An unmanned aerial vehicle autonomous safe landing control method based on end-side visual calculation cooperation

Abstract

The application discloses an unmanned aerial vehicle autonomous safe landing control method based on edge-side view calculation cooperation, belongs to the technical field of unmanned aerial vehicle control, and comprises the following steps: acquiring channel state information and descending speed of the unmanned aerial vehicle in real time, calculating a target compression sampling rate, and extracting an adaptive resolution ROI feature matrix according to the target compression sampling rate; calculating an edge semantic segmentation result matrix according to the adaptive resolution ROI feature matrix, and synchronously acquiring real-time wind resistance data of a station; calculating a comprehensive time delay window, evolving the edge semantic segmentation result matrix and the real-time wind resistance data to the future, generating a predicted space-time safety field matrix and a predicted wind shear compensation vector at a future moment; calculating a real-time confidence decay factor, and weighting and correcting the predicted space-time safety field matrix and the predicted wind shear compensation vector, and outputting a regulation and control instruction; and the application generates a future safety field and a wind shear compensation, thereby effectively offsetting physical time delay and ensuring the core effect of autonomous safe landing.

Description

Technical Field

[0001] This application belongs to the field of unmanned aerial vehicle (UAV) control technology, and specifically relates to an autonomous and safe landing control method for UAVs based on edge-view computing collaboration. Background Technology

[0002] With the booming development of the low-altitude economy, the application of logistics drones in e-commerce delivery, emergency material transportation, and other scenarios is becoming increasingly widespread. Autonomous and safe landing of drones is a core technological link in ensuring the smooth completion of the entire logistics loop. Existing drone landing control solutions typically rely on two mainstream architectures: one is independent onboard visual computing control, where the drone relies entirely on an onboard high-resolution camera, inertial navigation system, and onboard microprocessor to complete image recognition, semantic segmentation, and trajectory planning at the landing site locally; the other is cloud / edge-side assisted control, where the drone uploads the raw images or video streams of the landing area to a ground server or edge base station via a wireless communication link. Powerful edge computing then calculates the landing environment and sends control commands back to the drone.

[0003] However, in the actual complex logistics operation environment, the above-mentioned existing technologies have the following significant bottlenecks and defects: (1) Logistics drones have extremely strict limitations on airborne power consumption and load, and the airborne microprocessor cannot run complex deep learning semantic segmentation networks or multidimensional spatiotemporal prediction algorithms for a long time and at high frequency; (2) Traditional side-assisted schemes require uploading high-resolution original image data. In the actual descent process, due to multipath effect and Doppler frequency shift, the wireless channel state often fluctuates violently. If the transmission bandwidth drops suddenly, it is very easy to cause image data transmission lag and packet loss, resulting in the ground station being unable to provide control feedback in real time; (3) The wireless transmission delay from end to edge, combined with the forward propagation inference delay of the edge algorithm, constitutes an unavoidable comprehensive delay window. Since the UAV is in a dynamic high-speed motion state during descent, when the control command calculated by the edge is transmitted back to the airborne end, the landing field environment on which it is based has already lagged, resulting in spatiotemporal data disconnection, which cannot meet the safety requirements of high-frequency closed-loop control; (4) The low-altitude environment of the landing site is complex, and sudden local micro-meteorological wind fields can easily cause the UAV to deviate from the predetermined route in the final landing stage. The existing scheme lacks an effective mechanism to deeply integrate spatiotemporal obstacle prediction with real-time aerodynamic compensation. Summary of the Invention

[0004] To address the aforementioned issues, this application provides a method for autonomous and safe landing control of unmanned aerial vehicles (UAVs) based on edge-view computing collaboration, comprising the following steps: Step 100: Real-time acquisition of UAV channel state information and descent speed, calculation of target compression sampling rate, extraction of adaptive resolution ROI feature matrix of landing area based on target compression sampling rate, and sending to ground intelligent landing station; Step 200: The ground-based intelligent landing station calculates the edge semantic segmentation result matrix based on the adaptive resolution ROI feature matrix, and simultaneously acquires the station's real-time wind resistance data. Step 300: The ground intelligent landing station calculates the comprehensive delay window based on the transmission delay and algorithm inference delay, and uses the spatiotemporal prediction network to evolve the edge semantic segmentation result matrix and real-time wind resistance data into the future, generating the predicted spatiotemporal safety field matrix and predicted wind shear compensation vector for future moments and transmitting them back to the UAV. Step 400: The UAV calculates the real-time confidence decay factor of the side data based on the time difference between the current receiving time and the future time. Step 500: The UAV uses the real-time confidence decay factor to perform weighted correction on the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector, and outputs control commands based on the weighted correction of the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector.

