Unmanned aerial vehicle surveying and mapping method and system based on deep learning

Abstract

The invention relates to the technical field of unmanned aerial vehicle intelligent surveying and mapping, in particular to an unmanned aerial vehicle surveying and mapping method and system based on deep learning, and the method comprises the following steps: S1, collecting the obstacle information in the surrounding environment in real time, including the position, shape and motion state, and generating an environment data set; s2, classifying obstacles, and distinguishing static obstacles from dynamic obstacles; for a dynamic obstacle, using a recurrent neural network to predict a moving track of the dynamic obstacle; s3, updating the route plan of the unmanned aerial vehicle in real time; s4, adjusting the speed of the unmanned aerial vehicle according to the obstacle density and the dynamic obstacle frequency in the current area; and S5, when the route conflicts with the predicted path of the dynamic obstacle, adjusting the flight height or temporarily staying. According to the invention, accurate recognition and dynamic prediction of the obstacle are realized through deep learning, and the route planning, speed and data acquisition frequency of the unmanned aerial vehicle are optimized in real time, so that the safety and efficiency of surveying and mapping are significantly improved in a complex environment.

Description

Technical Field

[0001] The present invention relates to the technical field of UAV intelligent mapping, and particularly to a UAV mapping method and system based on deep learning. Background Art

[0002] With the development of UAV technology and deep learning algorithms, UAVs have been widely used in fields such as geographic information collection, environmental monitoring, and infrastructure inspection. UAV mapping has gradually replaced traditional manual or ground equipment mapping methods due to its flexibility and efficiency. However, in complex environments, such as areas with dense urban buildings, forests, and other areas with variable obstacles, UAV mapping faces great challenges, especially in identifying and avoiding dynamic obstacles. Current mapping systems mostly rely on GPS and simple path planning algorithms, lacking the ability to dynamically adapt to the environment, resulting in insufficient safety and mapping accuracy in complex environments.

[0003] There are several key problems in the existing technology of UAV mapping that need to be solved urgently: First, how to identify, classify, and predict the movement paths of obstacles in an environment with variable obstacles in real time to avoid potential collision risks. Second, how to reasonably adjust the flight speed and acquisition frequency in areas with unevenly distributed obstacles to improve mapping efficiency. Third, current UAVs are difficult to quickly adjust the flight path according to the real-time state of obstacles to ensure flight safety. Therefore, proposing a UAV mapping method and system based on deep learning has become an important technical requirement for improving mapping accuracy and safety. Summary of the Invention

[0004] Based on the above objectives, the present invention provides a UAV mapping method and system based on deep learning.

[0005] A UAV mapping method based on deep learning includes the following steps:

[0006] S1: During the flight of the UAV, in real time collect obstacle information in the surrounding environment, including position, shape, and motion state, and perform denoising and format standardization processing on the collected data to generate an environmental data set;

[0007] S2: Based on the data set generated in S1, classify the obstacles to distinguish static obstacles and dynamic obstacles; for dynamic obstacles, use a recurrent neural network to predict their movement trajectories to generate future movement paths and speed information of the obstacles;

[0008] S3: Based on the obstacle classification and prediction data in S2, update the flight path planning of the UAV in real time; for static obstacles, determine a safe flight area based on their positions; for dynamic obstacles, calculate the avoidance time and safe distance according to their predicted trajectories and generate a new flight path;

[0009] S4: Divide the flight area into several small areas, and adjust the UAV speed according to the obstacle density and dynamic obstacle frequency in the current area; when in an area with dense obstacles, reduce the flight speed and increase the data collection frequency; when in an area with sparse obstacles, perform the mapping task at the fastest speed;

[0010] S5: When the flight path conflicts with the predicted path of a dynamic obstacle, according to the obstacle prediction result of S2, choose to adjust the flight altitude or stay temporarily to ensure the UAV passes through the conflict area safely.

[0011] Optionally, the specific steps of S1 include:

[0012] S11: During the flight of the UAV, obtain the distance and three-dimensional spatial distribution information of obstacles in the environment through the configured lidar; obtain the two-dimensional image data and shape feature information of obstacles through the configured camera; obtain the real-time position information of obstacles through the ultrasonic sensor and confirm the relative distance between the obstacles and the UAV; to establish complete obstacle information, including the position, three-dimensional shape contour and its motion state of the obstacles;

[0013] S12: Perform fusion processing on the obstacle data of the lidar, camera and ultrasonic sensor obtained in step S11, match the spatial data of the lidar, the image data of the camera and the position information of the ultrasonic wave to ensure that the position information, shape features and motion states of each obstacle correspond to each other;

[0014] S13: Denoise the data fused in S12. Apply the Kalman filter algorithm to filter the distance data of the lidar and ultrasonic wave to eliminate the existing environmental noise; apply the Gaussian filter to the camera image data to remove image noise points;

[0015] S14: Convert the data after denoising in S13 into a unified format, specifically including converting the distance and position coordinates into a unified three-dimensional coordinate system, and formatting the image data into the size and resolution that meet the input requirements of the deep learning model to generate a standardized environmental data set.

[0016] Optionally, the specific steps of S13 include:

[0017] S131: Perform initial settings on the distance data collected by the lidar and ultrasonic sensor, define the state vector to represent the distance and distance change rate of the obstacle, and set the error covariance matrix to estimate the degree of measurement error;

[0018] S132: At the current moment, predict the state vector and the error covariance matrix. Describe the evolution of distance information over time through the state transition matrix, while considering the influence of control inputs on the state and the noise in the system process to obtain the predicted state at the next moment;

[0019] S133: Update the state using the measurement data from the lidar and ultrasonic sensors. Determine the balance between the actual measurement and the predicted state by calculating the Kalman gain, and correct the state vector and the error covariance matrix to obtain the denoised obstacle distance and speed data;

[0020] S134: Apply a Gaussian filter to the image data collected by the camera. The Gaussian filter is calculated at each pixel point of the image, and the influence of surrounding pixels is integrated through predetermined weights to remove the noise in the image.

[0021] Optionally, the specific steps of S2 include:

[0022] S21: Based on the standardized environmental dataset generated by S1, input the data into a deep learning model for feature extraction, including the spatial features, shape features, and relative motion features of the obstacles, to provide accurate obstacle attribute information for subsequent classification;

[0023] S22: Analyze the extracted obstacle features using a pre-trained convolutional neural network model, and classify the obstacles into two categories: static obstacles and dynamic obstacles. Specifically, by analyzing the position and shape changes of the obstacles in consecutive data frames, identify the obstacles with no obvious moving trend as static obstacles, and the obstacles with a continuous moving trend as dynamic obstacles;

[0024] S23: For the obstacles classified as dynamic, input their feature data into a recurrent neural network, and generate their future motion paths and speed information by analyzing the historical motion states of the obstacles to estimate the positions and directions of the obstacles in the next time period;

[0025] S24: Output the trajectory prediction results of S23 as the future motion paths and speed information of the dynamic obstacles, and transfer the prediction results to step S3 for real-time route planning adjustment.

