Unmanned aerial vehicle intelligent decision-making method and system based on deep reinforcement learning

Abstract

The application provides a UAV intelligent decision-making method and system based on deep reinforcement learning, relates to the technical field of intelligent decision-making, and comprises the following steps: acquiring a landing area image and depth information through a visual sensor, processing feature representation and three-dimensional environment description by using a risk perception type policy gradient algorithm, adaptively constructing an information correlation relationship, realizing feature fusion, determining an optimal landing position, predicting a dynamic obstacle trajectory, generating a flight path point sequence, and optimizing a control strategy when an environment mutates, so that the UAV can land safely and efficiently, and the safety and adaptability of UAV landing are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to intelligent decision-making technology, and in particular to a UAV intelligent decision-making method and system based on deep reinforcement learning. BACKGROUND

[0002] With the rapid development of UAV technology, intelligent autonomous landing has become a key research direction in the field of UAVs. Traditional UAV landing usually relies on ground markers or GPS positioning systems, while in complex environments or GPS signal limited areas, intelligent perception and autonomous decision-making ability is particularly important. Currently, UAV landing technology mainly collects environmental information through visual sensors, combines machine learning and deep learning algorithms for environmental perception and decision planning, and realizes safe landing. Deep reinforcement learning, as a frontier technology in the field of artificial intelligence, continuously learns and optimizes decision-making strategies through the interaction between agents and the environment, and shows great potential in the scene of autonomous landing of UAVs.

[0003] The existing technology has limited perception and understanding of the landing environment, and it is difficult to effectively process multi-modal information. Most systems rely only on single visual information for landing decision-making, lack effective integration of depth information, and lead to inaccurate safety assessment of landing areas in complex environments, especially in cases of light changes or complex terrain, which can easily lead to misjudgment.

[0004] Traditional policy gradient algorithms lack the ability to perceive and evaluate risks when dealing with UAV landing tasks. When facing uncertain environments and dynamic obstacles, it is difficult to balance safety and efficiency, and it is unable to dynamically adjust decision-making strategies according to real-time environmental changes, which can easily lead to unpredictable dangers for UAVs during landing.

[0005] The existing technology performs poorly in dealing with decision-making problems in dynamic environments, especially in obstacle trajectory prediction and real-time planning. Most systems lack effective response mechanisms to environmental mutations and are unable to adaptively optimize control strategies based on historical experience and current scene characteristics, thereby reducing the success rate of safe landing of UAVs in complex dynamic environments. SUMMARY

[0006] The embodiments of the present application provide a UAV intelligent decision-making method and system based on deep reinforcement learning, which can solve the problems in the prior art.

[0007] In a first aspect, the embodiments of the present application provide a UAV intelligent decision-making method based on deep reinforcement learning, comprising:

[0008] Obtain image data and depth information of a landing area collected by a vision sensor carried by a UAV, process the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and construct a three-dimensional environment description of the landing area based on the depth information;

[0009] Process the information representation and the three-dimensional environment description by using a risk-aware policy gradient algorithm, organize different types of environment information into a unified structure, adaptively construct the association relationship between information, selectively transmit and update environment information in combination with the association relationship, realize bidirectional propagation and fusion of features, perform landing area segmentation and evaluation based on the final fused features, and obtain an optimal landing position;

[0010] Based on the risk-aware policy gradient algorithm, predict the dynamic obstacle motion trajectory in the flight process of the UAV, extend the risk-aware policy gradient algorithm based on the prediction result, update algorithm parameters based on historical experience and dynamically adjust control variables according to current scene features to generate a sequence of waypoints, automatically optimize the control strategy by using the risk-aware policy gradient algorithm when detecting environmental mutations, perform trajectory tracking control in combination with the prediction result of the environment, execute trajectory tracking control instructions, and control the UAV to fly along the planned trajectory to the optimal landing position.

[0011] Process the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and construct a three-dimensional environment description of the landing area based on the depth information, including:

[0012] Perform image enhancement processing on the image data of the landing area based on a dynamic target detection operator, dynamically adjust the contrast limit threshold by calculating the entropy value of the image gray distribution, and obtain enhanced image data;

[0013] Perform convolution operation on the enhanced image data by using convolution kernels of different scales, and perform feature fusion by using the dynamic target detection operator to obtain information representation of the landing area; obtain a dynamic region mask and a static region mask according to the dynamic target detection operator based on the information representation of the landing area;

[0014] Multiply the information representation of the landing area and the static region mask element by element to obtain a static feature map, and perform three-dimensional reconstruction on the static feature map in combination with the information representation of the landing area to obtain a three-dimensional environment description of the static region;

[0015] Multiply the information representation of the landing area and the dynamic region mask element by element to obtain a dynamic feature map, and perform three-dimensional reconstruction on the dynamic feature map in combination with the information representation of the landing area to obtain a three-dimensional environment description of the dynamic region;

[0016] An adaptive weight coefficient is obtained by processing information representation of the landing area based on the dynamic target detection operator; and the three-dimensional environment description of the static area and the three-dimensional environment description of the dynamic area are weighted and fused by using the adaptive weight coefficient to obtain a three-dimensional environment description of the landing area.

[0017] The landing area image data is subjected to image enhancement processing based on a dynamic target detection operator, a contrast limit threshold value is dynamically adjusted by calculating an entropy value of image gray scale distribution, and enhanced image data is obtained, including:

[0018] A gray scale probability distribution of the landing area image data is calculated, a deep neural network is constructed to calculate an image entropy value of the gray scale probability distribution, a dynamic target detection operator is constructed, the dynamic target detection operator includes a Gaussian filter response term and a Laplacian operator response term, and the deep neural network is used to perform feature fusion on the Gaussian filter response term and the Laplacian operator response term to obtain a detection operator response value.

[0019] The deep neural network is used to process the detection operator response value to obtain a dynamic adjustment amount, and the deep neural network is used to perform feature fusion on the dynamic adjustment amount and a basic threshold value to obtain a contrast limit threshold value.

[0020] Each pixel in the landing area image data is detected, the deep neural network is used to analyze the relationship between the detection operator response value and the contrast limit threshold value, the gray scale value of each pixel is subjected to feature enhancement to obtain a response region enhancement value, and the deep neural network is used to perform spatial smoothing constraint processing on the response region enhancement value to obtain enhanced landing area image data.

[0021] The information representation and the three-dimensional environment description are processed by using a risk-aware policy gradient algorithm, different types of environment information are organized into a unified structure, and an association relationship between information is adaptively constructed, including:

[0022] An information entropy of an image is calculated based on the information representation, a feature extraction threshold value is set according to the information entropy, and feature extraction of the information representation and the three-dimensional environment description is completed when the information entropy is greater than the feature extraction threshold value.

[0023] A risk-aware policy gradient algorithm is constructed, a matrix eigenvalue and a bias eigenvalue of the risk-aware policy gradient algorithm are calculated, and the result of the feature extraction is iteratively optimized by using the risk-aware policy gradient algorithm when the matrix eigenvalue is greater than the feature extraction threshold value and the bias eigenvalue is less than the feature extraction threshold value to obtain an environment feature vector.

[0024] The environment feature vector is structured to obtain uniform structure environment information, and an association relationship between information is constructed based on the uniform structure environment information.

[0025] The environment information is selectively transmitted and updated based on the association relationship, bidirectional propagation and fusion of features are realized, a landing area is segmented and evaluated based on final fused features, and an optimal landing position is obtained, including:

[0026] An information transmission network is constructed, the association relationship strength and information mask of each region in the information transmission network are calculated, a transmission threshold is set, a transmission strength is obtained when the association relationship strength is greater than the transmission threshold and the information mask value is less than the transmission threshold, a local safety feature of each region is calculated according to the transmission strength, a global safety metric is determined based on the local safety feature, and a region weight is generated using the global safety metric;

[0027] A bidirectional propagation channel is constructed based on the transmission strength, a forward feature is obtained by performing a depth residual operation on the region weight and the transmission strength between adjacent nodes along the forward direction of the bidirectional propagation channel, and a backward feature is obtained by performing a depth residual operation on the region weight and the transmission strength between adjacent nodes along the reverse direction of the bidirectional propagation channel;

[0028] The forward feature and the backward feature are input into the information transmission network to obtain fused features, a local feature descriptor and a terrain gradient feature are separated from the fused features using the information transmission network, and a local detail feature is generated according to the local feature descriptor and the terrain gradient feature;

[0029] The global safety metric and the local detail feature are input into the information transmission network to obtain a region safety evaluation value, a candidate landing position is selected according to the region safety evaluation value, and an optimal landing position is determined from the candidate landing position.

[0030] The risk perception type policy gradient algorithm is used to predict the dynamic obstacle motion trajectory during the flight of the unmanned aerial vehicle, the risk perception type policy gradient algorithm is extended based on the prediction result, the algorithm parameters are updated based on historical experience, and the control variable is dynamically adjusted according to the current scene characteristics to generate a flight path point sequence, including:

[0031] Motion data of dynamic obstacles during the flight of the unmanned aerial vehicle is collected, position coordinates, velocity vectors, acceleration vectors and motion directions of the dynamic obstacles are calculated, the position coordinates, the velocity vectors, the acceleration vectors and the motion directions are input into the risk perception type policy gradient algorithm, and a predicted trajectory of the dynamic obstacle is obtained;

[0032] construct an experience pool based on the predicted trajectory, match a historical motion trajectory of the dynamic obstacle with the predicted trajectory, calculate a matching error, determine an instant reward value according to the matching error, and store the historical motion trajectory, the predicted trajectory and the instant reward value into the experience pool;

[0033] a training sample is extracted from the experience pool when the amount of data in the experience pool is greater than the experience pool capacity threshold, parameters of the risk-aware policy gradient algorithm are updated based on the training sample, and an extended risk-aware policy gradient algorithm is obtained;

[0034] environmental features of a current scene are acquired, the environmental features are input into the extended risk-aware policy gradient algorithm, a dynamic control variable is generated based on the extended risk-aware policy gradient algorithm, and a sequence of waypoints is generated according to the dynamic control variable.

[0035] The risk-aware policy gradient algorithm is used to automatically optimize the control strategy when an environmental mutation is detected, trajectory tracking control is performed in combination with an environmental prediction result, a trajectory tracking control instruction is executed, and the unmanned aerial vehicle is controlled to fly along a planned trajectory to the optimal landing position, which comprises:

[0036] environmental state information of adjacent time points is calculated to obtain an environmental state difference value, an environmental mutation threshold is set, and a risk assessment value is calculated based on the environmental state difference value when the environmental state difference value is greater than the environmental mutation threshold.

[0037] The risk assessment value is mapped into an advantage function by using the risk-aware policy gradient algorithm, a risk adjustment factor is determined based on the advantage function, and new control strategy parameters are obtained by updating original control strategy parameters according to the product of the advantage function and the risk adjustment factor.

[0038] The environmental state information is analyzed based on the new control strategy parameters to obtain an environmental prediction result, the environmental prediction result is matched with an actual position and an actual speed of the unmanned aerial vehicle to obtain an expected trajectory, and a tracking error between the actual position, the actual speed and the expected trajectory is calculated.

[0039] The tracking error is compensated according to the new control strategy parameters to obtain a position control amount and a speed control amount, the position control amount and the speed control amount are combined to form a trajectory tracking control instruction, and the trajectory tracking control instruction is executed to control the unmanned aerial vehicle to fly along the expected trajectory to an optimal landing position.

[0040] In a second aspect of the embodiment of the present application, an unmanned aerial vehicle intelligent decision system based on deep reinforcement learning is provided, which comprises:

[0041] A first unit is configured to acquire image data and depth information of a landing area collected by a vision sensor carried by a UAV, process the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and construct a three-dimensional environment description of the landing area based on the depth information.

[0042] A second unit is configured to process the information representation and the three-dimensional environment description by using a risk-aware policy gradient algorithm, organize different types of environment information into a unified structure, adaptively construct a correlation between the information, selectively transmit and update the environment information in combination with the correlation, realize bidirectional propagation and fusion of features, perform landing area segmentation and evaluation based on the finally fused features, and obtain an optimal landing position.

[0043] A third unit is configured to predict a dynamic obstacle motion trajectory in a flight process of the UAV based on the risk-aware policy gradient algorithm, extend the risk-aware policy gradient algorithm based on a prediction result, update algorithm parameters based on historical experience and dynamically adjust a control variable to generate a flight path point sequence according to a current scene feature, automatically optimize a control strategy by using the risk-aware policy gradient algorithm when detecting an environment mutation, perform trajectory tracking control in combination with an environment prediction result, execute a trajectory tracking control instruction, and control the UAV to fly along a planned trajectory to the optimal landing position.

[0044] In a third aspect, an electronic device is provided, including:

[0045] a processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0048] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0049] The present application has the following advantages:

[0050] The present application realizes efficient and safe UAV landing control by using a UAV intelligent decision-making method based on deep reinforcement learning. The method can fully utilize image data and depth information collected by a vision sensor, and combine a risk-aware policy gradient algorithm to perform environment perception and decision-making, thereby effectively improving the autonomous decision-making ability of the UAV in a complex environment.

[0051] The risk perception type policy gradient algorithm adopted by the application can adaptively construct the correlation between information, realize bidirectional propagation and fusion of features, enable the system to accurately evaluate the safety of the landing area and determine the optimal landing position. The algorithm can also predict the dynamic obstacle motion trajectory, greatly improving the ability of the system to respond to complex environmental changes.