[0005] Furthermore, step 100 specifically includes: Step 110: Channel estimation is performed using the pilot signal of the airborne wireless communication chip, and the current channel state information is output; simultaneously, airborne multi-source sensor data is acquired and filtered and fused, and the current actual descent rate is calculated and output in real time. Step 120: The channel state information and the descent rate are input into a preset two-parameter nonlinear mapping model for joint solution, and the target compression sampling rate is output. Step 130: Identify and lock the landing identifier of the ground landing station, obtain the bounding box coordinates of the key interest region, crop the original pixel matrix of the key interest region based on the bounding box coordinates, and use a sparse measurement matrix that matches the target compression sampling rate to perform linear projection dimensionality reduction on the original pixel matrix, and output the adaptive resolution ROI feature matrix. Step 140: Obtain the current transmission timestamp, encapsulate the adaptive resolution ROI feature matrix and the transmission timestamp into an uplink data packet, and transmit it to the ground intelligent landing station through the wireless communication link.

[0006] Furthermore, step 120 specifically includes: Step 121: Establish the input based on Shannon's theorem, the channel marginal effect, and the dynamic decay exponent theory. With target compression sampling rate The nonlinear mapping model between them is shown in the following formula: ; in, This is the channel gain scaling factor. The channel fading response coefficient is... As a speed risk decay index, It is the minimum sampling rate constant. This is a penalty coefficient for nonlinear computing power overhead. For channel state information, For the descent speed, It is a natural constant; Step 122: Based on the Newton-Raphson recursion principle, the nonlinear mapping model is iteratively solved to obtain the target compression sampling rate.

[0007] Furthermore, in step 200, the ground-based intelligent landing station calculates the edge semantic segmentation result matrix based on the adaptive resolution ROI feature matrix, specifically including the following steps: Step 210: The ground-based intelligent landing station converts the adaptive resolution ROI feature matrix into an input tensor that conforms to the preset neural network input boundary; Step 220: The ground intelligent landing station sends the input tensor into the neural network processor of the edge integrated machine, calls the deconvolution kernel parameters to perform cascaded sliding window back projection restoration, and outputs a multi-channel unnormalized score tensor. Step 230: The edge all-in-one machine performs activation function mapping on the multi-channel unnormalized score tensor and outputs the multi-channel normalized probability distribution tensor. Step 240: The edge integrated machine extracts the channel index corresponding to the maximum probability of each pixel position based on the multi-channel normalized probability distribution tensor to generate an initial label matrix; it extracts the single-channel probability distribution map belonging to the obstacle category from the multi-channel normalized probability distribution tensor, and compares the single-channel probability distribution map with the preset conservative safety confidence threshold. If the probability of a pixel belonging to an obstacle exceeds the conservative safety confidence threshold, the label of this pixel is mapped to the obstacle index value, and a binary edge semantic segmentation result matrix is ​​obtained.

[0008] Furthermore, in step 300, the calculation method for the comprehensive delay window of the ground intelligent landing station based on transmission delay and algorithm inference delay is as follows: The central control and processing unit of the ground-based intelligent landing station records the received timestamp as follows: And read the sending timestamp that the drone packaged and recorded during uplink. ; Will and The wireless transmission delay is calculated by performing a difference operation. The formula is as follows: ;in For wireless transmission latency; Read and accumulate the algorithm inference latency consumed by the current semantic segmentation inference. Calculate and output the overall delay window. The calculation formula is as follows: ;in This is a comprehensive delay window.

[0009] Furthermore, the steps for using a spatiotemporal prediction network to evolve the edge semantic segmentation result matrix and real-time wind resistance data into the future to generate the predicted spatiotemporal safety field matrix and predicted wind shear compensation vector for future moments are as follows: Step 310, calculate the algorithm inference delay. With the received timestamp Perform summation and output the corresponding future target timestamp. ;in A target timestamp for a future moment; Step 320: The target timestamp of the future moment is sent as the control step size input to the edge all-in-one machine; the edge all-in-one machine simultaneously accesses the edge semantic segmentation result matrix, real-time wind resistance data, and the current descent trajectory data of the UAV parsed and calculated from the uplink data packet as the initial state tensor; Step 330: The edge integrated machine inputs the initial state tensor into the pre-deployed spatiotemporal prediction network, uses the target timestamp of the future moment as the target boundary of the forward time extrapolation evolution, and performs nonlinear spatiotemporal feature extrapolation along the spatial geometric path of the current descent trajectory data of the UAV to obtain the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector for the future moment.