[0026] Optionally, the specific steps of S22 include:

[0027] S221: Initialize the pre-trained convolutional neural network model, and configure the network parameters to adapt to the input of the obstacle feature data. The input feature data includes the obstacle images collected in different time frames and their position coordinate information;

[0028] S222: Input the obstacle image data in consecutive frames into a convolutional neural network model, extract the image features of the obstacle through the convolutional layer to form a feature map; the feature map of each time frame contains the position and shape information of the obstacle; establish the feature sequence of the same obstacle in consecutive frames by matching with the obstacle features of adjacent frames;

[0029] S223: Analyze the position change of the feature sequence matched in S222. Let the current frame position coordinates be (x k , y k ), and the previous frame position coordinates be (x k-1 , y k-1 ), then the displacement amount Δd k of the obstacle between these two frames is calculated as follows: Compare Δd k with the set displacement threshold T d . If Δd k is less than or equal to T d , then the displacement of the obstacle in the current frame is not sufficient to be judged as moving, and it is marked as a static obstacle candidate;

[0030] S224: Analyze the shape change of the obstacle in consecutive frames. Let the shape features of the obstacle in the current frame and the previous frame be the area A k and A k-1 , and the shape change rate ΔA k is calculated as: Compare ΔA k with the shape change threshold T A . If ΔA k is less than or equal to T A , then the shape change of the obstacle is not obvious, and it continues to be marked as a static obstacle candidate;

[0031] S225: Make a comprehensive judgment on the position change Δd and shape change rate ΔA of the candidate obstacle in multiple consecutive frames; if the displacement amount and shape change rate of the candidate obstacle in several consecutive frames are both less than their respective thresholds T d and T A , then the obstacle is identified as a static obstacle; otherwise, it is identified as a dynamic obstacle.

[0032] Optionally, the specific steps of S23 include:

[0033] S231: Initialize the recurrent neural network model and set the input of the model to the position coordinates, speed, and motion direction angle of the dynamic obstacle in consecutive time frames;

[0034] S232: Normalize the position and speed data of the obstacle in historical time frames to reduce the impact of data differences on model training;

[0035] S233: Gradually input the processed historical trajectory data into the hidden layer of the recurrent neural network; for the input of each time frame, the recurrent neural network model will update its state based on the previous state and the current input, gradually establishing the motion pattern of the obstacle to identify the movement rule of the obstacle;

[0036] S234: According to the final state of the recurrent neural network model, output the predicted position and speed information of the dynamic obstacle, generating the future position coordinates and speed data;

[0037] S235: Based on the predicted future position and speed data, calculate the motion direction angle of the obstacle to determine its future traveling direction, and after combining the predicted direction with the speed, form the complete future motion path information.

[0038] Optionally, the specific steps of S3 are as follows:

[0039] S31: Based on the obstacle classification data of S2, for static obstacles, establish the safe flight area of the static obstacle by obtaining the accurate position of the obstacle; assume the boundary of the obstacle is circular or polygonal, and define the boundary distance of the safe flight area as the safety distance R extended outward from the obstacle boundary s , and the calculation formula is: R s =R o +D margin , where R o is the radius of the obstacle or the distance from the outermost edge to the center, and D margin is the set safety margin;

[0040] S32: For the obstacles classified as dynamic, calculate the time and safety distance required for the UAV to avoid the obstacle according to its predicted trajectory and speed information; assume the current speed of the UAV is v u , the predicted speed of the obstacle is v d , and their relative speed is v r ; then the calculation formulas for the avoidance time and safety distance are: D a =v u ×T a +D margin ; where D is the distance between the current UAV and the obstacle, D margin is the additional safety margin during avoidance, T a represents the shortest time required for avoidance, and D a represents the minimum safety distance that the UAV should maintain from the obstacle during the avoidance process;

[0041] S33: Based on the safety area and avoidance parameters calculated in S31 and S32, generate the flight path of the UAV in real time; for static obstacles, adjust the UAV flight path by flying around the outside of the safety area boundary; for dynamic obstacles, based on the avoidance time and safety distance, update the flight direction and speed of the UAV in real time to avoid in front of the movement direction of the obstacle and ensure a safe distance.

[0042] Optionally, the specific content of S4 includes:

[0043] S41: Divide the flight area within the UAV mission into several small areas; specifically, with the UAV flight path as the center, generate multiple rectangular or polygonal grids at a fixed distance, and each grid area is called a small area;

[0044] S42: In each small area, calculate the current obstacle density and the occurrence frequency of dynamic obstacles; assume the area of the small area is A, and the total number of obstacles is N, then the obstacle density D is calculated as: The dynamic obstacle frequency F d By calculating the number of dynamic obstacles N entering the small area within a fixed time window d , and calculated according to the window duration T, the formula is:

[0045] S43: According to the obstacle density D and dynamic obstacle frequency F of each small area d , adjust the flight speed and data acquisition frequency of the UAV in real time; the specific adjustment criteria include:

[0046] Criterion 1, when D exceeds the set dense threshold D thresh or F d exceeds the dynamic obstacle frequency threshold F thresh , reduce the flight speed of the UAV to the safe speed and increase the data acquisition frequency to the set maximum frequency to ensure more environmental information is obtained in the obstacle-dense area;

[0047] Criterion 2, when D is lower than the set sparse threshold D low and F d is lower than the dynamic obstacle frequency threshold F low , adjust the flight speed of the UAV to the maximum speed v max to perform the mapping task to speed up the operation speed in the low-risk area.

[0048] Optionally, the specific content of S5 includes:

[0049] S51: When the current flight path of the UAV conflicts with the predicted path of a dynamic obstacle, calculate the relative velocity and relative position between the UAV and the obstacle to determine the time point of approaching the obstacle on the path and the severity of the potential conflict.

[0050] S52: Based on the conflict assessment result in S51, the UAV selects an avoidance strategy according to the speed and path situation of the obstacle. The specific strategies include temporary stay and flight altitude adjustment. If the obstacle has a high speed and quickly passes on the flight path, it is preferred to choose a temporary stay. If the obstacle has a low speed or stays in the flight path area for a long time, choose to adjust the flight altitude to avoid the conflict area vertically.

[0051] S53: When choosing altitude adjustment, set the new flight altitude of the UAV to h new ; The calculation formula is: h new = h u + Δh, where h u is the current flight altitude; Δh is the safety altitude offset.