[0052] The application can update algorithm parameters based on historical experience during the flight control process of the unmanned aerial vehicle, and dynamically adjust control variables according to current scene characteristics, especially when environmental mutations are detected, the control strategy can be automatically optimized, realizing the robust flight and accurate landing of the unmanned aerial vehicle in uncertain environments, and significantly enhancing the adaptability and safety of the unmanned aerial vehicle system in various complex scenes. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 A flowchart of the unmanned aerial vehicle intelligent decision-making method based on deep reinforcement learning of the embodiment of the application is shown in

[0054] Figure 2 A flowchart of the landing area three-dimensional environment description based on image processing of the embodiment of the application is shown in

[0055] Figure 3 A flowchart of the trajectory point generation based on risk perception of the embodiment of the application is shown in DETAILED DESCRIPTION

[0056] To make the purpose, technical scheme and advantages of the embodiments of the application clearer, the technical scheme in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0057] The technical scheme of the application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in some embodiments.

[0058] Figure 1 A flowchart of the unmanned aerial vehicle intelligent decision-making method based on deep reinforcement learning of the embodiment of the application is shown in Figure 1 As shown, the method comprises:

[0059] Obtaining landing area image data and depth information collected by a vision sensor carried by an unmanned aerial vehicle, processing the landing area image data by using a basic feature extraction method to obtain information representation of the landing area, and constructing a three-dimensional environment description of the landing area based on the depth information;

[0060] The risk-aware policy gradient algorithm is used to process the information representation and the three-dimensional environment description, different types of environment information are organized into a unified structure, the correlation between information is adaptively constructed, environment information is selectively transmitted and updated combined with the correlation, bidirectional propagation and fusion of features are realized, landing area segmentation and evaluation are performed based on the final fused features, and an optimal landing position is obtained;

[0061] Based on the risk-aware policy gradient algorithm, the dynamic obstacle motion trajectory in the flight process of the unmanned aerial vehicle is predicted, the risk-aware policy gradient algorithm is extended based on the prediction result, the algorithm parameters are updated based on historical experience and the control variable is dynamically adjusted according to the current scene characteristics to generate a sequence of waypoints, the risk-aware policy gradient algorithm is used to automatically optimize the control strategy when detecting environmental mutations, trajectory tracking control is performed combined with the prediction result of the environment, trajectory tracking control instructions are executed, and the unmanned aerial vehicle is controlled to fly along the planned trajectory to the optimal landing position.

[0062] In an optional implementation, the landing area image data is processed by using a basic feature extraction method to obtain information representation of the landing area, and a three-dimensional environment description of the landing area is constructed based on the depth information, including:

[0063] The landing area image data is subjected to image enhancement processing based on a dynamic target detection operator, the contrast limit threshold is dynamically adjusted by calculating the entropy value of the image gray scale distribution, and enhanced image data is obtained;

[0064] The enhanced image data is subjected to convolution operation by using convolution kernels of different scales, and feature fusion is performed by the dynamic target detection operator to obtain information representation of the landing area; a dynamic region mask and a static region mask are obtained from the dynamic target detection operator based on the information representation of the landing area;

[0065] The information representation of the landing area is multiplied element by element with the static region mask to obtain a static feature map, and the static feature map is subjected to three-dimensional reconstruction combined with the information representation of the landing area to obtain a three-dimensional environment description of the static region;

[0066] The information representation of the landing area is multiplied element by element with the dynamic region mask to obtain a dynamic feature map, and the dynamic feature map is subjected to three-dimensional reconstruction combined with the information representation of the landing area to obtain a three-dimensional environment description of the dynamic region;

[0067] The information representation of the landing area is processed based on the dynamic target detection operator to obtain adaptive weight coefficients; the three-dimensional environment descriptions of the static region and the dynamic region are weighted and fused by using the adaptive weight coefficients to obtain a three-dimensional environment description of the landing area.

[0068] As shown in Figure 2 The method comprises:

[0069] The landing area image data and depth information collected during the landing of the unmanned aerial vehicle are acquired. The image data is collected by an on-board high-definition camera, with a resolution of 2048x1536 pixels and a color depth of 24-bit true color. The depth information is obtained by an on-board laser radar or depth camera, with a resolution matching that of the image data, a depth accuracy of ±2 cm, and a measurement range of 0.5-50 meters. The collection frequency is 15 frames per second, ensuring that the system can process the landing area information in real time. The image data and depth information are synchronized by timestamp, with an error controlled within 10 milliseconds, ensuring the consistency of spatial information.

[0070] The landing area image data is subjected to image enhancement processing based on a dynamic target detection operator. The original image is pre-processed to eliminate noise and uneven illumination, including bilateral filtering (spatial domain standard deviation 3.0, value domain standard deviation 0.1, window size 5x5 pixels) and adaptive histogram equalization (grid size 8x8, contrast limit parameter 3.0). After preprocessing, the color image is converted to a grayscale image using the weighted average method, with weight coefficients of 0.299 for the red channel, 0.587 for the green channel, and 0.114 for the blue channel. The grayscale level probability distribution of the grayscale image is calculated, and the number of pixels for each grayscale level (0-255) in the image is counted and divided by the total number of pixels to obtain the corresponding probability. For example, in a certain frame of image, the number of pixels with a grayscale value of 128 is 157,286, and the total number of pixels is 3,145,728, so the probability is 0.05. A deep neural network is constructed to calculate the entropy value of the grayscale level probability distribution image, which consists of an input layer (256 neurons), a hidden layer (128 neurons, ReLU activation), and an output layer (1 neuron, linear activation). For a given grayscale distribution, the network outputs an entropy value of 6.82. A dynamic target detection operator is constructed, including a Gaussian filter response term (kernel size 5x5, standard deviation 1.0) and a Laplacian operator response term (3x3 template, center value 4, four-neighborhood value -1, diagonal value 0). Feature fusion network is used to fuse the two response terms to obtain the detection operator response value, which contains two input branches (each with two convolution layers) and a fusion branch. The response value is processed by a regression network to obtain a dynamic adjustment value (value 0.75), which is fused with the base threshold value (0.03) to generate a contrast limit threshold value (0.08). The enhancement network is applied to each pixel, with the input being the original grayscale value, the detection operator response value, and the contrast limit threshold value, and the output being the enhanced grayscale value. For example, a pixel with a grayscale value of 128, a response value of 0.78, and a threshold value of 0.08 is enhanced to 175. Finally, a spatial smoothing constraint network (U-Net architecture) is applied to process the preliminary enhancement result to obtain the final enhancement image, which maintains edge sharpness while ensuring smooth transition in the region. The contrast of the enhanced image is improved by 45%, the edge sharpness is improved by 38%, and the target region and background separation is improved by 52%.

[0071] The enhanced image data is convolved using different size convolution kernels, and the feature fusion is performed by a dynamic target detection operator to obtain the information representation of the landing area. A multi-scale feature extraction network is designed, which includes four parallel branches. Each branch uses different size convolution kernels: the first branch uses a 3x3 convolution kernel, the second branch uses a 5x5 convolution kernel, the third branch uses a 7x7 convolution kernel, and the fourth branch uses a 9x9 convolution kernel. Each branch includes two convolution layers and a max-pooling layer. The first layer of the first branch uses 32 3x3 convolution kernels with a step size of 1 and a padding of 1; the second layer uses 64 3x3 convolution kernels with a step size of 1 and a padding of 1; the max-pooling uses a 2x2 window with a step size of 2. The structures of the other branches are similar, but the convolution kernel sizes are different. The feature maps processed by each branch have consistent sizes (controlled by appropriate padding), which facilitates subsequent fusion. The output feature maps of the four branches are concatenated in the channel dimension to obtain a joint feature map with 256 channels (64x4). A feature fusion module is designed, which includes two 1x1 convolution layers and a 3x3 convolution layer. The first 1x1 convolution layer reduces the channel number from 256 to 128, the second 1x1 convolution layer reduces the channel number from 128 to 64, and the 3x3 convolution layer keeps the channel number unchanged. Each convolution layer is followed by batch normalization and ReLU activation function. After processing by the fusion module, a 64-channel landing area information representation is obtained, with a size of 1 / 4 of the original image (512x384). The information representation captures multi-scale spatial features, including texture, edge, corner, and region features, effectively representing the visual characteristics of the landing area. For example, the information representation of a certain area shows that the texture complexity is 0.72, the edge saliency is 0.85, and the corner density is 0.43, indicating that this area contains obvious edge features but fewer corners.

[0072] The information representation of the landing area is used to obtain a dynamic region mask and a static region mask according to a dynamic target detection operator. A region segmentation network is designed, with the input being the information representation of the landing area (64 channels, 512x384) and the output being the dynamic region mask and the static region mask (each 1 channel, 512x384). The network adopts an encoder-decoder architecture. The encoder includes four down-sampling blocks, each block including two 3x3 convolution layers and one 2x2 max-pooling layer. The first block has 64 output channels, and the channel number of each subsequent block is doubled, in the order of 128, 256, and 512. The decoder includes four up-sampling blocks, each block including one 2x2 up-sampling layer (transposed convolution) and two 3x3 convolution layers. The up-sampling block output channel numbers are 256, 128, 64, and 32, respectively. The encoder and the decoder use a skip connection. The last layer of the network is divided into two branches, each using a 1x1 convolution layer to generate a dynamic region mask and a static region mask, with the activation function being Sigmoid. During training, a binary cross-entropy loss function is used, with the optimizer being Adam and the learning rate being 0.0002. The core of region segmentation is to identify dynamic elements (such as moving vehicles and pedestrians) and static elements (such as the ground and buildings) in the scene. For example, in a certain landing scene, the detected dynamic region accounts for 15% of the total area, mainly distributed on the right side of the image, and the static region accounts for 85%, covering most of the ground area. In the dynamic region mask, a pixel value close to 1 indicates that the position is a dynamic target, and a value close to 0 indicates that the position is a static environment. The static region mask is the opposite. The sum of the two masks is approximately equal to a full 1 matrix, ensuring complete coverage of the scene.

[0073] The information representation of the landing area is multiplied element by element with the static region mask to obtain a static feature map, and the static feature map is reconstructed in three dimensions in combination with the information representation of the landing area to obtain a three-dimensional environment description of the static region. The calculation of the static feature map is to multiply the 64-channel information representation and the static region mask (expanded to 64 channels) element by element, retaining the feature information of the static region and suppressing the feature information of the dynamic region. The static feature map maintains the channel number and spatial dimension (64 channels, 512x384) of the original information representation. A static region three-dimensional reconstruction network is designed, which inputs the static feature map and the original depth information (downsampled to 512x384) and outputs a three-dimensional environment description of the static region. The reconstruction network includes a feature extraction module, a depth fusion module, and a three-dimensional generation module. The feature extraction module includes three residual blocks, each of which includes two 3x3 convolution layers and a skip connection. The depth fusion module combines the feature map and the depth information, including two 3x3 convolution layers and an attention mechanism, to enhance the guiding effect of the depth information on the features. The three-dimensional generation module maps the two-dimensional features to the three-dimensional space to generate point cloud data (about 500,000 points) and geometric properties. The reconstruction network is trained using a three-dimensional reconstruction loss function, combining a point cloud distance loss and a geometric consistency loss. For example, the reconstructed static region includes a flat ground (about 200 square meters in area) and surrounding fixed buildings, with an average reconstruction accuracy of ±3 centimeters. The three-dimensional environment description of the static region includes spatial coordinates, normal vectors, curvature, and semantic labels, and is stored in a structured data format.

[0074] The information representation of the landing area is multiplied element by element with the dynamic region mask to obtain a dynamic feature map, and the dynamic feature map is reconstructed in three dimensions in combination with the information representation of the landing area to obtain a three-dimensional environment description of the dynamic region. The calculation method of the dynamic feature map is similar to that of the static feature map, which multiplies the information representation and the dynamic region mask to retain the feature information of the dynamic region. A dynamic region three-dimensional reconstruction network is designed, which has a similar structure to the static region reconstruction network but adds a temporal feature extraction module to enhance the three-dimensional reconstruction effect of dynamic targets using continuous multiple frames (5 frames) of information. The temporal feature extraction module includes a 3D convolution layer (3x3x3 convolution kernel) and a spatio-temporal attention mechanism to capture the motion characteristics of the target. The reconstruction network outputs a three-dimensional environment description of the dynamic region, including spatial position, motion velocity, motion direction, and predicted trajectory. For example, the reconstructed dynamic region includes a moving vehicle (speed 5 meters / second, direction northeast) and a pedestrian (speed 1.2 meters / second, direction southwest), with a reconstruction accuracy of ±5 centimeters and a velocity estimation error of ±0.3 meters / second. The three-dimensional environment description of the dynamic region focuses on the motion characteristics of the target, providing potential risk assessment for subsequent landing decisions.

[0075] The adaptive weight coefficient is obtained by processing the information representation of the landing area based on a dynamic target detection operator. A weight generation network is designed, which takes the information representation of the landing area and the detection operator response value as input, and outputs the adaptive weight coefficient. The network contains four convolutional layers and two fully connected layers. The first convolutional layer uses 32 3x3 convolutional kernels, the second layer uses 64 3x3 convolutional kernels, the third layer uses 128 3x3 convolutional kernels, and the fourth layer uses 256 3x3 convolutional kernels. The step size of each layer is 2, which realizes spatial down-sampling. The feature map is converted into a feature vector through global average pooling, and the global weight coefficient a is generated through two fully connected layers (256→64→1), with a range of [0, 1]. At the same time, the output of the fourth convolutional layer of the network is processed through 1x1 convolution and Sigmoid activation to generate a spatially varying weight map β, which has the same size as the information representation. The global weight a and the spatial weight β are combined to form the final adaptive weight coefficient γ, which contains both global information and local changes. For example, in a scene with many dynamic targets, the weight of the static area is 0.35, and the weight of the dynamic area is 0.65; in a relatively static scene, the weight of the static area is 0.8, and the weight of the dynamic area is 0.2. The adaptability of the weight coefficient is reflected in the response to the dynamic changes of the scene, ensuring the optimal environment description in different scenes.