[0010] Furthermore, step 330 specifically includes: Step 331: Using the current descent trajectory of the UAV as the central axis, expand outwards the preset safe pipe radius to construct and output a spatial geometric gated mask matrix; Step 332: Based on the time difference between the current sampling time and the target timestamp of the future time, calculate the time look-ahead span, and then calculate the number of discrete time iterations required for the autoregressive advancement of the spatiotemporal prediction network to generate the time extrapolation control vector; Step 333: The initial state tensor, spatial geometry gated mask matrix, and temporal extrapolation control vector are fed into the pre-deployed spatiotemporal prediction network as common inputs. Inside the recurrent unit of the spatiotemporal prediction network, the intermediate layer features are multiplied element-wise with the spatial geometry gated mask matrix to constrain the feature evolution within the safe pipeline radius. Spatiotemporal extrapolation is performed cyclically according to the number of iterations to output the spatiotemporal evolution tensor at the target time. Step 334: Extract the spatial distribution features of obstacles in the spatiotemporal evolution tensor and input them into the first parallel decoding head, convert them into virtual repulsion potential energy values ​​in discrete space, and output the predicted spatiotemporal safety field matrix. Step 335: Extract the micro-meteorological wind field features from the spatiotemporal evolution tensor and input them into the second parallel decoding head. Through global average pooling and fully connected layer calculation, output the predicted wind shear compensation vector for the side wind shear disturbance.

[0011] Furthermore, in step 400, the formula for calculating the real-time confidence decay factor of the side data is as follows: ; in, The absolute value of the time deviation is a scalar, and the calculation formula is: , This is the current timestamp. For future timestamps; The attenuation sensitivity coefficient, The preset upper limit threshold for failure time deviation. This is the real-time confidence decay factor for the side data.

[0012] Furthermore, in step 500, the calculation formula for the weighted correction of the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector by the UAV using the real-time confidence attenuation factor is as follows: ; ; in, and These are the weighted, corrected prediction spatiotemporal safety field matrix and the prediction wind shear compensation vector, respectively. and These are the pre-fixed system preset static full-risk safety field matrix and the system preset static safety balance vector, respectively. and These are the prediction spatiotemporal safety field matrix and the prediction wind shear compensation vector, respectively. This is the real-time confidence decay factor for the side data.

[0013] Furthermore, the output control command based on the weighted and corrected predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector includes the following steps: Divide the current physical coordinate scalar of the UAV by the preset grid physical resolution constant and round down to obtain the nearest neighbor grid index tuple; Using the nearest neighbor grid index tuple as the address index key, a single-point physical addressing read operation is performed on the weighted and corrected predicted spatiotemporal safety field matrix to output the high-frequency spatiotemporal danger potential energy scalar. High-frequency spatiotemporal danger potential scalar Critical threshold of pre-stored safety boundary potential energy Perform a one-dimensional size comparison; if the following conditions are met... If the condition is met, then write logic 1 to the mode register; otherwise, if the condition is not met, write logic 1 to the mode register. Then write logic 0 to the mode register; obtain the control mode switching Boolean flag. ; when When an adaptive descent rate adjustment command is issued, The go-around control command was issued at the appropriate time.

[0014] The beneficial effects of the technical solutions provided in this application include at least the following: (1) Input the real-time channel state information and the actual descent speed of the UAV into the two-parameter nonlinear mapping model to calculate the target compression sampling rate. Based on the target compression sampling rate, extract and project the feature matrix with an adaptive resolution that matches the current bandwidth, which greatly reduces the computation and transmission overhead at the end. (2) A comprehensive time delay window was constructed, and the semantic segmentation results and real-time wind resistance data were used to evolve along the current descent trajectory to the future using a spatiotemporal prediction network. It can directly output the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector for future moments, which perfectly offsets the physical time delay caused by wireless transmission and algorithm inference. (3) A spatial geometric gating mask matrix based on the current descent trajectory is introduced. During the network autoregressive propagation process, the hidden state feature map and the spatial geometric gating mask matrix perform a hardware Hadamard product operation by multiplying element by element. This forcibly constrains the development trend of future features within the preset physical flight safety pipeline spatial domain, effectively suppressing the noise expansion of irrelevant spatial domains and making the predicted safety field matrix output from the side more focused.

