAV in the air (sine maneuver + linear drift).

[0123] At this stage, the high-precision servo controller 310 calculated the visual lag to be approximately 12ms. By reviewing the historical inertial state queue, it corrected the lag error caused by high-speed flight. (See attached diagram) Figure 3 The black dashed line representing the traditional control method clearly shows that the traditional method is always half a beat slower than the actual trajectory (there is obvious phase lag); while the black dotted line representing the present invention (inertial deduction) almost completely coincides with the actual trajectory (solid line), achieving zero lag tracking.

[0124] Tracking error analysis (corresponding appendix) Figure 4 ):

[0125] As attached Figure 4 As shown in the (Real-time Tracking Error Comparison) diagram, the vertical axis represents the absolute tracking error (degrees), and the horizontal dotted line in the diagram represents the effective threshold for glare (0.5 degrees), which is the standard for judging whether the aiming is successful. At this stage, the solid line representing the error of the present invention is always close to the 0 axis and below the threshold line, ensuring that the laser can accurately suppress the target; while the dashed line representing the error of the traditional method fluctuates more and is prone to exceeding the effective threshold.

[0126] Phase 3: Graded Denial and Power Modulation (see appendix) Figure 5 ):

[0127] The system is based on the area of ​​the detection frame. The target distance is inverted, and the laser power is adaptively adjusted based on the distance.

[0128] Power modulation logic (corresponding appendix) Figure 5 ):

[0129] As attached Figure 5 As shown in the diagram (Adaptive Laser Power Modulation), this figure uses a dual Y-axis coordinate system. The left vertical axis represents the target distance (meters), corresponding to the black dotted line in the figure; the right vertical axis represents the laser duty cycle (%), corresponding to the thick black solid line in the figure. The system calculates the command duty cycle according to the following formula. :

[0130] ;

[0131] Combined with appendix Figure 5 The curve changes can be seen as follows:

[0132] As the drone gets farther away (the dotted line shows a peak, about 800m), the system increases the laser duty cycle to ensure sufficient suppression energy (the solid line shows a peak, outputting strong light, about 95%).

[0133] When the drone approaches (the dotted line shows a trough, approximately 200m), to prevent the laser thermal effect from burning the drone's sensors and causing it to crash and injure people, the system automatically reduces the duty cycle (the solid line shows a trough, outputting weak light, approximately 15%), so that the image is overexposed and cannot be transmitted back. The figure clearly shows the inverse control relationship of increasing power as the distance increases and decreasing power as the distance decreases.

[0134] Phase Four: Anti-Shake Inertial Endurance (See Appendix) Figure 3 and appendix Figure 5 ):

[0135] The drone flew over the flagpole behind the stadium, where the visual signal was interrupted and it entered a blind spot.

[0136] Blind spot trajectory deduction (corresponding appendix) Figure 3 ):

[0137] In the appendix Figure 3 In the diagram, the gray shaded area (from the 4th to the 5th second) represents the visual blind spot (visual occlusion area). At this time, the system detects occlusion and the flag is set. .

[0138] Observe the curves in this area: The traditional method (dashed line) stops updating after entering the shadow area, maintaining the dead value of the previous moment, which causes the target to be completely lost after flying out of the obstruction. However, the present invention (dotted line) uses the data of the inertial measurement unit 330 to perform flywheel inference and continues to predict the flight path along the tangential direction in the shadow area, maintaining the continuity of the motion trend.

[0139] Safety disconnection mechanism (corresponding appendix) Figure 5 ):

[0140] In the appendix Figure 5 Similarly, in the gray shaded area (marked as a safe cut-off state) for 4-5 seconds, the system forcibly reduces the laser duty cycle (solid line) to 0% because the target's status cannot be confirmed. This reflects the system's safety mechanism: although the gimbal is tracking within the blind zone, the laser must be turned off to prevent accidental damage.

[0141] Blind spot error control (corresponding appendix) Figure 4 ):

[0142] Corresponding Appendix Figure 4During the blind zone period, the error (dashed line) of the traditional method increases linearly, indicating complete loss of tracking; while the error (solid line) of this invention, although slightly diverging within the blind zone due to lack of observation (manifesting as a small drift), remains within a controllable range. After 0.6 seconds, the drone flies out of the obstruction. Due to the extremely small inference error (<2mrad), the system instantly passes through the appearance feature vector. Complete re-identification and seamlessly resume closed-loop tracking.