[0052] S54: When choosing a temporary stay, the UAV determines the waiting time according to the difference between the approaching time and the passing time of the obstacle. The formula is: T w = T d - T c , where T w is the waiting time, T d is the passing time of the obstacle, and T c is the approaching time. If T w > 0, the UAV will hover within the safe range from the conflict area and continue to fly along the original flight path after the obstacle has passed.

[0053] A UAV mapping system based on deep learning for implementing the above-mentioned UAV mapping method based on deep learning, including the following modules:

[0054] Environmental data acquisition module: including lidar, camera and ultrasonic sensor, used to collect obstacle information in the surrounding environment of the UAV in real time, including the position, shape and motion state of the obstacle, and transmit the collected data to the data preprocessing module;

[0055] Data preprocessing module: used to denoise and standardize the format of the data transmitted by the environmental data acquisition module to generate a standardized environmental data set;

[0056] Obstacle Analysis and Classification Module: Based on the environmental data set generated by the data preprocessing module, use a pre-trained convolutional neural network to extract features and classify obstacles, distinguishing between static and dynamic obstacles; for obstacles classified as dynamic, use a recurrent neural network to analyze their historical trajectories and predict their future movement paths and speed information;

[0057] Route Planning Module: Used to update the UAV's route plan in real time according to the obstacle classification results provided by the obstacle analysis and classification module and the movement prediction data of dynamic obstacles; for static obstacles, the route planning module sets a safe flight area based on the obstacle position; for dynamic obstacles, calculate the avoidance time and safe distance and generate a new flight path;

[0058] Flight Control and Speed Adjustment Module: Receive the flight path and avoidance plan generated by the route planning module, and combine the obstacle density and dynamic obstacle frequency in the current flight area of the UAV to adjust the flight speed and data acquisition frequency of the UAV in real time; specifically, when in an area with dense obstacles, reduce the flight speed and increase the data acquisition frequency; when in an area with sparse obstacles, increase the flight speed to improve the mapping efficiency;

[0059] Conflict Detection and Avoidance Module: Used to detect potential conflicts on the route in real time. When the UAV route intersects with the predicted path of a dynamic obstacle, select an appropriate avoidance strategy; including choosing to adjust the flight altitude or stay temporarily when the conflict is severe to ensure the UAV can safely pass through the conflict area.

[0060] Advantages of the present invention:

[0061] In the present invention, by combining deep learning technologies, real-time recognition, classification, and dynamic prediction of obstacles in complex environments are achieved; by using a convolutional neural network to extract and classify obstacle features, the system can accurately distinguish between static and dynamic obstacles, and use a recurrent neural network to analyze the historical trajectories of dynamic obstacles, thereby accurately predicting their future movement paths and speeds. This technology improves the path planning accuracy of UAVs in environments with changing obstacles, enables UAVs to make more flexible real-time route adjustments, effectively avoids potential collision risks, and ensures flight safety.

[0062] In the present invention, by analyzing the obstacle density and dynamic obstacle frequency, the flight speed and data acquisition frequency of the UAV are dynamically adjusted. The speed is reduced and the data acquisition frequency is increased in areas with dense obstacles, and the speed is increased in sparse areas to accelerate the mapping process, thereby significantly improving the mapping efficiency. Through this optimized regulation of speed and frequency, the UAV can achieve efficient and safe mapping tasks in complex environments and meet the needs of diverse application scenarios. Description of the Drawings

[0063] To more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the drawings in the following description are only those of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0064] Figure 1 Schematic diagram of the UAV mapping method according to an embodiment of the present invention;

[0065] Figure 2 Schematic diagram of the UAV mapping system according to an embodiment of the present invention. Detailed implementation manners

[0066] The present invention will be described in detail below with reference to the drawings and specific embodiments. At the same time, it should be noted here that in order to make the embodiments more detailed, the following embodiments are the best and preferred embodiments. For some well-known technologies, those skilled in the art can also adopt other alternative methods for implementation; moreover, the drawings are only for more specifically describing the embodiments, and are not intended to specifically limit the present invention.

[0067] It should be noted that in the specification, when referring to "an embodiment", "embodiment", "exemplary embodiment", "some embodiments", etc., it indicates that the described embodiment may include specific features, structures or characteristics, but not necessarily every embodiment includes such specific features, structures or characteristics. In addition, when combining an embodiment to describe a specific feature, structure or characteristic, implementing such a feature, structure or characteristic in combination with other embodiments (whether explicitly described or not) should be within the knowledge of those skilled in the relevant art.

[0068] Generally, terms can be understood at least in part from their use in the context. For example, at least in part depending on the context, the term "one or more" used herein can be used to describe any feature, structure or characteristic in a singular sense, or can be used to describe a combination of features, structures or characteristics in a plural sense. In addition, the term "based on" can be understood as not necessarily intended to convey a set of exclusive factors, but instead, at least in part depending on the context, allowing for the existence of other factors that may not be explicitly described.

[0069] As Figure 1 shown, a UAV mapping method based on deep learning includes the following steps:

[0070] S1: During the flight of the UAV, the obstacle information in the surrounding environment is collected in real time, including the position, shape and motion state, and the collected data is denoised and format standardized to generate an environmental data set;

[0071] S2: Based on the dataset generated in S1, classify the obstacles to distinguish static obstacles and dynamic obstacles; for dynamic obstacles, use a Recurrent Neural Network (RNN) to predict their movement trajectories to generate the future movement path and speed information of the obstacles;

[0072] S3: Based on the obstacle classification and prediction data in S2, update the flight route planning of the drone in real-time; for static obstacles, determine the safe flight area based on their positions; for dynamic obstacles, calculate the avoidance time and safe distance according to their predicted trajectories and generate a new flight path, and this flight route planning is continuously adjusted according to the prediction data during the flight;

[0073] S4: Divide the flight area into several small areas and adjust the drone speed according to the obstacle density and dynamic obstacle frequency in the current area; when in an area with dense obstacles, reduce the flight speed and increase the data collection frequency; when in an area with sparse obstacles, perform the mapping task at the fastest speed to improve the mapping efficiency;

[0074] S5: When the flight route conflicts with the predicted path of a dynamic obstacle, according to the obstacle prediction result in S2, choose to adjust the flight altitude or stay temporarily to ensure the drone passes through the conflict area safely.