[0015] Other features and advantages of this application will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the application. The objectives and other advantages of this application may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 A flowchart of an autonomous and safe landing control method for unmanned aerial vehicles based on edge-view computing collaboration is shown. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please see Figure 1 An autonomous and safe landing control method for unmanned aerial vehicles (UAVs) based on edge-view computing collaboration includes the following steps: Step 100: Real-time acquisition of UAV channel state information and descent speed, calculation of target compression sampling rate, extraction of adaptive resolution ROI (Region of Interest) feature matrix of landing area based on target compression sampling rate, and transmission to ground intelligent landing station; Step 200: The ground-based intelligent landing station calculates the edge semantic segmentation result matrix based on the adaptive resolution ROI feature matrix, and simultaneously acquires the station's real-time wind resistance data. Step 300: The ground intelligent landing station calculates the comprehensive delay window based on the transmission delay and algorithm inference delay, and uses the spatiotemporal prediction network to evolve the edge semantic segmentation result matrix and real-time wind resistance data into the future, generating the predicted spatiotemporal safety field matrix and predicted wind shear compensation vector for future moments and transmitting them back to the UAV. Step 400: The UAV calculates the real-time confidence decay factor of the side data based on the time difference between the current receiving time and the future time. Step 500: The UAV uses the real-time confidence decay factor to perform weighted correction on the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector, and outputs control commands based on the weighted correction of the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector.

[0020] The following is a detailed explanation of each step: Step 100 specifically includes: Step 110: Channel estimation is performed using the physical layer (PHY) base station pilot signal of the airborne wireless communication chip. The amplitude and phase distortion of the pilot symbols are calculated in real time, and the current channel state information is output. Simultaneously, the vertical accelerometer data, barometer pressure change rate, and multi-source positioning data of the Global Navigation Satellite System (GNSS) are read from the airborne inertial measurement unit (IMU), input into the extended Kalman filter fusion algorithm, and the current descent velocity is calculated and output in real time. Channel State Information (CSI) refers to the channel attribute characteristics of a mobile communication link, used to quantitatively reflect signal attenuation, multipath effects, and Doppler shift along the transmission path.

[0021] Step 120: The channel state information and descent rate are used as common input sources and input into a preset two-parameter nonlinear mapping model stored in the memory of the airborne microprocessor for joint solution, and the target compression sampling rate is output. Step 130: Use the airborne camera to acquire panoramic images and identify and lock the landing identifiers of the ground landing station, output the bounding box coordinates of the Region of Interest (ROI), and crop the original pixel matrix of the ROI based on the bounding box coordinates. The original pixel matrix of the ROI is linearly projected to reduce its dimensionality by calling a sparse measurement matrix that matches the target compression sampling rate, and the adaptive resolution ROI feature matrix is ​​extracted and output. ; Step 140: Call the real-time clock to obtain the microsecond-level system timestamp, which is used as the current transmission timestamp T1; then use the adaptive resolution ROI feature matrix. Together with the transmission timestamp T1 as the data payload field, the data is assembled into a complete uplink data packet according to the underlying communication protocol; the uplink data packet is pushed into the transmission buffer of the airborne radio transceiver and transmitted to the ground smart landing station through the wireless communication link.

[0022] Actual descent speed in step 110 The solution process includes: Step 111: The airborne flight control system acquires vertical accelerometer data within each preset sampling period. Barometer pressure change rate GNSS multi-source positioning data DGNSS containing vertical displacement and Doppler velocity; aggregate the above data and output a unified multi-source bottom-level raw observation dataset; Step 112, using vertical accelerometer data Perform time integration on the velocity state of the previous cycle and output the estimated vertical velocity state quantity. ; Step 113: Using the system process noise covariance matrix, analyze the predicted vertical velocity state variables. Perform error propagation deduction and output the predicted state error covariance matrix. ; The formula for calculating the prediction state error covariance matrix is: ,in, , For height variance, For vertical velocity variance, Let vertical height z be the vertical velocity. Covariance between Here is the state transition matrix. for transpose, Let be the process noise covariance matrix.

[0023] Step 114: Multiply the preset observation matrix by the estimated vertical velocity state variable. Converted to predicted measurements and compared with the barometer pressure change rate. The difference between GNSS multi-source positioning data and DGNSS data is calculated to output a measurement innovation residual vector that reflects the actual measurement deviation. ; Step 115, combining the predicted state error covariance matrix The optimal Kalman gain matrix is ​​calculated and output by performing matrix inversion operations on the preset sensor measurement noise covariance matrix. ; Step 116, using the Kalman gain matrix Measurement innovation residual vector Perform weighted calculations to obtain the state correction amount; Step 117: Add the state correction to the estimated vertical velocity state value. The system performs a posterior state update to eliminate accumulated errors caused by inertial navigation drift, outputs the optimal estimated three-dimensional vertical velocity, and uses this optimal estimate as the current actual descent velocity. .

[0024] Step 120 specifically includes: Step 121: Establish the input based on Shannon's theorem, the channel marginal effect, and the dynamic decay exponent theory. With target compression sampling rate The nonlinear mapping model between them is shown in the following formula: ; in, This is the channel gain scaling factor. The channel fading response coefficient is... As a speed risk decay index, It is the minimum sampling rate constant. This is a penalty coefficient for nonlinear computing power overhead. For channel state information, For the descent speed, It is a natural constant; Step 122: Based on the Newton-Raphson recursion principle, the nonlinear mapping model is iteratively solved to obtain the target compression sampling rate.