[0143] Phase Five: Geographic Fence Circuit Breaker

[0144] When the drone attempts to swoop down toward the audience, the encoder feedback angle of the high-precision two-dimensional motion platform 320 triggers a dynamic look-ahead boundary; when the controller predicts that the optical axis will touch the boundary of the audience area in 30ms, it triggers a hardware interrupt in advance, forcibly cutting off the laser drive circuit to ensure absolute safety.

[0145] Experimental verification and effect comparison summary:

[0146] To verify the effectiveness of the present invention, a hardware-in-the-loop simulation test platform was constructed for comparative experiments.

[0147] Experimental conditions: The simulated target performs a sinusoidal maneuver (simulating a serpentine maneuver) at a frequency of 0.5 Hz and a maximum angular velocity of 5° / s. The system is compared to a traditional PID photoelectric tracking system.

[0148] Data analysis of experimental results:

[0149] Tracking accuracy: During the unobstructed and stable tracking phase (see attached document) Figure 4 In the non-shaded area, the root mean square error (RMSE) of the control group was 1.2 mrad; while the present invention uses high-frequency kinematic state deduction to reduce the RMSE to 0.15 mrad, and the error curve is always suppressed below the effective threshold line of glare (0.5 degrees).

[0150] Blind zone robustness: During the experiment, a complete blockage interference was artificially introduced between the 4th and 5th second (see attached reference). Figure 3 (Gray shaded area), the control group's optical axis stopped after losing the visual signal; while the present invention, through covariance dilation processing and inertial endurance, achieves the goal of losing the image but not losing the step. When the target reappears, the aiming error is only 0.8mrad, which is still within the effective coverage range of the laser beam.

[0151] Energy Control: Experimental Data (See Appendix) Figure 5 The results show that the laser duty cycle (solid line) of the present invention exhibits a strict nonlinear inverse proportional adjustment with the target distance (dotted line), and automatically triggers safety cut-off (duty cycle returns to zero) during the blind zone, verifying the safety and intelligence of the system under complex working conditions.

Claims

1. A high-precision laser glare closed-loop control system based on deep learning vision algorithms, characterized in that, include: The laser glare rejection unit (100) and the high-precision servo tracking device (300) are connected via a signal transmission unit (200); the laser glare rejection unit (100) integrates a high-definition video image acquisition unit (110), a deep learning vision processing unit (130) and a laser emission unit; the high-precision servo tracking device (300) includes a high-precision servo controller (310), a high-precision two-dimensional motion platform (320) and an inertial measurement unit (330). The system adopts a dual-rate asynchronous closed-loop architecture: the deep learning vision processing unit (130) outputs target observation data with absolute acquisition timestamps in the slow perception loop; the high-precision servo controller (310) performs high-frequency kinematic state deduction based on the inertial data output by the inertial measurement unit (330) in the fast control loop, and performs spatiotemporally aligned asynchronous measurement update according to the lag time of the target observation data to obtain the final state estimate to drive the high-precision two-dimensional motion platform (320).

2. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, A state covariance matrix interaction mechanism is established between the deep learning vision processing unit (130) and the high-precision servo controller (310); The high-precision servo controller (310) is used to generate a prediction covariance matrix when performing the high-frequency kinematic state deduction, and to feed back the prediction covariance matrix to the deep learning vision processing unit (130) in real time. The deep learning vision processing unit (130) is used to read the prediction covariance matrix, extract the diagonal element values ​​corresponding to the azimuth and pitch directions, and calculate the region of interest size required for the next frame image of the high-definition video image acquisition unit (110) based on the diagonal element values. The deep learning vision processing unit (130) is also used to write the size of the region of interest into the register interface of the high-definition video image acquisition unit (110) to adjust the resolution and frame rate of the image acquisition.

3. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The high-precision servo controller (310) is equipped with a historical state buffer queue, which stores historical time, historical state estimate, historical covariance matrix and historical inertial data according to the control cycle. The specific method by which the high-precision servo controller (310) performs the spatiotemporally aligned asynchronous measurement update includes: In response to receiving the target observation data, the historical state record with the smallest absolute time difference from the absolute acquisition timestamp is retrieved from the historical state buffer queue; The Kalman gain is calculated using the target observation data, the historical covariance matrix in the historical state record, and the historical state estimate, and the historical state record is then corrected. Using the corrected historical state record as a starting point, all historical inertial data from the absolute acquisition timestamp to the current moment are extracted, and forward reintegration is performed through the state transition equation to recursively obtain the final state estimate at the current moment.

4. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 3, characterized in that, The high-precision servo controller (310) is used to monitor the validity of the target observation data to generate a sensor mode switching flag. When the sensor mode switching flag is in a valid state, the high-precision servo controller (310) enters the inertial flywheel mode; In the inertial flywheel mode, the high-precision servo controller (310) cuts off the measurement update loop based on visual data and maintains the inertial deduction of the state vector only based on the three-axis angular velocity data output by the inertial measurement unit (330). Meanwhile, the high-precision servo controller (310) superimposes blind zone process noise onto the prediction covariance matrix generated by the high-frequency kinematic state deduction, and performs blind zone expansion processing on the prediction covariance matrix.

5. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The deep learning vision processing unit (130) is used to run a deep learning object detection model to perform inference, the deep learning object detection model containing independent embedding branches; The embedding branch is used to map the deep semantic features of the image into a unitized feature vector; The deep learning vision processing unit (130) is used to perform identity consistency verification by calculating the cosine similarity of the unitized feature vector after the target is lost and recaptured. The deep learning vision processing unit (130) is also used to introduce a kinematic constraint filter, calculate the theoretical reachable area of ​​the detection box in the current frame based on the target state at the previous moment and the preset maximum maneuverability of the target, and remove detection boxes whose center coordinates exceed the theoretical reachable area.

6. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The laser emitting unit is driven by a pulse width modulation signal, and the signal transmission unit (200) includes a line that inverts the pulse width modulation signal and transmits it to the high-definition video image acquisition unit (110). The high-definition video image acquisition unit (110) is used to take the inverted pulse width modulation signal as an external trigger shutter signal and perform exposure integration only during the laser pulse off interval. The automatic exposure control loop of the high-definition video image acquisition unit (110) is set with a maximum allowable integration time, which is limited to less than the difference between the turn-off time calculated by the laser pulse period and duty cycle and the safety protection interval.

7. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The target observation data includes the target pixel area, and the high-precision servo controller (310) is used to generate a power modulation command based on the final state estimate, the power modulation command including the power modulation duty cycle; The power modulation duty cycle is non-linearly mapped based on the target pixel area; The high-precision servo controller (310) is used to increase the power modulation duty cycle when the target pixel area decreases, and decrease the power modulation duty cycle when the target pixel area increases; The high-precision servo controller (310) is also used to calculate the aiming error, and only when the aiming error is less than the preset effective threshold for glare, it sends a laser emitting enable signal to the laser glare rejection unit (100).

8. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The high-precision two-dimensional motion platform (320) includes an absolute encoder, and the high-precision servo controller (310) is used to perform geofence security verification. The high-precision servo controller (310) reads the angle fed back by the absolute encoder in real time and calculates the dynamic predicted angle including the total cut-off delay of the system in combination with the current angular velocity; The high-precision servo controller (310) compares the dynamic prediction angle with the preset no-emission zone angle range, and cuts off the laser emission circuit when the dynamic prediction angle touches the no-emission zone angle range.

9. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, A temperature sensor is arranged on the laser glare rejection unit (100), and the deep learning vision processing unit (130) is used to convert pixel coordinates into angular deviations. During the conversion process, the deep learning vision processing unit (130) reads the data output by the temperature sensor and calculates the thermal deformation correction amount based on the temperature axis deviation polynomial model. The deep learning vision processing unit (130) adds the thermal deformation correction amount to the angle deviation and introduces the line-of-sight optical axis installation error compensation amount, converting the result into azimuth angle deviation and pitch angle deviation relative to the laser emission coordinate system.

10. The high-precision laser glare closed-loop control system based on deep learning vision algorithm according to claim 1, characterized in that, The signal transmission unit (200) is equipped with a hardware heartbeat monitoring mechanism; The deep learning vision processing unit (130) is used to output periodic pulse signals; The high-precision servo controller (310) is used to periodically detect the pulse signal from the deep learning vision processing unit (130); When the pulse signal is not detected within the preset time window, the high-precision servo controller (310) triggers a low-level safety shutdown interrupt, forcibly pulling down the laser enable pin of the laser emitting unit.

Images