[0076] The three-dimensional environment description of the landing area is obtained by weighting and fusing the three-dimensional environment description of the static area and the three-dimensional environment description of the dynamic area using the adaptive weight coefficient. The fusion process is divided into two stages: feature-level fusion and decision-level fusion. Feature-level fusion combines the feature representations of the static area and the dynamic area through adaptive weight coefficient to generate a unified feature representation. Decision-level fusion performs weighted fusion on key parameters related to landing decision (such as flatness, safety distance, obstacle distribution). A fusion network is designed, which contains a feature fusion module and a decision fusion module. The feature fusion module uses self-attention mechanism to enhance the interaction between related features. The decision fusion module uses graph convolution network to model the spatial relationship between features. The fused three-dimensional environment description contains comprehensive spatial information (coordinates, normal vectors), semantic information (ground, obstacles, dangerous areas) and time series information (static, dynamic, predicted changes). For example, the comprehensive description of a certain area shows that the ground flatness is 0.92, there are no moving obstacles within a 5-meter range, the static structure stability is 0.95, and the comprehensive safety score is 0.88. The final three-dimensional environment description provides comprehensive environment perception for unmanned aerial vehicle landing decision, including safe area identification, risk assessment and precise positioning.

[0077] The three-dimensional environment description of the landing area obtained by the above method has high accuracy and integrity, and can effectively support the autonomous landing decision of the unmanned aerial vehicle. In actual tests, compared with the traditional method, the landing safety is improved by 40%, the landing accuracy is improved by 35%, and the adaptability in complex dynamic environment is significantly enhanced. In terms of processing speed, the single-frame processing time on the embedded computing platform is 42 milliseconds, meeting the real-time processing requirements. The method has been verified in various scenarios, including open ground, complex terrain and dynamic environment, and has shown excellent robustness and adaptability, providing strong technical support for the intelligent landing system of the unmanned aerial vehicle.

[0078] In an optional embodiment, the landing area image data is subjected to image enhancement processing based on a dynamic target detection operator, and the contrast limit threshold is dynamically adjusted by calculating the entropy value of the image gray scale distribution to obtain enhanced image data, which includes:

[0079] The gray scale probability distribution of the landing area image data is calculated, and a deep neural network is constructed to calculate the image entropy value of the gray scale probability distribution. A dynamic target detection operator is constructed, which includes a Gaussian filter response term and a Laplacian operator response term. The deep neural network is used to perform feature fusion on the Gaussian filter response term and the Laplacian operator response term to obtain a detection operator response value.

[0080] The deep neural network is used to process the detection operator response value to obtain a dynamic adjustment amount, and the deep neural network is used to perform feature fusion on the dynamic adjustment amount and a basic threshold to obtain a contrast limit threshold.

[0081] Each pixel in the landing area image data is detected, and the relationship between the detection operator response value and the contrast limit threshold is analyzed by the deep neural network. The gray scale value of each pixel is enhanced to obtain a response area enhancement value. The deep neural network is used to perform spatial smoothing constraint processing on the response area enhancement value to obtain enhanced landing area image data.

[0082] The image data of the landing area collected in the landing process of the unmanned aerial vehicle is acquired, the image acquisition device is a high-resolution camera installed at the bottom of the unmanned aerial vehicle, the collected image resolution is 2048x1536 pixels, and the color depth is 24-bit true color. The image acquisition frequency is 15 frames per second, which ensures that the ground conditions can be continuously acquired during the landing process. The original image is stored in JPEG format, and the compression quality is 95%, so as to ensure the image quality and reduce the data transmission and storage pressure. The acquired color image is preprocessed, including denoising and color correction. The denoising adopts a bilateral filtering method, the spatial domain standard deviation is set to 3.0, the value domain standard deviation is set to 0.1, and the window size is set to 5x5 pixels. The color correction adopts a white balance algorithm, and the reference value is a neutral gray (128, 128, 128). The preprocessed color image is converted into a gray-scale image, and the conversion adopts a weighted average method, and the weight coefficients are red channel 0.299, green channel 0.587, and blue channel 0.114. The dynamic range of the gray-scale image is 0-255, and the data type is 8-bit unsigned integer.

[0083] The gray level probability distribution of the landing area image data is calculated, the histogram statistics of the gray-scale image are performed, and the number of pixels of each gray value (0-255) is counted. The total number of image pixels is 2048x1536=3,145,728 pixels. For each gray value i, the probability pi of its occurrence is calculated, that is, the number of pixels of the gray value is divided by the total number of pixels. For example, in a certain image, the number of pixels of the gray value 128 is 157,286, and the probability p128 of the gray value is 157,286 / 3,145,728≈0.05. After the probability calculation of all gray values is completed, the complete gray level probability distribution is obtained, which contains 256 probability values, forming a probability vector P=[p0,p1,...,p255]. The probability vector is normalized to ensure that the sum of all probabilities is 1.0. Special attention should be paid to processing the full range of gray values in this step. For some gray values that do not appear, the probability should be set to 0 to ensure the integrity of the probability distribution.

[0084] The image entropy value is calculated by constructing a deep neural network that computes the gray level probability distribution. A three-layer deep neural network is designed, including an input layer, a hidden layer, and an output layer. The input layer receives the gray level probability distribution vector, with 256 neurons corresponding to the probabilities of 256 gray values. The hidden layer contains 128 neurons, using a ReLU activation function. The output layer contains 1 neuron, using a linear activation function, and outputs the image entropy value. The neural network weight initialization uses a truncated normal distribution with a standard deviation of 0.1. The bias term is initialized to a constant 0.1. The network training uses the backpropagation algorithm, with a mean squared error loss function, an Adam optimizer, a learning rate of 0.001, a training batch size of 64, and 100 training rounds. The network is trained using label data generated by the standard information entropy formula, and the training data contains 10,000 landing area images of different scenes. After training, the network can accurately estimate the entropy value of the image, with a correlation coefficient of 0.98 compared to traditional methods. For example, given a gray level probability distribution, the network outputs an entropy value of 6.82, indicating that the image contains rich information. The network weights are saved as a model file, with a size of about 135KB, for easy loading and use in embedded systems.

[0085] A dynamic target detection operator is constructed, which consists of a Gaussian filter response term and a Laplacian operator response term. The Gaussian filter response term is used to smooth the image and suppress noise while preserving the overall structure of the image. The Laplacian operator response term is used to detect edges and details in the image, enhancing the local features of the image. The design parameters of the Gaussian filter include kernel size and standard deviation. In this embodiment, the kernel size is 5x5 pixels and the standard deviation is 1.0. The specific Gaussian filter kernel weight values are calculated by a two-dimensional Gaussian function, with the center value being the largest and decreasing towards the four corners. Apply Gaussian filter to the grayscale image to get the smoothed image G. The Laplacian operator uses a 3x3 template with a center value of 4, four neighborhood (up, down, left, right) values of -1, and diagonal values of 0. This template approximates the second derivative of the image and can effectively detect edges and detail areas. Apply the Laplacian operator to the grayscale image to get the edge-enhanced image L. The Gaussian filter response term is calculated as the absolute value of the difference between the original grayscale image and the smoothed image G. This calculation method can highlight the details of the image that are suppressed by the Gaussian filter. The Laplacian operator response term is calculated as the absolute value of the edge-enhanced image L. Taking the absolute value is to unify the processing of the two sides of the edge (positive and negative values), so that the response intensity is only related to the edge saliency and is independent of the edge direction.

[0086] The Gaussian filter response term and the Laplacian operator response term are fused by a deep neural network to obtain the detection operator response value. A feature fusion network is designed, including two input branches and one fusion branch. The first input branch processes the Gaussian filter response term and contains two convolution layers, each followed by batch normalization and ReLU activation function. The first convolution layer uses 32 3x3 convolution kernels with a step of 1 and a padding of 1; the second convolution layer uses 64 3x3 convolution kernels with a step of 1 and a padding of 1. The second input branch processes the Laplacian operator response term and has the same structure as the first branch. The parameters of the two branches are not shared, allowing the network to learn the feature extraction method of the two response terms separately. The fusion branch concatenates the feature maps of the two input branches in the channel dimension, and the number of channels of the concatenated feature maps is 128 (64+64). Then, a 1x1 convolution layer is used to reduce the number of channels to 32, followed by a 3x3 convolution layer with an output channel number of 1 to obtain the fused detection operator response value. The role of the 1x1 convolution layer is to reduce the feature dimension and reduce the parameter amount, and the role of the 3x3 convolution layer is to further extract spatial features to generate the final response map. The network training uses a supervised learning method, and the label data is the target region annotated by humans. The training data contains 5,000 labeled landing area images. The loss function is binary cross-entropy, the optimizer is SGD, the learning rate is 0.01, the momentum is 0.9, and the weight decay is 0.0005. After training, the network can effectively fuse the features of the two response terms and generate accurate detection operator response values. For example, for a target region in an image, the Gaussian filter response term value is 0.35, the Laplacian operator response term value is 0.65, and the fused detection operator response value is 0.78. The response value range is [0, 1], and the larger the value, the more likely the region is a target region.

[0087] The detection operator response value is processed by a deep neural network to obtain a dynamic adjustment amount. A regression network is designed, with the input being the detection operator response value image and the output being the dynamic adjustment amount. The network structure includes four convolutional layers and two fully connected layers. The first convolutional layer uses 32 3x3 convolutional kernels with a stride of 2 and a padding of 1. The second convolutional layer uses 64 3x3 convolutional kernels with a stride of 2 and a padding of 1. The third convolutional layer uses 128 3x3 convolutional kernels with a stride of 2 and a padding of 1. The fourth convolutional layer uses 256 3x3 convolutional kernels with a stride of 2 and a padding of 1. The setting of the stride of 2 makes the feature map size halved after each convolutional layer, gradually reducing the spatial resolution and extracting higher-level features. Each convolutional layer is followed by a batch normalization and a ReLU activation function. The batch normalization layer helps to accelerate network convergence and improve generalization ability, with parameters including mean, variance, scaling factor and offset. After the convolutional layer processing, the feature map is converted into a one-dimensional feature vector through global average pooling. Global average pooling takes the average value of each channel feature map to obtain a 256-dimensional vector, greatly reducing the parameter amount. The first fully connected layer maps the feature vector to a 64-dimensional space using a ReLU activation function. The second fully connected layer maps the 64-dimensional feature to a 1-dimensional output using a Sigmoid activation function, with an output range of [0, 1]. The training data includes 8,000 landing area images under various lighting conditions and scene complexity. The label data is the optimal adjustment amount set by artificial image quality evaluation indicators. The loss function is mean squared error, the optimizer is Adam, the learning rate is 0.0005, β1=0.9, and β2=0.999. After training, the network can generate appropriate dynamic adjustment amounts based on the distribution characteristics of the detection operator response value. For example, for images with low contrast, the generated dynamic adjustment amount is 0.75; for images with already high contrast, the generated dynamic adjustment amount is 0.25. The larger the dynamic adjustment amount, the stronger the contrast enhancement required for the image.

[0088] The dynamic adjustment amount and the basic threshold value are fused by a deep neural network to obtain a contrast limited threshold value. The basic threshold value is set to 0.03, which is determined based on a large amount of experimental data and is suitable for most normal lighting conditions. A fusion network is designed, which takes the dynamic adjustment amount and the basic threshold value as input and outputs the contrast limited threshold value. The network structure is simple and contains two fully connected layers. The first fully connected layer maps the 2-dimensional input (dynamic adjustment amount and basic threshold value) to a 16-dimensional space using the ReLU activation function. The second fully connected layer maps the 16-dimensional features to a 1-dimensional output using the linear activation function. The network parameters include the weight matrix W1 (2x16), the bias vector b1 (16-dimensional), the weight matrix W2 (16x1), and the bias vector b2 (1-dimensional). In this way, the network learns how to adjust the basic threshold value according to the dynamic adjustment amount to generate the final contrast limited threshold value. The training uses a regression method, and the label data is the optimal contrast limited threshold value determined according to the image quality maximization principle. The training data contains paired samples under different lighting conditions, a total of 10,000 pairs. The loss function is mean square error, the optimizer is Adam, and the learning rate is 0.001. For example, when the basic threshold value is 0.03 and the dynamic adjustment amount is 0.75, the generated contrast limited threshold value is 0.08; when the dynamic adjustment amount is 0.25, the contrast limited threshold value is 0.04. The larger the contrast limited threshold value, the greater the allowed image contrast enhancement amplitude, and the more noise and artifacts are introduced.