[0025] Step 130 specifically includes: Step 131, based on the target compression sampling rate Addressing access to the preset measurement matrix reference dictionary, Perform range addressing matching with the sampling rate index key values ​​in the dictionary, and call and latch the sparse measurement matrix that matches the current sampling rate. ; Step 132: Synchronously retrieve and read the original pixel matrix of the ROI corresponding to the current frame from the high-speed image buffer. ; Step 133, sparse measurement matrix Column dimensions and the original pixel matrix of the ROI The row dimension is used to perform matrix multiplication dimension validity checks and clock synchronization, and the data pairs to be projected and aligned are combined and output. Step 134: Perform matrix multiplication on the data pairs to be projected and aligned to obtain the adaptive resolution ROI feature matrix. The calculation formula is: .

[0026] In step 200, the ground-based intelligent landing station calculates the edge semantic segmentation result matrix based on the adaptive resolution ROI feature matrix, specifically including the following steps: Step 210: The edge integrated machine of the ground intelligent landing station converts the adaptive resolution ROI feature matrix into a low-dimensional input tensor that conforms to the preset neural network input boundary. Step 220: The ground-based intelligent landing site feeds the low-dimensional input tensor into the neural network processor of the edge integrated machine, calls the deconvolution kernel parameters embedded in the on-chip weight cache, and performs hardware primitive-level matrix multiplication-addition-dot product pipeline operations: First, the spatial geometric pixel gaps of the input tensor are padded with zeros to achieve sparsity expansion. Then, the deconvolution kernel parameters are used to perform cascaded sliding window back projection restoration, which multiplies the spatial resolution of the low-dimensional feature map to a size corresponding to the original spatial dimension of the UAV landing site. Finally, the multi-channel unnormalized score tensor is calculated and output. ; Step 230, Edge All-in-One Machine along Multi-Channel Unnormalized Score Tensor The channel axis dimension performs parallel Softmax activation functions to numerically map each spatial pixel, transforming each spatial pixel on each channel into a discrete scalar that conforms to the probability distribution, normalizing and assembling it, and outputting a multi-channel normalized probability distribution tensor. Step 240: The edge computing unit extracts the channel physical index corresponding to the maximum probability of each pixel position along the channel axis based on the multi-channel normalized probability distribution tensor, and constructs a two-dimensional initial label matrix; it then extracts the single-channel probability distribution map belonging to the security-sensitive category (i.e., obstacle category) from the multi-channel normalized probability distribution tensor, and compares the single-channel probability distribution map with a preset conservative security confidence threshold. The comparison is performed to generate a binary mask matrix. If the probability of a pixel belonging to an obstacle exceeds the conservative safety confidence threshold (i.e., the corresponding value of the pixel in the binary mask matrix is ​​true), then the label of this pixel is mapped to the obstacle index value. After mapping, a binary edge semantic segmentation result matrix is ​​obtained.