[0075] S1 specifically includes:

[0076] S11: During the flight of the drone, obtain the distance and three-dimensional spatial distribution information of the obstacles in the environment through the configured lidar; obtain the two-dimensional image data and shape feature information of the obstacles through the configured camera; obtain the real-time position information of the obstacles through the ultrasonic sensor and confirm the relative distance between the obstacles and the drone; to establish complete obstacle information, including the position, three-dimensional shape contour and its motion state of the obstacles;

[0077] S12: Perform fusion processing on the obstacle data of the lidar, camera and ultrasonic sensor obtained in step S11, match the spatial data of the lidar, the image data of the camera and the position information of the ultrasonic wave to ensure that the position information, shape features and motion states of each obstacle correspond;

[0078] S13: Denoise the data fused in S12. Apply the Kalman filter algorithm to filter the distance data of the lidar and ultrasonic wave to eliminate the existing environmental noise; apply the Gaussian filter to the camera image data to remove image noise points to ensure the accuracy and clarity of the obstacle information data;

[0079] S14: Convert the data after denoising in S13 into a unified format, specifically including converting distances and position coordinates into a unified three-dimensional coordinate system, and formatting the image data into a size and resolution that meet the input requirements of the deep learning model to generate a standardized environmental dataset; Through the above steps, the environmental dataset contains complete, clear, and uniformly formatted obstacle information, which helps to improve the accuracy and reliability of subsequent obstacle classification and path planning, thereby enhancing the mapping effect of the drone in complex environments.

[0080] S13 specifically includes:

[0081] S131: Perform initial settings on the distance data collected by the lidar and ultrasonic sensors, define a state vector to represent the distance and rate of change of the distance of the obstacle, and set an error covariance matrix to estimate the degree of measurement error;

[0082] S132: At the current moment, predict the state vector and the error covariance matrix, describe the evolution of the distance information over time through the state transition matrix, and at the same time consider the influence of the control input on the state and the noise in the system process to obtain the predicted state at the next moment;

[0083] S133: Update the state using the measurement data of the lidar and ultrasonic waves, determine the balance between the actual measurement and the predicted state by calculating the Kalman gain, and correct the state vector and the error covariance matrix to obtain the denoised obstacle distance and speed data;

[0084] S134: Apply a Gaussian filter to the image data collected by the camera for processing. The Gaussian filter calculates at each pixel point of the image, synthesizes the influence of the surrounding pixels through a predetermined weight, removes the noise points in the image, and makes the image data clearer.

[0085] The specific calculation steps for denoising are as follows:

[0086] First, perform filtering on the distance data of the lidar and ultrasonic sensors, and set the initial state vector x0 and the initial error covariance matrix P0 of the Kalman filter. Among them, the state vector is representing the distance d of the obstacle k and the rate of change of the distance The error covariance matrix P k represents the estimated degree of measurement error, and the initial value is Among them, σ d is the standard deviation of the distance, is the standard deviation of the speed;

[0087] Then, at time k, for the state vector x k and the error covariance matrix P kMake a prediction and calculate the predicted state vector and the predicted error covariance matrix The formula is: where A is the state transition matrix, which is used to describe the evolution of the state vector over time; B is the control input matrix, which describes the influence of the control input on the state; u k is the control input vector, representing the external control influence on the system; Q is the process noise covariance matrix, representing the noise in the system process;

[0088] Next, use the measurement data z of the lidar and ultrasonic wave k to perform state update, calculate the Kalman gain K k and update the state vector and error covariance matrix. The formula is:

[0089]

[0090] where H is the measurement matrix, which transforms the state space to the measurement space; R is the measurement noise covariance matrix, representing the noise in the measurement process; I is the identity matrix, which is used to maintain the consistency of the matrix dimensions;

[0091] Finally, perform Gaussian filtering on the image data collected by the camera to remove noise; specifically, apply a Gaussian kernel at each pixel point (x, y). The specific formula is as follows: where G(x, y) is the value of the Gaussian kernel function, representing the filtering weight; σ is the standard deviation, which is used to control the smoothing intensity of the Gaussian filter; subsequently, perform convolution calculation on each pixel point I(i, j) on the image through the Gaussian kernel G(x, y) to calculate the filtered image pixel value I′(i, j). The formula is: where I(i, j) is the pixel value of the original image at the position (i, j); I′(i, j) is the pixel value of the filtered image at the position (i, j); k represents the radius of the filtering window, which is (N is the size of the Gaussian filter template); the above steps use Kalman filtering to achieve noise reduction processing of lidar and ultrasonic wave data, improving the accuracy of obstacle distance data; at the same time, processing image noise through a Gaussian filter improves the clarity of image data, thereby providing more reliable support for subsequent obstacle classification and recognition, and providing a high-quality data basis for the route planning of the drone.

[0092] S2 specifically includes:

[0093] S21: Based on the standardized environment dataset generated by S1, input the data into a deep learning model for feature extraction, including the spatial features, shape features, and relative motion features of obstacles, providing accurate obstacle attribute information for subsequent classification;

[0094] S22: Analyze the extracted obstacle features using a pre-trained convolutional neural network (CNN) model, and classify the obstacles into two categories: static obstacles and dynamic obstacles. Specifically, by analyzing the position and shape changes of the obstacles in consecutive data frames, identify the obstacles with no obvious movement trend as static obstacles, and the obstacles with a continuous movement trend as dynamic obstacles.

[0095] S23: For the obstacles classified as dynamic, input their feature data into a recurrent neural network (RNN), and use the time series processing ability of the RNN to predict the movement trajectory of the dynamic obstacles. By analyzing the historical movement states of the obstacles, generate their future movement paths and speed information to estimate the position and direction of the obstacles in the next time period.

[0096] S24: Output the trajectory prediction result of S23 as the future movement path and speed information of the dynamic obstacles, and transfer the prediction result to step S3 for real-time flight path planning adjustment to ensure that the UAV can timely adjust its flight path according to the movement states of the dynamic obstacles. Through the above steps, the accurate classification of obstacles is achieved using a deep learning model, distinguishing static and dynamic obstacles, and predicting their future movement paths and speeds based on the historical trajectories of the dynamic obstacles, providing a reliable basis for the real-time flight path planning of the UAV and improving the mapping accuracy and safety of the UAV in complex environments.

[0097] S22 specifically includes:

[0098] S221: Initialize the pre-trained convolutional neural network (CNN) model, and configure the network parameters to adapt to the input of the obstacle feature data. The input feature data includes the obstacle images and their position coordinate information collected in different time frames for analyzing the continuous changes of the obstacles.

[0099] S222: Input the obstacle image data in consecutive frames into the convolutional neural network model, extract the image features of the obstacles through the convolutional layer to form a feature map. Each feature map of each time frame contains the position and shape information of the obstacles. Establish a feature sequence of the same obstacle in consecutive frames by matching the obstacle features with adjacent frames.