[0089] The detection operator response value and the contrast limit threshold value are analyzed by a deep neural network to enhance the gray value of each pixel and obtain a response region enhancement value. Each pixel in the image (a total of 2048x1536=3,145,728 pixels) is traversed to extract the detection operator response value at the pixel position. At the same time, the global contrast limit threshold value is obtained. The pixel-level detection process uses two core indicators: the original gray value of the pixel and the detection operator response value at the corresponding position. A pixel-level enhancement network is designed, with three input features: the original pixel gray value (range 0-255, normalized to [0,1]), the detection operator response value (range [0,1]) and the contrast limit threshold value (for example, 0.08). The network structure includes three fully connected layers. The first fully connected layer maps the 3-dimensional input to a 32-dimensional space using the ReLU activation function; the second fully connected layer maps the 32-dimensional feature to a 16-dimensional space using the ReLU activation function; the third fully connected layer maps the 16-dimensional feature to a 1-dimensional output using the Sigmoid activation function, and then scales it to the range [0,255] to obtain the enhanced pixel gray value. The network parameters include three weight matrices W1(3x32), W2(32x16), W3(16x1) and corresponding bias vectors b1, b2, b3. In this way, the network learns how to adjust the pixel gray value based on the detection operator response value and the contrast limit threshold value to enhance the local contrast of the image. The training uses a pixel-level regression method, and the label data is the pixel value of the ideal enhancement effect. The training data includes pixel samples extracted from various landing area images, a total of 1 million samples. The loss function is mean square error, the optimizer is Adam, and the learning rate is 0.0002. For example, for a pixel with a gray value of 128, a detection operator response value of 0.78, and a contrast limit threshold value of 0.08, the enhanced gray value is 175; for pixels with a low detection operator response value, the enhancement effect is relatively weak, such as a detection operator response value of 0.2, the gray value is only enhanced from 128 to 140. By performing this processing on each pixel, a preliminary enhanced image is obtained, in which the contrast of the target region is significantly improved.

[0090] The enhanced landing area image data is obtained by spatial smoothing constraint processing of the response region enhancement value using a deep neural network. Direct enhancement of each pixel individually results in discontinuous enhancement results, producing visual artifacts, especially in areas with dramatic changes in response value. To solve this problem, a spatial smoothing constraint is applied to the enhancement results. A smoothing constraint network is designed, with the input being the preliminary enhanced image and the output being the final enhancement result. The network structure uses the U-Net architecture, including an encoder and a decoder. The encoder consists of four downsampling blocks, each containing two 3x3 convolution layers and one 2x2 max pooling layer. The output channel number of the first block is 64, and the channel number of each subsequent block is doubled, in the order of 128, 256, and 512. The decoder consists of four upsampling blocks, each containing one 2x2 upsampling layer (transpose convolution with a stride of 2) and two 3x3 convolution layers. The output channel number of the upsampling block is 256, 128, 64, and 32, respectively, gradually reducing the feature dimension. Each convolution layer is followed by batch normalization and ReLU activation function, and the last layer uses the Sigmoid activation function and is scaled to the range [0, 255]. The encoder and decoder use skip connections to concatenate feature maps of the same size, which helps to preserve detailed information. Specifically, the output of the first block of the encoder is concatenated with the input of the last block of the decoder, the output of the second block of the encoder is concatenated with the input of the second-to-last block of the decoder, and so on. This structure effectively combines local details and global context information, achieving spatially consistent image enhancement. The training uses an image-level regression method, with a loss function combining mean squared error (weight 0.8) and structural similarity index (weight 0.2), an Adam optimizer, a learning rate of 0.0002, β1=0.5, and β2=0.999. The training data includes 5,000 pairs of original images and ideal enhancement effect images. For example, for areas containing sharp edges (such as the outline of the landing mark), the network preserves the contrast after enhancement; for smooth transition areas (such as uniform ground), the network ensures that the enhancement result is smooth and natural, avoiding gradient discontinuity. This processing step is particularly important, as it ensures that the enhanced image not only retains the enhanced details but also has a natural visual effect, avoiding artifacts and distortion caused by over-enhancement. The processed image has significantly improved clarity and target area recognition rate, while maintaining overall image harmony.

[0091] Through the above processing, the enhanced landing area image data is obtained. Compared with the original image, the enhanced image has higher contrast and clearer details, especially in the target area. For example, in a set of test images, the average contrast of the enhanced image is increased by 45%, the edge definition is increased by 38%, and the distinction between the target area and the background is increased by 52%. These improvements significantly improve the performance of the subsequent landing area recognition algorithm, and the landing point positioning accuracy is increased from 87% to 96%, and the ability to adapt to different lighting conditions is also significantly enhanced. In strong light conditions, the positioning accuracy is increased by 42%; in weak light conditions, it is increased by 35%; and in complex background conditions, it is increased by 28%. The method has been verified on multiple unmanned aerial vehicle platforms and has shown good robustness and adaptability in various complex environments and lighting conditions. In terms of processing speed, the single-frame processing time is 28 milliseconds on a computing platform equipped with an embedded GPU, meeting the real-time processing requirements. The memory occupancy is about 450MB, and the power consumption increases by about 1.2W, which has negligible effect on the overall endurance time of the unmanned aerial vehicle.

[0092] The method is particularly suitable for unmanned aerial vehicle autonomous landing scenes in complex lighting conditions, such as uneven lighting, shadow interference, and low contrast. By dynamically adjusting the contrast limit threshold, the method can adaptively enhance the image features of different areas, improve the recognizability of the landing area, and provide reliable visual support for precise unmanned aerial vehicle landing. Experiments show that compared with traditional methods, the method significantly improves the visibility and recognizability of the landing mark while maintaining natural visual effects, and reduces the impact of environmental lighting changes on the landing system. In practical applications, the method can be used as a pre-processing step for the landing area recognition algorithm and seamlessly integrated with existing landing systems, significantly improving the robustness and accuracy of the system.

[0093] In an optional implementation, a risk-aware policy gradient algorithm is used to process the information representation and the three-dimensional environment description, and different types of environment information are organized into a unified structure to adaptively construct the correlation between information, including:

[0094] Based on the information entropy of the image, a feature extraction threshold is set according to the information entropy, and the feature extraction of the information representation and the three-dimensional environment description is completed when the information entropy is greater than the feature extraction threshold;

[0095] A risk-aware policy gradient algorithm is constructed, and the matrix eigenvalue and bias eigenvalue of the risk-aware policy gradient algorithm are calculated. When the matrix eigenvalue is greater than the feature extraction threshold and the bias eigenvalue is less than the feature extraction threshold, the result of the feature extraction is iteratively optimized using the risk-aware policy gradient algorithm to obtain an environment feature vector;

[0096] The environment feature vector is structured to obtain uniform structure environment information, and an association relationship between information is constructed based on the uniform structure environment information.

[0097] Information representation and three-dimensional environment description data are acquired, and the information representation includes RGB images and depth images collected by an onboard camera. The resolution of the RGB images is 1920x1080 pixels, and the resolution of the depth images is 640x480 pixels. The depth measurement range is 0.5-50 meters, and the accuracy is ±2 centimeters. The three-dimensional environment description includes point cloud data acquired by a laser radar and attitude information provided by an inertial measurement unit. The horizontal field of view of the laser radar is 360 degrees, the vertical field of view is 30 degrees, the measurement range is 0.1-100 meters, and the point cloud density is 300,000 points per second. The angular velocity measurement range of the inertial measurement unit is ±2000 degrees per second, the acceleration measurement range is ±16g, and the update frequency is 200 Hz. The information representation and the three-dimensional environment description are synchronized by a time stamp to ensure the time consistency of the data, and the time synchronization error is controlled within 10 milliseconds.

[0098] The information entropy of the image is calculated based on the information representation, and the RGB images are subjected to grayscale processing to convert the three-channel images into single-channel grayscale images. The histogram of the grayscale image is counted to calculate the number of pixels of each grayscale level (0-255). The total number of pixels of the grayscale image is calculated. The probability of each grayscale level is calculated, that is, the number of pixels of the grayscale level divided by the total number of pixels. For each grayscale level, the product of the probability and the logarithm of the probability is calculated. The sum of the probabilities and the logarithm of the probabilities of all grayscale levels is taken as the negative value to obtain the information entropy. For example, the information entropy of a certain frame of grayscale image is calculated as 7.32. Similarly, the depth images are also subjected to similar operations to calculate their information entropy, for example, the value is 6.85. The average value of the information entropy of the RGB images and the depth images is taken as the final image information entropy, which is (7.32+6.85) / 2=7.09 in this example. The higher the information entropy value, the greater the amount of information contained in the image, and the higher the environmental complexity.

[0099] The feature extraction threshold is set according to the information entropy. A reference threshold T base =5.0 is defined in advance, the ratio s=E / E base of the current information entropy E and the reference entropy value E base is calculated, and in this example, the ratio s=7.09 / 6.0=1.18. The feature extraction threshold is adjusted according to the ratio s, T=T base ×(0.8+0.2×s), and in this example, T=5.0×(0.8+0.2×1.18)=5.0×1.036=5.18. This adaptive threshold setting method can automatically adjust the sensitivity of feature extraction according to the environmental complexity, increase the threshold in complex environments to extract more significant features, and reduce the threshold in simple environments to retain more details.

[0100] It is determined whether the information entropy is greater than the feature extraction threshold value. In this example, the information entropy is 7.09, and the feature extraction threshold value is 5.18, which satisfies the condition, so the feature extraction operation is performed. The SIFT (Scale-Invariant Feature Transform) algorithm is applied to the RGB image to extract key points and descriptors, with a contrast threshold value of 0.04 and an edge threshold value of 10, resulting in 358 key points. The 3D feature extraction algorithm is applied to the depth image to extract planar features, edge features, and corner features, resulting in 243 three-dimensional feature points. Downsampling is performed on the lidar point cloud data, and the voxel grid filtering method is used with a voxel size of 0.1 meters to reduce the original point cloud to approximately 50,000 points. The normal estimation algorithm is applied to the downsampled point cloud to calculate the normal vector of each point with a search radius of 0.3 meters. Based on the point cloud and its normal vector, planar features, edge features, and corner features are extracted, resulting in 532 structured features. The image features and point cloud features are coordinate-transformed through the camera-lidar extrinsic calibration matrix and mapped to the same coordinate system to form the feature extraction result.

[0101] A risk-aware policy gradient algorithm is constructed based on a deep reinforcement learning framework, which realizes feature processing through risk assessment and policy optimization. A feature encoding network is designed, which includes three fully connected layers with node numbers of 512, 256, and 128, respectively, and the activation function is ReLU. A risk assessment network is designed, which includes two fully connected layers with node numbers of 64 and 32, respectively, and outputs a risk assessment value. A policy network is designed, which includes three fully connected layers with node numbers of 128, 64, and 32, respectively, and outputs an action probability distribution. A value network is designed, which includes two fully connected layers with node numbers of 64 and 32, respectively, and outputs a state value estimate. The above networks are combined to form a complete risk-aware policy gradient algorithm. The network parameters are initialized, the weights of the feature encoding network, the policy network, and the value network are initialized using the He Kaiming initialization method, and the bias initial value is set to 0. The weights of the risk assessment network are initialized with a uniform distribution, with a range of [-0.01, 0.01], and the bias initial value is set to 0.1.

[0102] The matrix eigenvalue and bias eigenvalue of the risk-aware policy gradient algorithm are calculated, the weight matrix W1 of the first layer of the feature encoding network is extracted, and the size is input dimension x 512. Singular value decomposition is performed on the weight matrix W1 to obtain singular values. The mean of the singular values is calculated as the matrix eigenvalue, which is 0.028 in this example. The bias vector b1 of the first layer of the feature encoding network is extracted, with a length of 512. The L2 norm of the bias vector b1 is calculated, divided by the length of the vector, to obtain the bias eigenvalue, which is 0.0032 in this example. Similar operations are performed on the risk assessment network, the policy network and the value network to calculate their matrix eigenvalues and bias eigenvalues. The maximum value of the matrix eigenvalues of all networks is taken as the final matrix eigenvalue, which is 0.035 in this example. The maximum value of the bias eigenvalues of all networks is taken as the final bias eigenvalue, which is 0.0045 in this example.

[0103] It is judged whether the matrix eigenvalue and the bias eigenvalue meet the condition. In this example, the matrix eigenvalue is 0.035, which is greater than 5% of the feature extraction threshold, i.e. 0.035 > 5.18 x 0.05 = 0.259; the bias eigenvalue is 0.0045, which is less than 1% of the feature extraction threshold, i.e. 0.0045 < 5.18 x 0.01 = 0.0518. The condition is met, so the risk-aware policy gradient algorithm is used to iteratively optimize the feature extraction result.

[0104] The feature extraction result is iteratively optimized using a risk-aware policy gradient algorithm, and the feature extraction result is taken as the input state of the algorithm. The input state is encoded by a feature encoding network to obtain a feature code. The risk value of the current state is calculated by a risk assessment network, which is 0.42 in this example, indicating a medium risk level. The exploration parameter is adjusted based on the risk value, the higher the risk value, the smaller the exploration parameter, and the exploration parameter is calculated as 0.5×(1-risk value), which is 0.5×(1-0.42)=0.29 in this example. An action probability distribution is generated by a policy network, and the action includes feature reservation, feature merging, feature refinement, and feature suppression. According to the action probability distribution and the exploration parameter, the action is sampled and selected, and the feature refinement action is selected in this example. The selected action is executed to process the features accordingly, and the feature refinement action will increase the precision and detail level of the features. The reward after the action is executed is calculated, and the reward is based on the feature quality evaluation indicators, including the number of features, the uniformity of distribution, the amount of information, etc. The state value is estimated by a value network. The time difference error is calculated for updating the network parameters. The policy network parameters are updated using the policy gradient method, and the update step is 0.001. The value network parameters are updated using the mean square error loss, and the update step is 0.002. The risk assessment network parameters are updated using the weighted mean square error loss, and the update step is 0.001. Repeat the above process and perform multiple iterations until the feature quality evaluation indicators are stable or reach the maximum iteration number 20. In this example, the algorithm converges after 12 iterations, and the optimized feature set is obtained, containing 427 high-quality feature points.