[0027] Step 300 specifically includes the following steps: Step 310, the central control and processing unit of the ground intelligent landing station records the received timestamp as follows: And read the sending timestamp that the drone packaged and recorded during uplink. ;Will and The wireless transmission delay is calculated by performing a difference operation. The formula is as follows: ;in For wireless transmission latency; read and accumulate the algorithm inference latency consumed by the current semantic segmentation inference. Calculate and output the overall delay window. The calculation formula is as follows: ;in As a comprehensive latency window; to include algorithm inference latency With the received timestamp Perform summation and output the corresponding future target timestamp. ;in A target timestamp for a future moment; Step 320: Set the target timestamp for the future moment. The control step size is input to the edge integrated machine; the edge integrated machine synchronously accesses the edge semantic segmentation result matrix. Real-time wind resistance data And the current descent trajectory data of the drone obtained by parsing and calculating from the uplink data packets. as an initial state tensor ; Step 330: The edge appliance inputs the initialization state tensor into the pre-deployed spatiotemporal prediction network, in order to... As the target boundary for forward time extrapolation, along the descent trajectory data The spatial geometric path is used to extrapolate nonlinear spatiotemporal characteristics to obtain the predicted spatiotemporal safety field matrix for future times. and predicted wind shear compensation vector Step 330 specifically includes the following steps: Step 331: The edge computing device extracts the drone's descent trajectory data. The corresponding three-dimensional discrete path point coordinates, in Using the three-dimensional geometric pathline as the central axis, a preset physical flight safety conduit radius is extended outward. Mesh cells falling within the conduit radius are marked as logic 1, and mesh cells outside the conduit are marked as logic 0, thereby calculating, constructing, and outputting the spatial geometric gating mask matrix. ; Step 332: The edge appliance's clock manager captures the current system sampling time. And load the target timestamp for the specified future time. ;Will and Perform subtraction to calculate the time-lookahead scalar. ;Will Divide by the unit rolling time step of the pre-deployed spatiotemporal prediction network The discrete-time iterations N required for the network to perform autoregressive propagation are calculated by rounding down, and N is encapsulated as the core control constraint. The time extrapolation control vector is then calculated and output. ; Step 333: Initialize the state tensor Spatial geometric gated mask matrix Time extrapolation control vector As a common cascaded input, it is fed into the pre-deployed spatiotemporal prediction network in the neural network processor of the edge device; the spatiotemporal prediction network initiates an autoregressive multi-step forward rolling prediction procedure: to initialize the state tensor As The input base map at each time step, within each level of the Long Short-Term Memory (LSTM) recurrent unit (such as ConvLSTM) of the network, combines the hidden state feature maps of the intermediate layers with... Perform element-wise multiplication using hardware Hadamard product operations to force the development trend of future features to be constrained to the descent trajectory data. Within the locked descent channel spatial domain, noise expansion in irrelevant spatial domains is suppressed; the spatiotemporal prediction network, based on... The above spatiotemporal extrapolation is performed repeatedly for the specified number of iterations N, until the target boundary is reached after N autoregressive iterations are exhausted. Stop at time, extrapolate across time, and output the high-dimensional spatiotemporal evolution tensor of the target time. ; Step 334: Extract the high-dimensional spatiotemporal evolution tensor at the target time. The probability spatial distribution of future obstacles in the image is input to the first parallel decoding head. The three-dimensional transpose convolution operator (Conv3D-Transpose) is called to perform boundary smoothing and safety margin topology expansion on the probability distribution, converting it into a virtual repulsive potential scalar value in a discrete spatial grid. After array encapsulation, the corresponding predicted spatiotemporal safety field matrix is ​​output on-site. ; Step 335: Extract the high-dimensional spatiotemporal evolution tensor at the target time. The future local micro-meteorological wind field channel is input to the second parallel decoding head. After the spatial dimension is compressed by the global average pooling layer, it is sent to a one-dimensional multi-layer linear fully connected layer (MLP). The rigid body aerodynamic model is implicitly solved at the feature level. The reverse control three-dimensional correction amount required to completely offset the wind shear disturbance torque at the future spatial location is quantitatively decoded and calculated. The predicted wind shear compensation vector is then encapsulated and output. .

[0028] Step 340, predict the spatiotemporal security field matrix With predicted wind shear compensation vector and future target timestamps Encapsulation yields downlink control calibration data packets .

[0029] In step 400, the formula for calculating the real-time confidence decay factor of the side data is as follows: ; in, The absolute value of the time deviation is a scalar, and the calculation formula is: , The timestamp of the current moment; The attenuation sensitivity coefficient, The preset upper limit threshold for failure time deviation. This is the real-time confidence decay factor for the side data.

[0030] Step 500 specifically includes the following steps: Step 510: Read the pre-fixed system preset static full-risk safety field matrix. and the system's preset static safety balance vector The system has a preset static full-risk safety field matrix. This refers to a discrete numerical matrix in which, under the three-dimensional physical grid of the landing site, the repulsive potential energy values ​​of all spatial grids except the preset absolute central geometric safety passage are assigned the maximum risk physical extreme value; the system preset static safety trim vector It refers to the basic three-dimensional control vector of a drone in a standard windless environment or nominal hovering state; Step 520: Perform a weighted correction on the predicted spatiotemporal security field matrix. The correction formula is as follows: ; in, This is the weighted and corrected prediction spatiotemporal security field matrix. A static full-risk safety field matrix is ​​preset for the pre-fixed system. To predict the spatiotemporal security field matrix.

[0031] Step 530: Perform a weighted correction on the predicted wind shear compensation vector. The correction formula is as follows: ; in, This is the weighted and corrected predicted wind shear compensation vector. Pre-set static safety balance vectors for the pre-fixed system. This is for predicting wind shear compensation vectors.

[0032] Step 540: Scalarize the current physical coordinates of the drone. Divide by the preset grid physical resolution constant d and round down to obtain the nearest neighbor grid index tuple. ; Step 550, generate the nearest neighbor grid index tuple As an address index key input, for Perform a single-point physical addressing read operation to directly copy and read the discrete potential energy values ​​stored in the specific grid address, and output the high-frequency spatiotemporal danger potential energy scalar. The calculation formula is: ; Step 560, scalarize the high-frequency spatiotemporal danger potential energy. Critical threshold of pre-stored safety boundary potential energy Perform a one-dimensional size comparison; if the following conditions are met... If the condition is met, it is determined that there is an insurmountable sudden risk obstacle in the airspace ahead, and a logic 1 is written to the mode register; otherwise, if the condition is not met, it is determined that there is no such obstacle. If the current space environment is deemed safe and reachable, a logic 0 is written to the mode register; the control mode switching Boolean flag is then obtained. ; Step 570, when When an adaptive descent rate adjustment command is issued, The go-around control command was issued at the appropriate time.