[0100] S223: Analyze the position changes of the feature sequence matched in S222. Let the position coordinates of the current frame be (x k , y k ), and the position coordinates of the previous frame be (x k-1 , y k-1 ). Then the displacement amount Δd k of the obstacle between these two frames is calculated by the formula: Take Δdk Compare with the set displacement threshold T d If Δd k is less than or equal to T d at this time, the displacement of the obstacle in the current frame is not sufficient to be judged as moving, and it is marked as a static obstacle candidate;

[0101] S224: Analyze the shape change of the obstacle within consecutive frames. Let the shape features of the obstacle in the current frame and the previous frame be the area A k and A k-1 , and the shape change rate ΔA k is calculated as: Compare ΔA k with the shape change threshold T A If ΔA k is less than or equal to T A , the shape change of the obstacle is not obvious, and it continues to be marked as a static obstacle candidate;

[0102] S225: Make a comprehensive judgment on the position change Δd and shape change rate ΔA of the candidate obstacle in multiple consecutive frames; if the displacement amount and shape change rate of the candidate obstacle within several consecutive frames are both less than their respective thresholds T d and T A , then the obstacle is identified as a static obstacle; otherwise, it is identified as a dynamic obstacle; Through the above steps, the convolutional neural network is used to analyze the position and shape changes of the obstacle in consecutive frames, realizing the accurate classification of static and dynamic obstacles, providing reliable data support for subsequent path prediction and route adjustment, and ensuring the flight safety and path optimization of the drone during the mapping process.

[0103] S23 specifically includes:

[0104] S231: Initialize the recurrent neural network (RNN) model, and set the input of the model to the position coordinates, speed, and motion direction angle of the dynamic obstacle in consecutive time frames;

[0105] S232: Normalize the position and speed data of the obstacle in historical time frames to reduce the impact of data differences on model training;

[0106] S233: Gradually input the processed historical trajectory data into the hidden layer of the recurrent neural network; for the input of each time frame, the recurrent neural network model will update its state according to the previous state and the current input, gradually establishing the motion pattern of the obstacle to identify the moving rule of the obstacle;

[0107] S234: According to the final state of the recurrent neural network model, output the predicted position and speed information of the dynamic obstacle, generating the future position coordinates and speed data;

[0108] S235: Based on the predicted future position and velocity data, calculate the moving direction angle of the obstacle to determine its future traveling direction. After combining the predicted direction with the velocity, form the complete future motion path information to help the drone judge the moving trend of the dynamic obstacle.

[0109] The calculation steps for generating its future motion path and velocity information using a recurrent neural network model are as follows:

[0110] Recurrent neural network initialization: Initialize the recurrent neural network (RNN), and set the network input as the position coordinates and velocity information of the dynamic obstacle in consecutive time frames; Let the input vector for each time frame be [x t , y t , v t , θ t , where x t and y t represent the position coordinates of the obstacle at time frame t, v t represents the velocity, and θ t represents the moving direction angle;

[0111] Input data preprocessing: Normalize the position and velocity data in the historical time frames to reduce the impact of data differences on model training; Assume the maximum values of the position coordinates are X max and Y max, , and the maximum value of the velocity is V max , then the normalized input vector is calculated as follows: This normalized input vector is used to train the model to predict the moving trend of the dynamic obstacle;

[0112] Historical trajectory input and state update: Input the normalized historical trajectory into the hidden layer of the RNN, and gradually transmit the data of each time frame; For time frame t, the state vector h t is updated according to the state h t-1 of the previous time frame and the current input vector, and the formula is: h t = f(W h · h t-1 + W x · x t + b h ), where h t is the hidden state of the current time frame; W h and W x represent the network weight matrices; b h is the bias term; f is the activation function used to generate the output of the hidden layer;

[0113] Future Motion Path and Speed Prediction: Based on the final state vector h of the RNN t+n , the predicted positions and speeds of dynamic obstacles are output, and the predicted future position coordinates (x t+n , y t+n ) and speed v t+n are used as outputs. The specific generation formulas are: x t+n = g(W o · h t+n + b o ); y t+n = g(W o · h t+n + b o ); v t+n = g(W v · h t+n + b v ); where W o and W v are the weight matrices of the output layer; b o and b v are the biases of the output layer; g is the output activation function used to predict the motion positions and speeds in future time frames;

[0114] Future Motion Direction Estimation: Using the predicted speeds and position coordinates obtained above, calculate the motion direction angle θ t+n in the future time frame; assuming the predicted position is (x t+n 'y t+n ) and the current position is (x t , y t ), the calculation formula for the direction angle θ t+n is: And combine this predicted direction angle with the speed to construct the future motion path information of the obstacle; through the above steps, the historical motion states of dynamic obstacles are analyzed and predicted using the recurrent neural network model to generate the motion path, speed, and direction in the future time period, providing accurate real-time basis for the route adjustment and avoidance decision-making of the UAV, thereby effectively improving the safety and mapping efficiency of the UAV in the dense obstacle environment.

[0115] S3 specifically includes:

[0116] S31: Based on the obstacle classification data of S2, for static obstacles, by obtaining the accurate positions of the obstacles, establish the safe flight areas of static obstacles; assume the boundaries of the obstacles are circular or polygonal, and define the boundary distance of the safe flight area as the safety distance R s extending outward from the obstacle boundary; this distance meets the minimum safety interval requirements between the UAV flight and static obstacles, and the calculation formula is: R s = R o + D margin, where R o is the radius of the obstacle or the distance from the outermost edge to the center, and D margin is the set safety margin; the UAV flight path should be planned outside this safety area to ensure a safe distance from static obstacles during flight;

[0117] S32: For obstacles classified as dynamic, calculate the time and safety distance required for the UAV to avoid the obstacle based on its predicted trajectory and speed information; assume the current speed of the UAV is v u , the predicted speed of the obstacle is v d , and the relative speed of the two is v r ; then the calculation formulas for the avoidance time and safety distance are: D a = v u ×T a + D margin ; where D is the distance between the current UAV and the obstacle, and D margin is the additional safety margin during avoidance, T a represents the shortest time required for avoidance, and D a represents the minimum safety distance that the UAV should maintain from the obstacle during the avoidance process;

[0118] S33: Based on the safety area and avoidance parameters calculated in S31 and S32, generate the UAV flight path in real time; for static obstacles, the flight path of the UAV is adjusted by bypassing the outside of the safety area boundary; for dynamic obstacles, based on the avoidance time and safety distance, the flight direction and speed of the UAV are updated in real time to avoid in front of the moving direction of the obstacle and ensure a safe distance; the flight path planning update is executed based on continuous obstacle prediction data to ensure the real-time and accuracy of the UAV flight path adjustment in a complex environment; through the above steps, the safety area and avoidance path are determined for static and dynamic obstacles respectively in real time, realizing the dynamic adjustment of the UAV flight path. This planning method ensures the safety of the UAV in a complex flight environment and improves the navigation safety and path optimization effect of the UAV without affecting the task execution efficiency.