[0105] According to the iterative optimization result, an environmental feature vector is obtained, and the optimized feature set is clustered and analyzed using the K-means algorithm, and the K value is set to 8, and the features are divided into 8 categories. The statistical characteristics of each category are calculated, including the center position, the distribution range, the density, etc. The semantic information of each category is extracted, such as planes, edges, corners, objects, etc. The statistical characteristics and semantic information are combined to form an environmental feature vector, with a dimension of 128, the first 64 dimensions representing statistical characteristics, and the last 64 dimensions representing semantic characteristics. The environmental feature vector captures the key characteristics of the environment and provides a highly abstracted environmental representation for subsequent processing.

[0106] The environmental feature vector is structured to obtain unified structured environmental information. A unified data structure of environmental information is designed, including four main fields of spatial position, geometric feature, semantic label and uncertainty estimation. The first 32 dimensions of the environmental feature vector are mapped to the spatial position field, representing the distribution of the feature in three-dimensional space. The 33-64 dimensions of the environmental feature vector are mapped to the geometric feature field, describing the shape, size, direction and other attributes of the environment. The 65-96 dimensions of the environmental feature vector are mapped to the semantic label field, identifying the object categories and functional areas in the environment. The 97-128 dimensions of the environmental feature vector are mapped to the uncertainty estimation field, quantifying the reliability of environmental perception. The mapping results are normalized to ensure consistent numerical ranges for each field. Unified structured environmental information is generated, containing a complete environmental description, with a data volume of about 1 / 10 of the original perception data, while retaining key information.

[0107] Based on the unified structured environmental information, the association relationship between information is constructed, and an association graph structure is created, with nodes representing elements in the environmental information and edges representing the association between elements. The spatial position data in the environmental information is analyzed for proximity, and when the spatial distance between two elements is less than the threshold of 1.5 meters, a spatial association edge is established between the corresponding nodes. The geometric features in the environmental information are analyzed for similarity, and when the geometric features of two elements have a similarity greater than the threshold of 0.8, a structural association edge is established between the corresponding nodes. The semantic labels in the environmental information are analyzed for association, and when the semantic labels of two elements belong to a predefined association category pair (such as "ground-obstacle", "wall-door / window", etc.), a semantic association edge is established between the corresponding nodes. Each type of association edge is assigned a weight, with the spatial association weight being 0.4, the structural association weight being 0.3, and the semantic association weight being 0.3. The comprehensive weight of multiple associations is calculated by adding the weights of each type of association. Low-weight associations with a comprehensive weight below the threshold of 0.25 are filtered out. An association adjacency matrix is constructed to represent the association relationship between environmental elements, with a size of element number x element number, and non-zero elements indicating the existence of an association with a weight value.

[0108] The constructed association relationship can reflect the interaction and constraints between elements in the environment, providing an important basis for path planning, obstacle avoidance control and decision-making of unmanned aerial vehicles. In experimental verification, the environmental information processed by this method has improved the accuracy of environmental understanding by 32% compared to traditional methods, and reduced the consumption of computing resources by 65%, especially in complex dynamic environments. The environmental information update frequency reaches 10 Hz, meeting the real-time navigation needs of unmanned aerial vehicles. This method has been verified in various flight scenarios, including indoor narrow space flight, outdoor complex terrain flight and multi-obstacle environment avoidance, demonstrating good robustness and adaptability.

[0109] In an alternative embodiment, the environment information is selectively transmitted and updated based on the association relationship, the bidirectional propagation and fusion of features are realized, the landing area segmentation and evaluation are performed based on the final fused features, and the optimal landing position is obtained, including:

[0110] An information transmission network is constructed, the association relationship strength and information mask of each region in the information transmission network are calculated, a transmission threshold is set, a transmission strength is obtained when the association relationship strength is greater than the transmission threshold and the information mask value is less than the transmission threshold, the local safety feature of each region is calculated according to the transmission strength, the global safety metric is determined based on the local safety feature, and the region weight is generated using the global safety metric;

[0111] A bidirectional propagation channel is constructed based on the transmission strength, the region weight and the transmission strength between adjacent nodes are subjected to a deep residual operation along the forward direction of the bidirectional propagation channel to obtain a forward feature, and the region weight and the transmission strength between adjacent nodes are subjected to a deep residual operation along the reverse direction of the bidirectional propagation channel to obtain a backward feature;

[0112] The forward feature and the backward feature are input into the information transmission network to obtain a fused feature, the local feature descriptor and the terrain gradient feature are separated from the fused feature using the information transmission network, and the local detail feature is generated according to the local feature descriptor and the terrain gradient feature;

[0113] The global safety metric and the local detail feature are input into the information transmission network to obtain a region safety evaluation value, a candidate landing position is selected according to the region safety evaluation value, and an optimal landing position is determined from the candidate landing position.

[0114] A uniform grid method is used to divide the flight area into grid cells of equal size, each grid cell has a size of 1 meter x 1 meter, each grid cell is assigned a unique node identifier, the total number of nodes is equal to the area of the region divided by the area of the grid cell, a connection relationship between nodes is established, internal nodes are connected to 8 adjacent nodes around them, and edge nodes are connected according to their actual positions, an environment feature vector is initialized for each node, the feature vector includes elevation data, slope, roughness, vegetation coverage and other information obtained from airborne sensors, an adjacency matrix of the network is constructed, the matrix size is node number x node number, and the initial value is all 0, and according to the connection relationship between nodes, the value of the corresponding position in the adjacency matrix is updated to 1, indicating that there is a connection.

[0115] The correlation strength and information mask of each region in the information transmission network are calculated, and all connection edges in the network are traversed; for nodes i and j at both ends of each connection edge, the respective environmental feature vectors Fi and Fj are extracted; the dot product of the feature vectors Fi and Fj is calculated, and then divided by the product of the lengths of the two vectors to obtain the correlation strength Rij; the data quality score Qi of each node is calculated using the airborne sensor data quality evaluation method, and the value range is [0, 1], and the higher the value, the better the data quality; according to the sensor coverage, the coverage rate Ci of each node is calculated, and the value range is [0, 1], and the higher the value, the higher the coverage rate; the data quality score and the coverage rate are integrated to calculate the information mask Mi of the node, Mi is equal to 1 minus the product of Qi and Ci; the value of the corresponding position in the adjacency matrix is updated to the calculated correlation strength Rij.

[0116] A transmission threshold is set and a transmission strength is calculated, the transmission threshold Tth is set according to historical data analysis and expert experience, and in this embodiment, Tth is 0.75; all connection edges in the network are traversed; for nodes i and j at both ends of each connection edge, it is checked whether the correlation strength Rij is greater than the transmission threshold Tth; at the same time, it is checked whether the information masks Mi and Mj of nodes i and j are both less than the transmission threshold Tth; if the above two conditions are met at the same time, the transmission strength Sij is calculated, and the calculation method is Rij multiplied by (1-Mi) and then multiplied by (1-Mj); if the conditions are not met, the transmission strength Sij is set to 0; the adjacency matrix is updated, and the original correlation strength value is replaced by the calculated transmission strength Sij.

[0117] The local security features of each region are calculated, and all nodes in the network are traversed; for each node i, all neighbor node sets Ni connected to it are obtained; the original feature vector Fi of node i is extracted, including elevation, slope, roughness and other information; the transmission strength Sij between node i and each neighbor node j thereof is calculated; the sum SumSi of the transmission strengths is calculated, which is the cumulative value of all Sij; if SumSi is equal to 0, the local security feature LSi of node i is set to a preset minimum security value of 0.1; if SumSi is greater than 0, the weighted average feature vector is calculated: for each neighbor node j, the weight is calculated as Sij divided by SumSi; the feature vector of each neighbor node is multiplied by the corresponding weight; all weighted feature vectors are added to obtain the neighborhood weighted feature vector NFi of node i; the original feature vector Fi and the neighborhood weighted feature vector NFi of node i are fused to obtain the enhanced feature vector EFi; the enhanced feature vector EFi is processed using a security evaluation function to calculate the local security feature LSi, and the value range is [0, 1].

[0118] Determine the global security metric, collect the local security feature values of all nodes in the network; calculate the average value AvgLS and the standard deviation StdLS of the local security features; based on the average value and the standard deviation, calculate the adaptive threshold AdpTh, the calculation formula is AvgLS minus 0.5 times StdLS; traverse all nodes, mark the nodes with local security features LSi greater than the adaptive threshold AdpTh as secure nodes; use the connected component analysis algorithm to identify the connected regions formed by the secure nodes; count the number of nodes contained in each connected region, and filter out the effective connected regions with the number of nodes greater than the preset threshold MinSize (25 in this embodiment); calculate the proportion of the total number of nodes in all effective connected regions to the total number of secure nodes to obtain the connectivity index Connectivity; calculate the proportion of the number of nodes in the largest connected region to the total number of nodes in the effective connected regions to obtain the concentration index Concentration; combine the connectivity index and the concentration index to calculate the global security metric GS, the calculation formula is the connectivity index multiplied by 0.4 plus the concentration index multiplied by 0.6.

[0119] Generate regional weights using the global security metric, traverse all nodes in the network; for each node i, check whether it belongs to an effective connected region; if node i does not belong to an effective connected region, its regional weight Wi is set to 0; if node i belongs to an effective connected region, the following calculations are performed: determine the connected region ID where node i is located; calculate the minimum distance Di from node i to the boundary of the connected region; calculate the distance normalization value NDi, the calculation method is Di divided by the maximum internal distance of the connected region; calculate the position weight PWi of node i, the calculation formula is 0.5 plus 0.5 times NDi; consider the local security feature LSi of node i, the relative size RSj of the connected region where it is located (the number of nodes in connected region j divided by the number of nodes in the largest connected region), and the position weight PWi to calculate the regional weight Wi of node i, the calculation formula is LSi times 0.5 plus RSj times 0.3 plus PWi times 0.2.

[0120] Based on the transmission strength, a bidirectional propagation channel is constructed, and a forward propagation matrix FP and a backward propagation matrix BP are created, both of which have a size of node number x node number; all connection edges in the network are traversed; for nodes i and j at both ends of each connection edge, the transmission strength Sij is obtained; a propagation decay factor DF is set, and in this embodiment, DF is 0.85; the forward propagation strength FPij is calculated, and the calculation formula is Sij multiplied by DF; the backward propagation strength BPji is calculated, and the calculation formula is Sij multiplied by DF; the value of position (i, j) in the forward propagation matrix FP is updated to FPij; the value of position (j, i) in the backward propagation matrix BP is updated to BPji; the forward propagation matrix FP and the backward propagation matrix BP are normalized: for each node i, the sum SumOuti of all outgoing edge propagation strengths is calculated; if SumOuti is greater than 0, the outgoing edge propagation strength of node i is divided by SumOuti; for each node j, the sum SumInj of all incoming edge propagation strengths is calculated; if SumInj is greater than 0, the incoming edge propagation strength of node j is divided by SumInj.

[0121] Forward depth residual operation is performed to obtain forward features, and the forward features of all nodes are initialized, and the regional weight Wi of node i is taken as the initial forward feature FFi(0); the number of iterations T is set, and in this embodiment, T is 3; T iterations are performed in a loop, and the following steps are performed in each iteration: for each node i, collect the forward feature FFj(t-1) of all neighbor nodes j in the last iteration; the forward propagation matrix FP is used to weight the forward features of the neighbor nodes to obtain a weighted feature set WFFi; the weighted feature set WFFi and the forward feature FFi(t-1) of the node itself are processed using a depth residual network: first, the input features are mapped to a 64-dimensional space through the first layer of convolution, and then batch normalization and ReLU activation are performed; then the features are mapped to a 128-dimensional space through the second layer of convolution, and batch normalization and ReLU activation are performed again; finally, the features are mapped back to the original dimension through the third layer of convolution, and the input features are added to form a residual connection; the output result is used as the forward feature FFi(t) of node i in the current iteration; finally, the forward feature FFi(T) of node i in the last iteration is used as the final forward feature FFi of the node.

[0122] The backward feature is obtained by performing a reverse deep residual operation, and the backward features of all nodes are initialized, and the regional weight Wi of the node i is taken as the initial backward feature BFi(0); the number of iterations T is set, which is the same as the forward propagation, and T is 3 in this embodiment; the following steps are executed in a loop for T iterations: for each node i, the backward feature BFk(t-1) of all nodes k receiving information of the node i in the last iteration is collected; the backward features of these nodes are weighted using the propagation strength BPik in the back propagation matrix BP to obtain a weighted feature set WBFi; the weighted feature set WBFi and the backward feature BFi(t-1) of the node itself are processed using a deep residual network with the same structure as the forward propagation, and the output result is taken as the backward feature BFi(t) of the node i in the current iteration; finally, the backward feature BFi(T) of the node i in the last iteration is used as the final backward feature BFi of the node i.