[0033] Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration, characterized in that, Includes the following steps: Step 100: Real-time acquisition of UAV channel state information and descent speed, calculation of target compression sampling rate, extraction of adaptive resolution ROI feature matrix of landing area based on target compression sampling rate, and sending to ground intelligent landing station; Step 200: The ground-based intelligent landing station calculates the edge semantic segmentation result matrix based on the adaptive resolution ROI feature matrix, and simultaneously acquires the station's real-time wind resistance data. Step 300: The ground intelligent landing site calculates the comprehensive delay window based on the transmission delay and algorithm inference delay, and uses the spatiotemporal prediction network to evolve the edge semantic segmentation result matrix and real-time wind resistance data into the future, generating the predicted spatiotemporal safety field matrix and predicted wind shear compensation vector for future moments and transmitting them back to the UAV. Step 400: The UAV calculates the real-time confidence decay factor of the side data based on the time difference between the current receiving time and the future time. Step 500: The UAV uses the real-time confidence decay factor to perform weighted correction on the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector, and outputs control commands based on the weighted correction of the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector.

2. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 1, characterized in that, Step 100 specifically includes: Step 110: Channel estimation is performed using the pilot signal of the airborne wireless communication chip, and the current channel state information is output; simultaneously, airborne multi-source sensor data is acquired and filtered and fused, and the current actual descent rate is calculated and output in real time. Step 120: The channel state information and the descent rate are input into a preset two-parameter nonlinear mapping model for joint solution, and the target compression sampling rate is output. Step 130: Identify and lock the landing identifier of the ground landing station, obtain the bounding box coordinates of the key interest region, crop the original pixel matrix of the key interest region based on the bounding box coordinates, and use a sparse measurement matrix that matches the target compression sampling rate to perform linear projection dimensionality reduction on the original pixel matrix, and output the adaptive resolution ROI feature matrix. Step 140: Obtain the current transmission timestamp, encapsulate the adaptive resolution ROI feature matrix and the transmission timestamp into an uplink data packet, and transmit it to the ground intelligent landing station through the wireless communication link.

3. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 2, characterized in that, Step 120 specifically includes: Step 121: Establish the input based on Shannon's theorem, the channel marginal effect, and the dynamic decay exponent theory. With target compression sampling rate The nonlinear mapping model between them is shown in the following formula: ； in, This is the channel gain scaling factor. The channel fading response coefficient is... As a speed risk decay index, It is the minimum sampling rate constant. This is a penalty coefficient for nonlinear computing power overhead. For channel state information, For the descent speed, It is a natural constant; Step 122: Based on the Newton-Raphson recursion principle, the nonlinear mapping model is iteratively solved to obtain the target compression sampling rate.

4. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 1, characterized in that, In step 200, the calculation of the edge semantic segmentation result matrix by the ground intelligent landing station based on the adaptive resolution ROI feature matrix specifically includes the following steps: Step 210: The ground-based intelligent landing station converts the adaptive resolution ROI feature matrix into an input tensor that conforms to the preset neural network input boundary; Step 220: The ground intelligent landing station sends the input tensor into the neural network processor of the edge integrated machine, calls the deconvolution kernel parameters to perform cascaded sliding window back projection restoration, and outputs a multi-channel unnormalized score tensor. Step 230: The edge all-in-one machine performs activation function mapping on the multi-channel unnormalized score tensor and outputs the multi-channel normalized probability distribution tensor. Step 240: The edge integrated machine extracts the channel index corresponding to the maximum probability of each pixel position based on the multi-channel normalized probability distribution tensor to generate an initial label matrix; it extracts the single-channel probability distribution map belonging to the obstacle category from the multi-channel normalized probability distribution tensor, and compares the single-channel probability distribution map with the preset conservative safety confidence threshold. If the probability of a pixel belonging to an obstacle exceeds the conservative safety confidence threshold, the label of this pixel is mapped to the obstacle index value, and a binary edge semantic segmentation result matrix is ​​obtained.

5. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 2, characterized in that, In step 300, the calculation method for the comprehensive delay window of the intelligent landing site based on transmission delay and algorithm inference delay is as follows: The central control and processing unit of the ground-based intelligent landing station records the received timestamp as follows: And read the sending timestamp that the drone packaged and recorded during uplink. ; Will and The wireless transmission delay is calculated by performing a difference operation. The formula is as follows: ;in For wireless transmission latency; Read and accumulate the algorithm inference latency consumed by the current semantic segmentation inference. Calculate and output the overall delay window. The calculation formula is as follows: ;in This is a comprehensive delay window.

6. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 5, characterized in that, The steps for using a spatiotemporal prediction network to evolve the edge semantic segmentation result matrix and real-time wind resistance data into the future, generating a predicted spatiotemporal safety field matrix and a predicted wind shear compensation vector for future moments, are as follows: Step 310, calculate the algorithm inference delay. With the received timestamp Perform summation and output the corresponding future target timestamp. ;in A target timestamp for a future moment; Step 320: Input the target timestamp of the future moment as the control step size into the edge all-in-one machine; The edge all-in-one machine synchronously accesses the edge semantic segmentation result matrix, real-time wind resistance data, and the current descent trajectory data of the drone parsed and calculated from the uplink data packet as the initial state tensor; Step 330: The edge integrated machine inputs the initial state tensor into the pre-deployed spatiotemporal prediction network, uses the target timestamp of the future moment as the target boundary of the forward time extrapolation evolution, and performs nonlinear spatiotemporal feature extrapolation along the spatial geometric path of the current descent trajectory data of the UAV to obtain the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector for the future moment.

7. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 6, characterized in that, Step 330 specifically includes: Step 331: Using the current descent trajectory of the UAV as the central axis, expand outwards the preset safe pipe radius to construct and output a spatial geometric gated mask matrix; Step 332: Based on the time difference between the current sampling time and the target timestamp of the future time, calculate the time look-ahead span, and then calculate the number of discrete time iterations required for the autoregressive advancement of the spatiotemporal prediction network to generate the time extrapolation control vector; Step 333: The initial state tensor, spatial geometry gated mask matrix, and temporal extrapolation control vector are fed into the pre-deployed spatiotemporal prediction network as common inputs. Inside the recurrent unit of the spatiotemporal prediction network, the intermediate layer features are multiplied element-wise with the spatial geometry gated mask matrix to constrain the feature evolution within the safe pipeline radius. Spatiotemporal extrapolation is performed cyclically according to the number of iterations to output the spatiotemporal evolution tensor at the target time. Step 334: Extract the spatial distribution features of obstacles in the spatiotemporal evolution tensor and input them into the first parallel decoding head, convert them into virtual repulsion potential energy values ​​in discrete space, and output the predicted spatiotemporal safety field matrix. Step 335: Extract the micro-meteorological wind field features from the spatiotemporal evolution tensor and input them into the second parallel decoding head. Through global average pooling and fully connected layer calculation, output the predicted wind shear compensation vector for the side wind shear disturbance.

8. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 1, characterized in that, In step 400, the formula for calculating the real-time confidence decay factor of the side data is as follows: ； in, The absolute value of the time deviation is a scalar quantity, and the calculation formula is: , This is the current timestamp. For future timestamps; The attenuation sensitivity coefficient, The preset upper limit threshold for failure time deviation. This is the real-time confidence decay factor for the side data.

9. The method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 1, characterized in that, In step 500, the calculation formula for the weighted correction of the predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector by the UAV using the real-time confidence attenuation factor is as follows: ； ； in, and These are the weighted, corrected prediction spatiotemporal safety field matrix and the prediction wind shear compensation vector, respectively. and These are the pre-fixed system preset static full-risk safety field matrix and the system preset static safety balance vector, respectively. and These are the prediction spatiotemporal safety field matrix and the prediction wind shear compensation vector, respectively. This is the real-time confidence decay factor for the side data.

10. A method for autonomous and safe landing control of unmanned aerial vehicles based on edge-view computing collaboration according to claim 9, characterized in that, The step of outputting control commands based on the weighted modified predicted spatiotemporal safety field matrix and the predicted wind shear compensation vector includes the following steps: Divide the current physical coordinate scalar of the UAV by the preset grid physical resolution constant and round down to obtain the nearest neighbor grid index tuple; Using the nearest neighbor grid index tuple as the address index key, a single-point physical addressing read operation is performed on the weighted and corrected predicted spatiotemporal safety field matrix to output the high-frequency spatiotemporal danger potential energy scalar. High-frequency spatiotemporal danger potential scalar Critical threshold of pre-stored safety boundary potential energy Perform a one-dimensional size comparison; if the following conditions are met... If the condition is met, then write logic 1 to the mode register; otherwise, if the condition is not met, write logic 1 to the mode register. Then write logic 0 to the mode register; obtain the Boolean flag for controlling mode switching. ; when When an adaptive descent rate adjustment command is issued, The go-around control command was issued at the appropriate time.

Images