[0119] S4 specifically includes:

[0120] S41: Divide the flight area within the UAV mission scope into several small areas; specifically, with the UAV flight path as the center, generate multiple rectangular or polygonal grids at a fixed distance, where each grid area is called a small area; the size of each small area is fixed, and the boundaries of adjacent areas are clear, which are used to independently count the obstacle distribution information;

[0121] S42: In each small area, calculate the current obstacle density and the dynamic obstacle appearance frequency; assuming the area of the small area is A and the total number of obstacles is N, the obstacle density D is calculated as: Dynamic obstacle frequency F d By calculating the number N of dynamic obstacles entering the small area within a fixed time window d , and obtained according to the window duration T, the formula is: Obstacle density D and dynamic obstacle frequency F d Will be used as the reference standard for subsequent speed adjustment;

[0122] S43: According to the obstacle density D and the dynamic obstacle frequency F of each small area d , adjust the flight speed and data acquisition frequency of the UAV in real time; the specific adjustment criteria include:

[0123] Standard 1, when D exceeds the set dense threshold D thresh or F d exceeds the dynamic obstacle frequency threshold F thresh , reduce the flight speed of the UAV to the safe speed and increase the data acquisition frequency to the set maximum frequency to ensure more environmental information is obtained in the obstacle-dense area;

[0124] Standard 2, when D is lower than the set sparse threshold D low and F d is lower than the dynamic obstacle frequency threshold F low , adjust the flight speed of the UAV to the maximum speed v max Execute the mapping task to speed up the operation speed in the low-risk area and improve the mapping efficiency; during the flight of the UAV, continuously monitor the obstacle density and dynamic obstacle frequency data of the small area to adjust the flight speed and acquisition frequency in real time; when the UAV enters the obstacle-dense area from the obstacle-sparse area, gradually reduce the speed and increase the data acquisition frequency; conversely, when entering the sparse area, gradually increase the speed and reduce the acquisition frequency to ensure smooth speed adjustment and real-time response to changes in the surrounding environment; through the above steps, the UAV can flexibly adjust the flight speed and data acquisition frequency in the environment with uneven obstacle distribution, thus ensuring flight safety in the obstacle-dense area and improving the mapping efficiency in the sparse area, significantly improving the mapping effect and task execution efficiency of the UAV in the complex environment.

[0125] S5 specifically includes:

[0126] S51: When the current flight path of the UAV conflicts with the predicted path of the dynamic obstacle, judge the time point of approaching the obstacle on the path and the severity of the potential conflict by calculating the relative speed and relative position of the UAV and the obstacle; the specific calculation method is as follows:

[0127] Relative speed calculation: Let the speed of the drone be v u And the predicted speed of the obstacle is v d , the relative speed v between the two r Calculated as: v r =|v u -v d |,Relative speed is used to determine the relative approach speed between the UAV and the obstacle in order to predict the intersection time between the two;

[0128] Approach time calculation: Let the current distance between the drone and the obstacle be D, then the approach time is T c Calculated as: Approach time T c Indicates the estimated time it takes for the drone to encounter an obstacle under the current speed conditions;

[0129] Conflict determination: will approach time T c and the preset safety time threshold T s Compare, if T c ≤T s , it is determined that there is a potential conflict risk and avoidance measures need to be taken; otherwise, no avoidance measures are required.

[0130] S52: Based on the conflict assessment results of S51, the drone selects an avoidance strategy based on the speed and path of the obstacle. Specific strategies include temporary stopping and flight altitude adjustment. If the obstacle is moving at a high speed and passing quickly on the route, temporary stopping is preferred. The drone will make a short stop based on the time it takes the obstacle to pass, avoiding the conflict zone. If the obstacle is moving at a low speed or stays in the route for a long time, the drone will adjust its flight altitude to avoid the conflict zone vertically.

[0131] S53: When altitude adjustment is selected, set the new flight altitude of the drone to h new Meet the safety interval requirements; the calculation formula is: h new =h u +Δh, where h u is the current flight altitude; Δh is the safe altitude offset, ensuring a safe vertical distance from obstacles;

[0132] S54: When choosing to temporarily stop, the drone determines the waiting time based on the difference between the approach time and the obstacle passing time. The formula is: T w =T d -T c , where T w is the waiting time, T d is the obstacle passing time, T c is the approach time; if T w>0, the UAV will hover within a safe range from the conflict area and continue to fly along the original route after the obstacle has passed; through the above steps, the UAV can select the best avoidance strategy based on the proximity time, relative speed, and flight altitude calculated in real time to ensure safe passage in the dynamic obstacle conflict area and achieve safe and efficient flight in complex environments.

[0133] As Figure 2 shown, a UAV mapping system based on deep learning is used to implement the above-mentioned UAV mapping method based on deep learning, including the following modules:

[0134] Environmental data acquisition module: includes lidar, camera, and ultrasonic sensors, which are used to collect obstacle information in the surrounding environment of the UAV in real time, including the position, shape, and motion state of the obstacles, and transmit the collected data to the data preprocessing module;

[0135] Data preprocessing module: used to denoise and standardize the format of the data transmitted by the environmental data acquisition module to generate a standardized environmental data set;

[0136] Obstacle analysis and classification module: based on the environmental data set generated by the data preprocessing module, use a pre-trained convolutional neural network to extract features and classify the obstacles, distinguishing static obstacles and dynamic obstacles; for obstacles classified as dynamic, use a recurrent neural network to analyze their historical trajectories and predict their future motion paths and speed information;

[0137] Route planning module: used to update the UAV's route plan in real time according to the obstacle classification results and dynamic obstacle motion prediction data provided by the obstacle analysis and classification module; for static obstacles, the route planning module sets a safe flight area based on the obstacle position; for dynamic obstacles, calculate the avoidance time and safe distance and generate a new flight path to ensure the safety and path optimization of the UAV;

[0138] Flight control and speed adjustment module: receives the flight path and avoidance plan generated by the route planning module, and combines the obstacle density and dynamic obstacle frequency in the current flight area of the UAV to adjust the flight speed and data acquisition frequency of the UAV in real time; specifically, in areas with dense obstacles, reduce the flight speed and increase the data acquisition frequency; in areas with sparse obstacles, increase the flight speed to improve the mapping efficiency;

[0139] Conflict detection and avoidance module: used to detect potential conflicts on the route in real time, and when the UAV route intersects with the predicted path of a dynamic obstacle, select an appropriate avoidance strategy; including selecting flight altitude adjustment or temporary stop in case of severe conflicts to ensure the UAV can safely pass through the conflict area.

[0140] The present invention covers any alternatives, modifications, equivalent methods, and solutions made within the spirit and scope of the present invention. To enable the public to have a thorough understanding of the present invention, specific details are described in detail in the following preferred embodiments of the present invention. However, those skilled in the art can fully understand the present invention even without the description of these details. In addition, well-known methods, processes, flows, components, and circuits are not described in detail to avoid unnecessary confusion about the essence of the present invention.