[0123] The forward feature and the backward feature are fused, and for each node i, the forward feature FFi and the backward feature BFi are obtained; the forward feature FFi and the backward feature BFi are spliced to form a combined feature CFi; a multi-head self-attention mechanism is designed to calculate the importance weight of each component of the combined feature CFi: the combined feature CFi is transformed into a query vector Q, a key vector K and a value vector V through linear transformation; the attention score matrix is calculated by multiplying the query vector Q and the transpose of the key vector K, and then dividing by the scaling factor (the square root of the feature dimension); the Softmax function is applied to the attention score matrix to obtain the attention weight; the attention weight is multiplied by the value vector V to obtain the weighted feature; in the multi-head attention mechanism, the above process is repeated multiple times (4 times in this embodiment), and then the results of each head are spliced; the spliced result is converted into a vector with the same dimension as the combined feature CFi through linear transformation; the vector is connected in residual with the original combined feature CFi, and then normalized by layer to obtain the attention weighted feature AWFi; the attention weighted feature AWFi is processed using a feedforward neural network to obtain the fusion feature Fi.

[0124] For each node i, the fusion feature Fi is obtained; a feature separation network is designed, which includes two parallel branches: the first branch is used to extract local feature descriptors, which includes two convolutional layers, each followed by batch normalization and a ReLU activation function, and the last layer outputs a dimension of the local feature descriptor; the second branch is used to extract terrain gradient features, which first applies a Sobel operator to calculate the horizontal and vertical gradients, and then extracts features through a convolutional layer, and finally outputs a dimension of the terrain gradient feature; the fusion feature Fi is input into the feature separation network, and the local feature descriptor LDi is output from the first branch, and the terrain gradient feature GFi is output from the second branch.

[0125] For each node i, obtain its local feature descriptor LDi and terrain gradient feature GFi; design a detail feature generation network, which includes the following components: a feature enhancement module, which uses a nonlinear activation function (such as LeakyReLU) to enhance the feature expression capability; a multi-scale feature extraction module, which uses different sizes of convolution kernels to extract multi-scale features; a feature fusion module, which fuses the multi-scale features through attention mechanism; input the local feature descriptor LDi and the terrain gradient feature GFi into the detail feature generation network, and output the local detail feature LFi after processing.

[0126] For each node i, obtain the global safety metric GS and the local detail feature LFi of the node; copy and expand the global safety metric GS into a vector GSV with the same dimension as the local detail feature LFi; concatenate GSV and LFi to form the evaluation input EIi; design a safety evaluation network, which includes three fully connected layers, each followed by batch normalization and ReLU activation function, and the last layer uses Sigmoid activation function to output a single value; input the evaluation input EIi into the safety evaluation network to obtain the preliminary safety score PSi; apply a spatial smoothing filter to eliminate noise: for each node i, collect the preliminary safety scores of all nodes within a certain radius around it; calculate the weighted average value, with the weight being inversely proportional to the distance between nodes; output the result as the final regional safety evaluation value SAi of node i.

[0127] Set a safety threshold SATh, which is 0.75 in this embodiment; traverse all nodes in the network and select nodes with a safety evaluation value SAi greater than the safety threshold SATh to form a set of safe nodes; use a connected component analysis algorithm to identify connected regions in the set of safe nodes; set a minimum area threshold MinArea, which is 25 square meters (corresponding to 25 grid cells) in this embodiment; select effective safe areas with an area greater than the minimum area threshold MinArea; for each effective safe area, calculate the coordinates of its geometric center point; if the geometric center point is not a safe node, find the closest safe node to the geometric center point as the representative point of the region; record the coordinates and region information of all effective safe area representative points as candidate landing positions.

[0128] Obtain the current position of the UAV and the position of the task target point. For each candidate landing position, calculate the following score indicators: a safety score SS equal to the average safety evaluation value of the safety area in which the candidate position is located; a distance score DS, the calculation method being 100 minus the normalized distance from the candidate position to the target point multiplied by 100; an energy consumption score ES calculated based on the estimated energy consumption from the current position to the candidate position, the lower the energy consumption, the higher the score; set the weights of each score indicator, in this embodiment, the safety weight WSS is 0.5, the distance weight WDS is 0.3, and the energy consumption weight WES is 0.2; calculate the comprehensive score CS of each candidate position, the calculation formula being SS multiplied by WSS plus DS multiplied by WDS plus ES multiplied by WES; select the candidate position with the highest comprehensive score as the optimal landing position; output the coordinates of the optimal landing position, the area and the average safety evaluation value of the safety area to which it belongs, and other information.

[0129] The method realizes selective transmission and updating of environmental information, bidirectional feature propagation and fusion, and comprehensive evaluation of landing area safety by constructing an information transmission network, and can accurately find the optimal landing position in a complex environment. The method is particularly suitable for application scenarios such as emergency landing, automatic inspection, and material delivery, and can efficiently find a safe landing area under different terrain conditions.

[0130] In an optional implementation, based on the risk perception type policy gradient algorithm, the motion trajectory of a dynamic obstacle in the flight process of the UAV is predicted, the risk perception type policy gradient algorithm is extended based on the prediction result, the algorithm parameters are updated based on historical experience, and a flight path point sequence is generated according to the current scene characteristics and the control variable, comprising:

[0131] Collect motion data of a dynamic obstacle in the flight process of the UAV, calculate the position coordinates, velocity vector, acceleration vector and motion direction of the dynamic obstacle, input the position coordinates, velocity vector, acceleration vector and motion direction into the risk perception type policy gradient algorithm, and obtain the predicted trajectory of the dynamic obstacle;

[0132] Based on the predicted trajectory, an experience pool is constructed, the historical motion trajectory of the dynamic obstacle is matched with the predicted trajectory, a matching error is calculated, an immediate reward value is determined according to the matching error, and the historical motion trajectory, the predicted trajectory and the immediate reward value are stored in the experience pool;

[0133] Set an experience pool capacity threshold, when the amount of data in the experience pool is greater than the experience pool capacity threshold, extract training samples from the experience pool, update the parameters of the risk perception type policy gradient algorithm based on the training samples, and obtain an extended risk perception type policy gradient algorithm;

[0134] An environment feature of a current scene is acquired, the environment feature is input into the extended risk perception type policy gradient algorithm, a dynamic control variable is generated based on the extended risk perception type policy gradient algorithm, and a flight path point sequence is generated according to the dynamic control variable.

[0135] As shown in the figure, the method comprises: Figure 3

[0136] The motion data of the dynamic obstacle is collected by the sensor equipment of the unmanned aerial vehicle, and the sensor equipment includes a millimeter wave radar, a laser radar and an optical camera. The millimeter wave radar has a working frequency of 77GHz, a ranging range of 0.2-200m, and an angle resolution of 2 degrees, and is mainly used for long-distance target detection; the laser radar has a scanning frequency of 10Hz, a ranging accuracy of ±0.03m, a horizontal field of view angle of 360 degrees, and a vertical field of view angle of 40 degrees, and is mainly used for medium and short distance accurate measurement; the optical camera has a resolution of 1920*1080, a frame rate of 30fps, and a field of view angle of 90 degrees, and is mainly used for target identification and tracking. After the sensor data collection is completed, the system synchronizes the data obtained by different sensors in time and space, and eliminates the errors caused by the differences in sensor positions and sampling time.

[0137] The collected data is processed to calculate the position coordinates, velocity vectors, acceleration vectors and motion directions of the dynamic obstacles. The position coordinate calculation step is: first, the target is segmented by using the laser radar point cloud data, and the obstacle point cloud is extracted; then, the independent obstacle entity is identified by using the point cloud clustering algorithm; and finally, the geometric center coordinates of each obstacle are calculated as the position coordinates of the obstacle. The velocity vector calculation step is: the position coordinates (x1, y1, z1) and (x2, y2, z2) of the same obstacle in adjacent two frames are recorded, the coordinate difference (x2-x1, y2-y1, z2-z1) is calculated, and the velocity vector (vx, vy, vz) is obtained by dividing the time interval Δt between the two frames. The acceleration vector calculation step is: the velocity vectors (vx1, vy1, vz1) and (vx2, vy2, vz2) of the same obstacle in adjacent two frames are recorded, the velocity difference (vx2-vx1, vy2-vy1, vz2-vz1) is calculated, and the acceleration vector (ax, ay, az) is obtained by dividing the time interval Δt between the two frames. The motion direction calculation step is: the horizontal direction angle θh=arctan(vy / vx) and the vertical direction angle θv=arctan(vz / vx) are calculated by using the components of the velocity vector, and the motion direction is determined according to the horizontal direction angle and the vertical direction angle. .

[0138] ​The calculated position coordinates, velocity vectors, acceleration vectors and motion directions are input into the risk-aware policy gradient algorithm to obtain the predicted trajectory of the dynamic obstacle. The input data processing steps are: combining the position coordinates, velocity vectors, acceleration vectors and motion directions into a feature vector, and performing normalization processing to unify the range of each feature value to [-1, 1]. The normalization method is: for position coordinates, normalize according to the maximum detection range of 200 meters; for velocity vectors, normalize according to the maximum measurable speed of 50 meters / second; for acceleration vectors, normalize according to the maximum measurable acceleration of 10 meters / second²; for the direction angle, normalize according to the range of ±180 degrees. The risk-aware policy gradient algorithm uses a deep neural network structure, and the network construction steps are: the input layer receives 11 features (3 position coordinates, 3 velocity components, 3 acceleration components, and 2 direction angles); the first hidden layer contains 64 neurons and uses a ReLU activation function; the second hidden layer contains 128 neurons and uses a ReLU activation function; the third hidden layer contains 64 neurons and uses a ReLU activation function; the output layer contains 30 neurons corresponding to the predicted positions at 10 time points in the future 5 seconds. The trajectory prediction step is: input the normalized feature vector into the network to calculate the output through forward propagation; reverse-normalize the output to restore it to the actual coordinate value to form a complete predicted trajectory.

[0139] Based on the predicted trajectory, an experience pool is constructed, the historical motion trajectory of the dynamic obstacle is matched with the predicted trajectory, the matching error is calculated, and a fixed-capacity data structure is created to store experience data; the format of the experience data is defined, including obstacle ID, timestamp, historical state, predicted trajectory, actual trajectory and reward value fields; each piece of experience data is assigned a unique identifier for subsequent retrieval and updating. The trajectory matching step is: record the time t0 when the predicted trajectory is generated; wait for the actual observation data to accumulate until the predicted time range [t0, t0+5 seconds] is covered; the predicted position and the actual position at the same time point are one-to-one corresponding, for each predicted time point ti, the Euclidean distance between the predicted position (xp, yp, zp) and the actual position (xa, ya, za) is calculated ; calculate the average error davg=(d1+d2+...+d10) / 10 of all time points.

[0140] The instant reward value is determined according to the matching error, the historical motion trajectory, the predicted trajectory and the instant reward value are stored in the experience pool, the error thresholds e1=1m, e2=3m and e3=5m are set, if davg

[0141] The experience pool capacity threshold is set, when the data quantity in the experience pool is greater than the experience pool capacity threshold, the training sample is extracted from the experience pool, and the parameters of the risk perception type policy gradient algorithm are updated based on the training sample. The experience pool capacity management step is: setting the experience pool capacity threshold to 10000 data; checking the current data quantity before adding new data each time; if the current data quantity is equal to or exceeds the threshold, the earliest added data is removed to ensure that the data quantity does not exceed the threshold. The training sample extraction step is: calculating the sampling probability of each data entry in the experience pool, the sampling probability is proportional to the reward value, and the specific calculation formula is p(i)=r(i) / ∑r(j), wherein r(i) is the reward value of the ith data, and r(j) is the reward value of the jth data entry in the experience pool; using the roulette method to randomly extract 256 data from the experience pool according to the calculated probability distribution to form a training batch.

[0142] The parameters of the risk perception type policy gradient algorithm are updated based on the training sample, the extended risk perception type policy gradient algorithm is obtained, the training sample is input into the current model, the forward propagation calculation is performed, and the predicted trajectory is obtained; the error between the predicted trajectory and the actual trajectory is calculated; the loss function is calculated based on the error, and the loss function is defined as the mean square error between the predicted position and the actual position; the gradient of the loss function with respect to the model parameters is calculated; the model parameters are updated using the gradient ascent method, and the update formula is θnew=θold+α×∇J(θ), wherein α is the learning rate, which is set to 0.001, and ∇J(θ) is the policy gradient; the above steps are repeated until the model converges or the preset iteration number is reached. The extended algorithm implementation step is: adding an environment feature input interface based on the original algorithm; expanding the network structure to add an environment feature processing module; adjusting the output layer to add a dynamic control variable generation function.

[0143] The environmental characteristics of the current scene are acquired, and the environmental characteristics are input into the extended risk perception type policy gradient algorithm. Real-time environmental data is acquired by using an airborne sensor, including wind speed, wind direction, visibility, and other meteorological information. The scene type is determined by a scene recognition algorithm, such as an open space, a complex obstacle region, and the like. The number and distribution of the currently detected dynamic obstacles are counted. The distribution of the static obstacles is analyzed in combination with preloaded map data. The environmental characteristic processing steps are as follows: numerical processing is performed on various environmental characteristics; the numerical characteristic values are normalized to unify the ranges of the characteristic values; and the normalized environmental characteristics and obstacle motion characteristics are combined to form a complete input vector.

[0144] The dynamic control variables are generated based on the extended risk perception type policy gradient algorithm. The environmental characteristics and obstacle prediction trajectories are input into the extended algorithm. The output vector is obtained by forward propagation calculation. The output vector is analyzed to obtain the values of the control variables, including the track point spacing, the turning angle limit, the speed range, and the acceleration limit. The generated control variables are checked for legality to ensure that they are within a safe range. If necessary, the control variables that exceed the safe range are corrected. The control variable optimization steps are as follows: the control variables are dynamically adjusted based on the current risk assessment results. In an obstacle-dense region, the track point spacing is reduced, the turning angle limit is relaxed, and the speed range is reduced. In an open region, the track point spacing is increased, the turning angle limit is tightened, and the speed range is increased.

[0145] A rapid expansion random tree is initialized, with the starting point being the current position of the unmanned aerial vehicle and the end point being the target position. Sampling is randomly performed in the state space at a set sampling frequency. For each sampling point, the nearest node in the tree is found. The new node is generated by expanding from the nearest node in the direction of the sampling point. It is checked whether the new node collides with an obstacle. If there is a collision, the new node is discarded. The legal new node is added to the tree. The above process is repeated until the tree contains the end point or the maximum number of iterations is reached. The path extraction steps are as follows: when the tree successfully connects the starting point and the end point, the path from the end point to the starting point is backtracked. The original path is pruned to remove redundant nodes. The path is smoothed using a B-spline curve. The track point sequence is generated by equally sampling the smoothed path according to the set track point spacing. The generated track points are checked for safety to ensure that they meet the constraint conditions of the dynamic control variables.

[0146] The curvature distribution of the sequence of track points is analyzed to identify sections with excessively sharp turns, the sections with excessively sharp turns are locally optimized to make the steering angle change more gently, the distance between adjacent track points is checked to ensure compliance with the track point spacing requirement, the total length of the sequence of track points and the estimated flight time are calculated to evaluate the path efficiency, and if necessary, the sequence of track points is globally optimized to balance safety and efficiency. The track execution step is: transmitting the optimized sequence of track points to the UAV flight control system; the flight control system generates control instructions according to the track points; the UAV executes the control instructions and flies according to the planned track; the system continuously monitors the motion state of dynamic obstacles; when a significant change in the motion trajectory of an obstacle is detected, the trajectory prediction and track planning are re-executed.

[0147] In practical applications, the system workflow is: sensor data is collected every 50 milliseconds to update the obstacle state; trajectory prediction is performed every 200 milliseconds to update the predicted trajectory; the safety of the current track is evaluated every 500 milliseconds, and if necessary, it is re-planned; training samples are extracted from the experience pool every 5 seconds to update the algorithm parameters; experience pool data is saved every 10 minutes to ensure that data is not lost due to system restart.

[0148] In a complex environment containing multiple dynamic obstacles, the obstacle avoidance success rate of this method reaches 95%, which is 15% higher than that of traditional methods; the average obstacle avoidance margin is 8.5 meters, which is 40% higher; the trajectory smoothness is 30% higher than that of traditional methods, reducing the energy consumption of the UAV; the flight efficiency is improved by 25%, shortening the task completion time; the online learning ability of the algorithm continuously improves the prediction accuracy as experience accumulates, and the prediction error is reduced by 45% after long-term operation compared to the initial stage.

[0149] This method is particularly suitable for use in complex environments where dynamic obstacles frequently appear, such as densely populated areas and multi-UAV cooperative work areas. The system can accurately predict the motion trajectory of obstacles and adjust the track planning in a timely manner to ensure flight safety while improving task efficiency. After a large number of experiments, this method has shown excellent performance in various complex scenarios, providing reliable technical support for UAV autonomous flight.

[0150] In an alternative embodiment, the risk-aware policy gradient algorithm is used to automatically optimize the control strategy when an environmental mutation is detected, trajectory tracking control is performed in combination with the environmental prediction result, trajectory tracking control instructions are executed, and the UAV is controlled to fly along the planned trajectory to the optimal landing position.

[0151] The environmental state information at adjacent time instants is calculated to obtain an environmental state difference value, an environmental mutation threshold value is set, and when the environmental state difference value is greater than the environmental mutation threshold value, a risk assessment value is calculated based on the environmental state difference value;

[0152] mapping the risk assessment value into an advantage function by using a risk-aware policy gradient algorithm, determining a risk adjustment factor based on the advantage function, updating parameters of an original control policy according to a product of the advantage function and the risk adjustment factor to obtain new control policy parameters;

[0153] analyzing the environment state information based on the new control policy parameters to obtain an environment prediction result, matching the environment prediction result with an actual position and an actual speed of the unmanned aerial vehicle to obtain an expected trajectory, and calculating a tracking error between the actual position, the actual speed and the expected trajectory;

[0154] compensating the tracking error according to the new control policy parameters to obtain a position control amount and a speed control amount, combining the position control amount and the speed control amount to form a trajectory tracking control instruction, and executing the trajectory tracking control instruction to control the unmanned aerial vehicle to fly along the expected trajectory to an optimal landing position.

[0155] The unmanned aerial vehicle obtains environment state information through a sensor system carried by the unmanned aerial vehicle. The sensor system includes a barometer, an anemometer, a humidity sensor, a GPS positioning module, etc. The sampling frequency of the barometer is 10 Hz, the measurement range is 300-1100 hundred pascals, and the accuracy is ±0.1 hundred pascals; the sampling frequency of the anemometer is 5 Hz, the measurement range is 0-60 meters / second, and the accuracy is ±0.5 meters / second; the sampling frequency of the humidity sensor is 1 Hz, the measurement range is 0-100%, and the accuracy is ±2%; and the positioning accuracy of the GPS positioning module is ±1.5 meters. The sensor data is transmitted to a flight control computing unit through a CAN bus, and the computing unit has a processing frequency of 100 Hz.

[0156] The environment state information at adjacent time instants is calculated to obtain an environment state difference value. Environment data is collected at time instants t and t-1, including wind speeds Vt and Vt-1, air pressures Pt and Pt-1, and humidities Ht and Ht-1. The wind speed difference value AV = |Vt-Vt-1|, the air pressure difference value AP = |Pt-Pt-1|, and the humidity difference value AH = |Ht-Ht-1| are calculated. For example, in a certain flight, the wind speed Vt = 7.5 meters / second is measured at time instant t, and the wind speed Vt-1 = 4.5 meters / second is measured at time instant t-1, so the wind speed difference value AV = 3.0 meters / second; the air pressure Pt = 985 hundred pascals is measured at time instant t, and the air pressure Pt-1 = 987 hundred pascals is measured at time instant t-1, so the air pressure difference value AP = 2.0 hundred pascals; and the humidity Ht = 65% is measured at time instant t, and the humidity Ht-1 = 60% is measured at time instant t-1, so the humidity difference value AH = 5.0%.

[0157] A threshold of environmental mutation is set to determine whether the environment has mutated, the threshold of wind speed mutation is AVth=2.0 m / s, the threshold of air pressure mutation is APth=1.5 hPa, and the threshold of humidity mutation is AHth=4.0%. The system determines whether the difference in environmental state is greater than the threshold of environmental mutation: if AV>AVth or AP>APth or AH>AHth, it is considered that the environment has mutated. In the above example, AV=3.0 m / s>AVth=2.0 m / s, AP=2.0 hPa>APth=1.5 hPa, and AH=5.0%>AHth=4.0%, all three indicators exceed the threshold, and the system determines that the environment has mutated.

[0158] When the environment mutates, a risk assessment value R is calculated based on the difference in environmental state, which is obtained by assigning weights to each difference in environmental state and calculating the weighted sum. In this embodiment, the weight of the wind speed difference is wV=0.5, the weight of the air pressure difference is wP=0.3, and the weight of the humidity difference is wH=0.2. The risk assessment value R is calculated as: R=wV×(AV / AVth)+wP×(AP / APth)+wH×(AH / AHth). In the above example, R=0.5×(3.0 / 2.0)+0.3×(2.0 / 1.5)+0.2×(5.0 / 4.0)=0.5×1.5+0.3×1.33+0.2×1.25=0.75+0.4+0.25=1.4. The range of the risk assessment value R is usually between 0 and 5, and the larger the value, the more serious the degree of environmental mutation.

[0159] The risk assessment value R is mapped to an advantage function A using a risk-aware policy gradient algorithm, which represents the value increase of the current state relative to the average state in this embodiment, and the calculation method is: when R<1, A=R; when 1≤R<3, A=R×R; when R≥3, A=9+(R-3)×0.5. In the above example, R=1.4 is in the range of 1≤R<3, so A=R×R=1.4×1.4=1.96.

[0160] A risk adjustment factor λ is determined based on the advantage function A, which reflects the sensitivity of the system to risk, and the calculation method is: λ=1 / (1+exp(-A+3)). In the above example, A=1.96, λ=1 / (1+exp(-1.96+3))=1 / (1+exp(1.04))=1 / (1+2.83)=1 / 3.83=0.26. The range of the risk adjustment factor λ is between 0 and 1, and the closer the value to 1, the more sensitive the system to risk.

[0161] The parameters of the original control strategy are updated according to the product of the advantage function A and the risk adjustment factor λ to obtain new control strategy parameters. Assume that the original control strategy parameters are θ = [θ1, θ2, θ3, θ4], which respectively represent the pitch control parameters, roll control parameters, throttle control parameters and yaw control parameters of the unmanned aerial vehicle. The original parameter values are θ1= 0.65, θ2= 0.58, θ3= 0.42, and θ4= 0.35. The control strategy parameter update formula is θ' = θ + α × λ × A × ∇J(θ), where α is the learning rate, which is set to 0.05, and ∇J(θ) is the policy gradient, which is set to [0.2, 0.15, 0.25, 0.1] in this example. Calculation gives: θ' = θ + 0.05 × 0.26 × 1.96 × [0.2, 0.15, 0.25, 0.1]. For θ1: θ1' = 0.65 + 0.05 × 0.26 × 1.96 × 0.2 = 0.65 + 0.0051 = 0.6551; similarly, θ2' = 0.58 + 0.0038 = 0.5838, θ3' = 0.42 + 0.0064 = 0.4264, and θ4' = 0.35 + 0.0025 = 0.3525.

[0162] The environment state information is analyzed based on the new control strategy parameters to obtain the environment prediction result. The environment prediction uses a time series analysis method, considers the historical data trend and the new control strategy parameters, and predicts the environment state in the next 5 seconds. In this example, the system predicts that the wind speed will change to [7.8, 8.0, 8.1, 8.0, 7.8] m / s and the wind direction will change to [75, 78, 80, 82, 85] degrees (0 degrees is the north direction, increasing clockwise) and the air pressure will change to [984, 983, 982, 982, 981] hundred pascals in the next 5 seconds.

[0163] The environment prediction result is matched with the actual position and actual speed of the UAV to obtain the expected trajectory. Assuming that the current position of the UAV is longitude 114.05 degrees, latitude 22.54 degrees, height 80 meters, horizontal speed 6 meters / second, direction angle 65 degrees, and vertical speed -1.5 meters / second, and the target landing point position is longitude 114.06 degrees, latitude 22.55 degrees. The system calculates the expected trajectory according to the predicted wind field information, the current state of the UAV, and the target landing point. The expected trajectory includes position and speed information at a series of time points, with an interval of 0.5 seconds, generating 10 trajectory points. The trajectory points include: (114.052, 22.542, 75), (114.054, 22.544, 65), (114.056, 22.546, 50), (114.057, 22.548, 35), (114.058, 22.549, 20), (114.059, 22.5495, 10), (114.06, 22.55, 0). The corresponding expected speed is: (6.2, -1.8), (6.0, -2.0), (5.5, -3.0), (4.5, -3.0), (3.0, -3.0), (2.0, -2.0), (0.5, -1.0).

[0164] The tracking error between the actual position and actual speed of the UAV and the expected trajectory is calculated. Assuming that at a certain time, the actual position of the UAV is (114.053, 22.543, 70), the actual speed is (6.1, -1.5), the corresponding expected position is (114.054, 22.544, 65), and the expected speed is (6.0, -2.0). The position error calculation is: longitude error = 114.054 - 114.053 = 0.001 degrees, latitude error = 22.544 - 22.543 = 0.001 degrees, height error = 65 - 70 = -5 meters. The speed error calculation is: horizontal speed error = 6.0 - 6.1 = -0.1 meters / second, vertical speed error = -2.0 - (-1.5) = -0.5 meters / second.

[0165] The position control amount and the speed control amount are obtained by compensating the tracking error according to the new control strategy parameters. The position control amount is calculated by the product of the position error and the position control gain, and the speed control amount is calculated by the product of the speed error and the speed control gain. The position control gain is Kp=[0.02, 0.02, 0.05], and the speed control gain is Kv=[0.6, 1.2]. The position control amount is calculated as follows: ΔPosCtrl=Kp×[longitude error, latitude error, height error]=[0.02×0.001, 0.02×0.001, 0.05×(-5)]=[0.00002, 0.00002, -0.25]. The speed control amount is calculated as follows: ΔVelCtrl=Kv×[horizontal speed error, vertical speed error]=[0.6×(-0.1), 1.2×(-0.5)]=[-0.06, -0.6].

[0166] The position control amount and the speed control amount are combined to form the trajectory tracking control instruction, and the control instruction includes four dimensions: the pitch control instruction, the roll control instruction, the throttle control instruction and the yaw control instruction. The combination method is as follows: the pitch control instruction=θ1'×(ΔPos y +ΔVel y )=0.6551×(0.00002+(-0.06))=-0.039; the roll control instruction=θ2'×(ΔPos x +ΔVel x )=0.5838×(0.00002+(-0.06))=-0.035; the throttle control instruction=θ3'×(ΔPos z +ΔVel y )=0.4264×(-0.25+(-0.6))=-0.363; and the yaw control instruction=θ4'×0=0; wherein ΔPos x is the X-axis position deviation amount, ΔPos y is the Y-axis position deviation amount, ΔPos z is the Z-axis position deviation amount, ΔVel x is the X-axis speed deviation amount, ΔVel y is the Y-axis speed deviation amount, and ΔVel z is the Z-axis speed deviation amount. The finally generated control instruction is [-0.039, -0.035, -0.363, 0].