[0141] The above are only the preferred embodiments of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the principle of the present invention, several improvements and refinements can be made, and these improvements and refinements should also be regarded as the protection scope of the present invention.

Claims

1. A drone mapping method based on deep learning, characterized in that, It includes the following steps: S1: During the flight of the drone, the obstacle information in the surrounding environment is collected in real time, including the position, shape and motion state, and the collected data is denoised and format standardized to generate an environmental data set; S2: Based on the data set generated in S1, the obstacles are classified to distinguish static obstacles and dynamic obstacles; For dynamic obstacles, a recurrent neural network is used to predict their movement trajectories to generate the future movement path and speed information of the obstacles; S3: Based on the obstacle classification and prediction data in S2, the flight route planning of the drone is updated in real time; for static obstacles, a safe flight area is determined based on their positions; for dynamic obstacles, the avoidance time and safe distance are calculated according to their predicted trajectories and a new flight path is generated; S4: The flight area is divided into several small areas, and the drone speed is adjusted according to the obstacle density and dynamic obstacle frequency in the current area; when in an area with dense obstacles, the flight speed is reduced and the data collection frequency is increased; when in an area with sparse obstacles, the mapping task is performed at the fastest speed; S5: When the flight route conflicts with the predicted path of a dynamic obstacle, according to the obstacle prediction result in S2, the flight altitude is selected to be adjusted or a temporary stop is made to ensure the safe passage of the drone through the conflict area.

2. The drone mapping method based on deep learning according to claim 1, wherein The specific content of S1 includes: S11: During the flight of the drone, the distance and three-dimensional spatial distribution information of the obstacles in the environment are obtained through the configured lidar; the two-dimensional image data and shape feature information of the obstacles are obtained through the configured camera; the real-time position information of the obstacles is obtained through the ultrasonic sensor and the relative distance between the obstacles and the drone is confirmed; to establish complete obstacle information, including the position, three-dimensional shape contour and its motion state of the obstacles; S12: The obstacle data of the lidar, camera and ultrasonic sensor obtained in step S11 are fused, and the spatial data of the lidar, the image data of the camera and the position information of the ultrasonic are matched to ensure that the position information, shape features and motion state of each obstacle correspond to each other; S13: The data fused in S12 is denoised. The distance data of the lidar and ultrasonic are filtered by applying the Kalman filter algorithm to eliminate the existing environmental noise; the Gaussian filter is applied to the camera image data to remove image noise; S14: The data after denoising in S13 is converted into a unified format, specifically including converting the distance and position coordinates into a unified three-dimensional coordinate system, and formatting the image data into a size and resolution that meet the input requirements of the deep learning model to generate a standardized environmental data set.

3. The drone mapping method based on deep learning according to claim 2, wherein, The specific content of S13 includes: S131: The distance data collected by the lidar and ultrasonic sensor are initially set, the state vector is defined to represent the distance and distance change rate of the obstacle, and the error covariance matrix is set to estimate the degree of measurement error; S132: At the current moment, predict the state vector and the error covariance matrix. Describe the evolution of distance information over time through the state transition matrix, while considering the influence of control inputs on the state and the noise in the system process to obtain the predicted state at the next moment; S133: Update the state using the measurement data from the lidar and ultrasonic sensors. Determine the balance between the actual measurement and the predicted state by calculating the Kalman gain, and correct the state vector and the error covariance matrix to obtain the denoised obstacle distance and speed data; S134: Apply a Gaussian filter to the image data collected by the camera. The Gaussian filter calculates at each pixel point of the image, synthesizes the influence of surrounding pixels through predetermined weights, and removes the noise points in the image.

4. The method for UAV mapping based on deep learning according to claim 1, characterized in that, The specific steps of S2 include: S21: Based on the standardized environmental dataset generated in S1, input the data into a deep learning model for feature extraction, including the spatial features, shape features, and relative motion features of the obstacles, to provide accurate obstacle attribute information for subsequent classification; S22: Use a pre-trained convolutional neural network model to analyze the extracted obstacle features, and classify the obstacles into two categories: static obstacles and dynamic obstacles. Specifically, by analyzing the position and shape changes of the obstacles in consecutive data frames, identify the obstacles with no obvious moving trend as static obstacles, and the obstacles with a continuous moving trend as dynamic obstacles; S23: For the obstacles classified as dynamic, input their feature data into a recurrent neural network. By analyzing the historical motion states of the obstacles, generate their future motion paths and speed information to estimate the positions and directions of the obstacles in the next time period; S24: Output the trajectory prediction results of S23 as the future motion paths and speed information of the dynamic obstacles, and transfer the prediction results to step S3 for real-time route planning adjustment.

5. The method for UAV mapping based on deep learning according to claim 4, characterized in that, The specific steps of S22 include: S221: Initialize the settings of the pre-trained convolutional neural network model, and configure the network parameters to adapt to the input of the obstacle feature data. The input feature data includes the obstacle images collected in different time frames and their position coordinate information; S222: Input the obstacle image data in consecutive frames into the convolutional neural network model, extract the image features of the obstacles through the convolutional layer to form feature maps. Each feature map of each time frame contains the position and shape information of the obstacles. Establish the feature sequence of the same obstacle in consecutive frames by matching with the obstacle features in adjacent frames; S223: Analyze the position change of the feature sequence matched in S222. Assume the current frame position coordinates are (x k , y k ), and the previous frame position coordinates are (x k-1 , y k-1 ). Then the displacement Δd k of the obstacle between these two frames is calculated as follows: Compare Δd k with the set displacement threshold T d . If Δd k is less than or equal to T d , then the displacement of the obstacle in the current frame is not sufficient to be judged as moving, and it is marked as a static obstacle candidate; S224: Analyze the shape change of the obstacle within consecutive frames. Let the shape features of the obstacle in the current frame and the previous frame be the area A k and A k-1 , and the shape change rate ΔA k is calculated as: Compare ΔA k with the shape change threshold T A . If ΔA k is less than or equal to T A , the shape change of the obstacle is not obvious, and it continues to be marked as a static obstacle candidate; S225: Comprehensively judge the position change Δd and shape change rate ΔA of the candidate obstacle in multiple consecutive frames; if the displacement amount and shape change rate of the candidate obstacle within several consecutive frames are both less than their respective thresholds T d and T A , then identify the obstacle as a static obstacle; otherwise, identify it as a dynamic obstacle.

6. The method for unmanned aerial vehicle mapping based on deep learning according to claim 5, wherein, The specific steps of S23 include: S231: Initialize the recurrent neural network model, and set the input of the model to the position coordinates, speed, and motion direction angles of the dynamic obstacles in consecutive time frames; S232: Normalize the position and speed data of the obstacles in historical time frames to reduce the influence of data differences on model training; S233: Gradually input the processed historical trajectory data into the hidden layer of the recurrent neural network; for the input of each time frame, the recurrent neural network model updates its state based on the previous state and the current input, gradually establishing the motion pattern of the obstacle to identify the movement rule of the obstacle; S234: According to the final state of the recurrent neural network model, output the predicted position and speed information of the dynamic obstacle, generating the future position coordinates and speed data; S235: Based on the predicted future position and speed data, calculate the movement direction angle of the obstacle to determine its future traveling direction, and after combining the predicted direction with the speed, form the complete future movement path information.

7. A method for UAV mapping based on deep learning according to claim 1, characterized in that The specific content of S3 includes: S31: Based on the obstacle classification data in S2, for static obstacles, by obtaining the exact positions of the obstacles, establish the safe flight areas of the static obstacles; assume the boundaries of the obstacles are circular or polygonal, and define the boundary distance of the safe flight area as the safety distance R extended outward from the obstacle boundary s , and the calculation formula is: R S = R o + D margin , where R o is the radius of the obstacle or the distance from the outermost edge to the center, and D margin is the set safety margin; S32: For the obstacles classified as dynamic, calculate the time and safety distance required for the UAV to avoid the obstacle based on its predicted trajectory and speed information; assume the current speed of the UAV is v u , the predicted speed of the obstacle is v d , and the relative speed of the two is v r ; then the calculation formulas for the avoidance time and safety distance are: D a = v u × T a + D margin ; where D is the distance between the current UAV and the obstacle, D margin is the additional safety margin during avoidance, T a represents the shortest time required for avoidance, and D a represents the minimum safety distance that the UAV should maintain from the obstacle during the avoidance process; S33: Based on the safety area and avoidance parameters calculated in S31 and S32, generate the flight path of the UAV in real time; for static obstacles, the route planning adjusts the UAV flight path by bypassing the outside of the boundary of its safety area; for dynamic obstacles, based on the avoidance time and safety distance, update the flight direction and speed of the UAV in real time to make it avoid in front of the movement direction of the obstacle and ensure a safe distance.

8. A method for UAV mapping based on deep learning according to claim 1, characterized in that, The specific content of S4 includes: S41: Divide the flight area within the UAV mission range into several small areas; specifically, centered on the UAV flight route, generate multiple rectangular or polygonal grids at a fixed distance, and each grid area is called a small area; S42: In each small area, calculate the current obstacle density and the dynamic obstacle appearance frequency; assuming the area of the small area is A and the total number of obstacles is N, the obstacle density D is calculated as: The dynamic obstacle frequency F d By calculating the number N of dynamic obstacles entering the small area within a fixed time window d , and obtained according to the window duration T, the formula is: S43: According to the obstacle density D and dynamic obstacle frequency F of each small area d , adjust the flight speed and data collection frequency of the UAV in real time; the specific adjustment criteria include: Standard 1, when D exceeds the set density threshold D thres h or F d exceeds the dynamic obstacle frequency threshold F thres h the flight speed of the drone is reduced to a safe speed, and the data acquisition frequency is increased to the set maximum frequency to ensure more environmental information is obtained in the area with dense obstacles; Standard 2, when D is lower than the set sparse threshold D low and F d is lower than the dynamic obstacle frequency threshold F low adjust the flight speed of the drone to the maximum speed v max Execute the mapping task to speed up the operation in low-risk areas.

9. The method for unmanned aerial vehicle mapping based on deep learning according to claim 5, characterized in that The specific content of S5 includes: S51: When the current flight route of the UAV conflicts with the predicted path of the dynamic obstacle, judge the time point of approaching the obstacle on the path and the severity of the potential conflict by calculating the relative speed and relative position of the UAV and the obstacle; S52: Based on the conflict assessment result of S51, the UAV selects an avoidance strategy according to the speed and path situation of the obstacle; the specific strategies include temporarily staying and adjusting the flight altitude; if the obstacle speed is high and it passes quickly on the flight route, preferentially select to temporarily stay; if the obstacle speed is low or it stays in the flight route area for a long time, select to adjust the flight altitude to avoid the conflict area in the vertical direction; S53: When altitude adjustment is selected, set the new flight altitude of the drone to h new ; The calculation formula is: h new = h u + Δh, where h u is the current flight altitude; Δh is the safety altitude offset; S54: When choosing to make a temporary stop, the drone determines the waiting time based on the difference between the approach time and the obstacle passage time. The formula is: T w = T d - T c , where T w is the waiting time, T d is the obstacle passage time, and T c is the approach time; if T w > 0, the drone will hover within the safe range from the conflict area and continue to fly along the original route after the obstacle has passed.

10. A drone mapping system based on deep learning, which is used to implement a drone mapping method based on deep learning according to any one of claims 1-9, characterized in that, It includes the following modules: Environmental data acquisition module: It includes lidar, camera and ultrasonic sensor, which are used to collect the obstacle information in the environment around the UAV in real time, including the position, shape and motion state of the obstacle, and transmit the collected data to the data preprocessing module; Data preprocessing module: It is used to perform denoising and format standardization processing on the data transmitted by the environmental data acquisition module to generate a standardized environmental data set; Obstacle analysis and classification module: Based on the environmental data set generated by the data preprocessing module, use the pre-trained convolutional neural network to extract features and classify the obstacles, distinguishing static obstacles and dynamic obstacles; for the obstacles classified as dynamic, use the recurrent neural network to analyze their historical trajectories and predict their future movement paths and speed information; Route Planning Module: It is used to update the route planning of the UAV in real time according to the obstacle classification results provided by the Obstacle Analysis and Classification Module and the motion prediction data of dynamic obstacles; for static obstacles, the Route Planning Module sets a safe flight area based on the obstacle position; for dynamic obstacles, it calculates the avoidance time and safe distance and generates a new flight path; Flight Control and Speed Adjustment Module: It receives the flight path and avoidance plan generated by the Route Planning Module, and combines the obstacle density and dynamic obstacle frequency in the current flight area of the UAV to adjust the flight speed and data acquisition frequency of the UAV in real time; specifically, when in an area with dense obstacles, it reduces the flight speed and increases the data acquisition frequency; when in an area with sparse obstacles, it increases the flight speed to improve the mapping efficiency; Conflict Detection and Avoidance Module: It is used to detect potential conflicts on the flight route in real time. When the UAV flight route intersects with the predicted path of a dynamic obstacle, it selects an appropriate avoidance strategy; This includes selecting flight altitude adjustment or temporary stay when the conflict is severe to ensure that the UAV safely passes through the conflict area.

Images