[0167] The trajectory tracking control instruction is executed to control the UAV to fly along the desired trajectory to the optimal landing position, and the control instruction is converted into motor speed instruction by the flight control system to drive the UAV to move. The system continuously monitors the environment state and the UAV state at a frequency of 50Hz, calculates the tracking error in real time and updates the control instruction. When the UAV approaches the ground (the height is lower than 5m), the system starts the slow descent mode, controls the vertical speed within-0.5m / s, and ensures smooth landing.

[0168] During the landing process, the system continuously detects the difference of the environment state, and if the environmental mutation is detected again (such as the wind speed suddenly increases during the landing process), the system repeats the above risk assessment, strategy optimization and trajectory tracking control process to ensure the safe landing of the UAV. During the whole landing process, the system controls the position error of the UAV within ±2m and the speed error within ±0.5m / s.

[0169] In actual tests, the method has been tested for landing 100 times under different environmental conditions. In a stable environment (wind speed less than 4m / s), the landing accuracy is ±0.8m; in a slightly disturbed environment (wind speed 4-8m / s), the landing accuracy is ±1.5m; and in a strong disturbed environment (wind speed 8-15m / s), the landing accuracy is ±2.3m. Compared with the traditional PID control method, the landing success rate of the method in the strong disturbed environment is increased by 42%, and the average landing accuracy is increased by 35%. The experimental results show that the method can effectively cope with environmental mutations and ensure the safe landing of the UAV.

[0170] In a second aspect of the embodiment of the application, a UAV intelligent decision system based on deep reinforcement learning is provided, comprising:

[0171] A first unit is configured to acquire image data and depth information of a landing area collected by a vision sensor carried by the UAV, process the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and construct a three-dimensional environment description of the landing area based on the depth information.

[0172] A second unit is configured to process the information representation and the three-dimensional environment description by using a risk perception type strategy gradient algorithm, organize different types of environmental information into a unified structure, adaptively construct the association relationship between the information, selectively transmit and update the environmental information in combination with the association relationship, realize bidirectional propagation and fusion of features, perform landing area segmentation and evaluation based on the finally fused features, and obtain an optimal landing position.

[0173] The third unit is configured to predict a dynamic obstacle movement track in the flight process of the UAV based on the risk-aware policy gradient algorithm, extend the risk-aware policy gradient algorithm based on the prediction result, update algorithm parameters based on historical experience, dynamically adjust control variables according to current scene characteristics to generate a flight path point sequence, automatically optimize a control strategy when detecting an environmental mutation by using the risk-aware policy gradient algorithm, perform trajectory tracking control in combination with an environmental prediction result, execute a trajectory tracking control instruction, and control the UAV to fly along a planned trajectory to the optimal landing position.

[0174] In a third aspect, an electronic device is provided, including:

[0175] a processor;

[0176] a memory for storing processor-executable instructions;

[0177] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0178] In a fourth aspect, a computer-readable storage medium is provided, which stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0179] The present application can be a method, device, system and / or computer program product. The computer program product can include a computer-readable storage medium having stored thereon computer-readable program instructions that, when executed by a computer, cause the computer to carry out various aspects of the present application.

[0180] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent decision of unmanned aerial vehicle based on deep reinforcement learning, characterized in that, The method comprises the following steps: acquiring image data and depth information of a landing area collected by a vision sensor carried by a UAV, processing the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and constructing a three-dimensional environment description of the landing area based on the depth information; processing the information representation and the three-dimensional environment description by using a risk-aware policy gradient algorithm, organizing different types of environment information into a unified structure, adaptively constructing the correlation between information, selectively transmitting and updating environment information in combination with the correlation, realizing bidirectional propagation and fusion of features, performing landing area segmentation and evaluation based on the final fused features to obtain an optimal landing position, comprising: constructing an information transmission network, calculating the correlation strength and information mask of each region in the information transmission network, setting a transmission threshold, obtaining a transmission strength when the correlation strength is greater than the transmission threshold and the information mask value is less than the transmission threshold, calculating a local safety feature of each region according to the transmission strength, determining a global safety metric based on the local safety feature, and generating a region weight by using the global safety metric; constructing a bidirectional propagation channel based on the transmission strength, performing a depth residual operation on the region weight and the transmission strength between adjacent nodes along the forward direction of the bidirectional propagation channel to obtain a forward feature, and performing a depth residual operation on the region weight and the transmission strength between adjacent nodes along the reverse direction of the bidirectional propagation channel to obtain a backward feature; inputting the forward feature and the backward feature into the information transmission network to obtain a fused feature, separating a local feature descriptor and a terrain gradient feature from the fused feature by using the information transmission network, and generating a local detail feature according to the local feature descriptor and the terrain gradient feature; inputting the global safety metric and the local detail feature into the information transmission network to obtain a region safety evaluation value, and selecting a candidate landing position according to the region safety evaluation value to determine an optimal landing position from the candidate landing position; predicting a dynamic obstacle motion trajectory in a flight process of the UAV based on the risk-aware policy gradient algorithm, extending the risk-aware policy gradient algorithm based on a prediction result, updating algorithm parameters based on historical experience and dynamically adjusting a control variable according to a current scene feature to generate a flight path point sequence, automatically optimizing a control strategy by using the risk-aware policy gradient algorithm when an environmental mutation is detected, performing trajectory tracking control in combination with an environmental prediction result, executing a trajectory tracking control instruction, and controlling the UAV to fly along a planned trajectory to the optimal landing position.

2. The method of claim 1, wherein, processing the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and constructing a three-dimensional environment description of the landing area based on the depth information comprises: performing image enhancement processing on the landing area image data based on a dynamic target detection operator, dynamically adjusting a contrast limit threshold by calculating the entropy value of the image gray distribution to obtain enhanced image data; Convolution operations are performed on the enhanced image data by using convolution kernels of different sizes, and feature fusion is performed by the dynamic target detection operator to obtain information representation of the landing area; based on the information representation of the landing area, a dynamic area mask and a static area mask are obtained according to the dynamic target detection operator; The information representation of the landing area is multiplied element by element with the static area mask to obtain a static feature map, and three-dimensional reconstruction is performed on the static feature map in combination with the information representation of the landing area to obtain a three-dimensional environment description of the static area; The information representation of the landing area is multiplied element by element with the dynamic area mask to obtain a dynamic feature map, and three-dimensional reconstruction is performed on the dynamic feature map in combination with the information representation of the landing area to obtain a three-dimensional environment description of the dynamic area; Based on the dynamic target detection operator, adaptive weight coefficients are obtained by processing the information representation of the landing area; and the three-dimensional environment descriptions of the static area and the dynamic area are weighted and fused by using the adaptive weight coefficients to obtain a three-dimensional environment description of the landing area.

3. The method of claim 2, wherein, The image data of the landing area is subjected to image enhancement processing based on a dynamic target detection algorithm, and the entropy value of the image gray scale distribution is calculated to dynamically adjust the contrast limit threshold, to obtain enhanced image data including: The gray scale probability distribution of the landing area image data is calculated, a deep neural network is constructed to calculate the image entropy value of the gray scale probability distribution, a dynamic target detection algorithm is constructed, the dynamic target detection algorithm includes a Gaussian filter response term and a Laplacian operator response term, and the deep neural network is used to perform feature fusion on the Gaussian filter response term and the Laplacian operator response term to obtain a detection algorithm response value; The deep neural network is used to process the detection algorithm response value to obtain a dynamic adjustment amount, and the deep neural network is used to perform feature fusion on the dynamic adjustment amount and a basic threshold value to obtain a contrast limit threshold; Each pixel in the landing area image data is detected, the relationship between the detection algorithm response value and the contrast limit threshold is analyzed by using the deep neural network, the gray scale value of each pixel is enhanced to obtain a response region enhancement value, and the deep neural network is used to perform spatial smoothing constraint processing on the response region enhancement value to obtain enhanced landing area image data.

4. The method of claim 1, wherein, The information representation and the three-dimensional environment description are processed by using a risk-aware policy gradient algorithm, different types of environmental information are organized into a unified structure, and the association relationship between the information is adaptively constructed including: Based on the information representation, the information entropy of the image is calculated, and the feature extraction threshold is set according to the information entropy; when the information entropy is greater than the feature extraction threshold, the feature extraction of the information representation and the three-dimensional environment description is completed; constructing a risk perception type policy gradient algorithm, calculating a matrix eigenvalue and a bias eigenvalue of the risk perception type policy gradient algorithm, when the matrix eigenvalue is greater than the feature extraction threshold and the bias eigenvalue is less than the feature extraction threshold, iteratively optimizing the result of the feature extraction by using the risk perception type policy gradient algorithm to obtain an environmental feature vector; structurally processing the environmental feature vector to obtain environmental information with a unified structure, and constructing an association relationship between information based on the environmental information with the unified structure.

5. The method of claim 1, wherein, Based on the risk perception type policy gradient algorithm, the dynamic obstacle motion trajectory in the flight process of the unmanned aerial vehicle is predicted, the risk perception type policy gradient algorithm is extended based on the prediction result, the algorithm parameters are updated based on historical experience, and the control variable is dynamically adjusted according to the current scene characteristics to generate a flight path point sequence, including: Collecting motion data of dynamic obstacles in the flight process of the unmanned aerial vehicle, calculating the position coordinates, velocity vectors, acceleration vectors and motion directions of the dynamic obstacles, and inputting the position coordinates, velocity vectors, acceleration vectors and motion directions into the risk perception type policy gradient algorithm to obtain the predicted trajectory of the dynamic obstacles; Based on the predicted trajectory, an experience pool is constructed, the historical motion trajectory of the dynamic obstacle is matched with the predicted trajectory, the matching error is calculated, the immediate reward value is determined according to the matching error, and the historical motion trajectory, the predicted trajectory and the immediate reward value are stored in the experience pool; Setting an experience pool capacity threshold, when the amount of data in the experience pool is greater than the experience pool capacity threshold, extracting training samples from the experience pool, updating the parameters of the risk perception type policy gradient algorithm based on the training samples, and obtaining an extended risk perception type policy gradient algorithm; Obtaining the environmental characteristics of the current scene, inputting the environmental characteristics into the extended risk perception type policy gradient algorithm, generating a dynamic control variable based on the extended risk perception type policy gradient algorithm, and generating a flight path point sequence according to the dynamic control variable.

6. The method of claim 1, wherein, Using the risk perception type policy gradient algorithm to automatically optimize the control strategy when detecting environmental mutations, combining the environmental prediction results to perform trajectory tracking control, executing trajectory tracking control instructions, and controlling the unmanned aerial vehicle to fly along the planned trajectory to the optimal landing position, including: Calculating the environmental state information of adjacent time to obtain an environmental state difference value, setting an environmental mutation threshold, when the environmental state difference value is greater than the environmental mutation threshold, calculating a risk assessment value based on the environmental state difference value; Using the risk perception type policy gradient algorithm to map the risk assessment value into an advantage function, determining a risk adjustment factor based on the advantage function, and updating the parameters of the original control strategy according to the product of the advantage function and the risk adjustment factor to obtain new control strategy parameters; analyzing the environment state information based on the new control strategy parameter to obtain an environment prediction result, matching the environment prediction result with an actual position and an actual speed of the UAV to obtain an expected trajectory, and calculating a tracking error between the actual position, the actual speed and the expected trajectory; compensating the tracking error based on the new control strategy parameter to obtain a position control amount and a speed control amount, combining the position control amount and the speed control amount to form a trajectory tracking control instruction, and executing the trajectory tracking control instruction to control the UAV to fly along the expected trajectory to an optimal landing position.

7. A drone intelligent decision system based on deep reinforcement learning for implementing the method of any of the preceding claims 1-6, characterized in that, The method comprises the following steps: a first unit is configured to acquire image data and depth information of a landing area collected by a vision sensor carried by a UAV, process the image data of the landing area by using a basic feature extraction method to obtain information representation of the landing area, and construct a three-dimensional environment description of the landing area based on the depth information; a second unit is configured to process the information representation and the three-dimensional environment description by using a risk perception type policy gradient algorithm, organize different types of environment information into a unified structure, adaptively construct a correlation relationship between the information, selectively transmit and update the environment information in combination with the correlation relationship, realize bidirectional propagation and fusion of features, perform landing area segmentation and evaluation based on finally fused features, and obtain an optimal landing position; a third unit is configured to predict a dynamic obstacle motion trajectory in a flight process of the UAV based on the risk perception type policy gradient algorithm, extend the risk perception type policy gradient algorithm based on a prediction result, update algorithm parameters based on historical experience and dynamically adjust a control variable according to a current scene feature to generate a flight path point sequence, automatically optimize a control strategy by using the risk perception type policy gradient algorithm when an environment mutation is detected, perform trajectory tracking control in combination with an environment prediction result, execute a trajectory tracking control instruction, and control the UAV to fly along a planned trajectory to the optimal landing position.

8. An electronic device, comprising: The method comprises the following steps: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions are executed by the processor to implement the method of any one of claims 1 to 6.

Images