A deep learning-based unmanned aerial vehicle (UAV) safe flight path planning method in an urban environment

Abstract

The present application relates to the technical field of unmanned aerial vehicle path planning, and belongs to a kind of unmanned aerial vehicle safe flight path planning method in urban environment based on deep learning, aiming at solving the safety and flight path optimization problem of unmanned aerial vehicle when flying in complex urban environment. The method realizes the automatic generation of optimal safe flight path by combining airspace initialization module, environment perception module, dynamic risk assessment module, path planning module and deep learning module through deep reinforcement learning algorithm, avoids obstacle collision and optimizes flight efficiency. This method can dynamically adjust the flight path, respond to environmental changes in real time, and ensure the safe and efficient flight of unmanned aerial vehicles in urban environment. It is suitable for intelligent transportation, logistics distribution, environmental monitoring and other fields, and has wide application prospect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle path planning, and belongs to a method for safe flight path planning of unmanned aerial vehicles in urban environments based on deep learning. BACKGROUND

[0002] With the rapid development of low-altitude economy, the application scenarios of unmanned aerial vehicles are continuously expanding, gradually penetrating from traditional aerial photography and surveying to high-value-added fields such as logistics distribution, urban monitoring, and emergency rescue. In particular, in the logistics industry, the development of low-altitude economy has promoted the improvement of "last mile" distribution efficiency and provided an innovative path for urban traffic relief. In urban environments, unmanned aerial vehicles not only can quickly respond to emergencies, but also can undertake tasks such as environmental monitoring and medical supply distribution, providing new technical support for smart city construction. However, the prosperity of low-altitude economy also puts higher requirements on the safety of unmanned aerial vehicles. The high-density buildings, complex traffic flow, dynamic obstacles, and variable weather conditions in cities make the flight path planning of unmanned aerial vehicles face severe challenges. Traditional path planning methods are difficult to meet the real-time and safety requirements in complex urban scenarios, hindering the further development of low-altitude economy. The present application combines deep learning and dynamic environment perception technology to adjust the flight path planning of unmanned aerial vehicles in real time, providing a more efficient and safer solution for the booming low-altitude economy and laying a solid foundation for promoting the intelligent and efficient use of urban airspace resources. SUMMARY

[0003] To solve the above technical problems, the present application provides a method for safe flight path planning of unmanned aerial vehicles in urban environments based on deep learning.

[0004] To achieve the above purpose, the present application provides the following technical solutions:

[0005] On the one hand, the present application provides a method for safe flight path planning of unmanned aerial vehicles in urban environments based on deep learning, comprising the following steps:

[0006] S1, airspace initialization: the flight area is rasterized into a series of equidistant three-dimensional square cells, the flight area boundary is constructed, and the functional area is divided; based on the functional area division, an initial risk coefficient is assigned to each grid, a risk coefficient table is generated, and the flight boundary and start and end point coordinates are marked;

[0007] S2, input flight task, specify start point coordinates and end point coordinates;

[0008] S3, dynamic flight path generation and update:

[0009] S31, initial path planning: based on the risk coefficient table, an improved A* algorithm is used to generate an initial path;

[0010] S32, path smoothing processing: using cubic spline curve to smooth the path nodes to generate a smooth static route, meeting the kinematic constraints of the UAV;

[0011] S33, flight execution and perception: the UAV flies according to the smoothed static route, while collecting environmental data in real time;

[0012] S34, dynamic risk update: update the grid risk coefficient by fusing perception data;

[0013] S35, deep learning prediction: organize historical risk data into a time series matrix and input it into an LSTM model to predict future risk coefficients and update the risk coefficient table. When the risk suddenly increases or the path is blocked, return to step S31 to regenerate the route.

[0014] As a preferred, step S1 airspace initialization process: each cell is defined by its lower left corner coordinates (x, y, z) and edge length L, where the edge length L is determined by the physical size of the UAV, flight accuracy and average flight height:

[0015]

[0016] Where: d is the diameter of the UAV, p is the positioning system error range, and h is the average flight height.

[0017] As a preferred, step S1 airspace initialization process:

[0018] The functional area division includes densely populated areas, building concentrated areas and traffic dense areas;

[0019] The initial risk coefficient is assigned by matching the grid position with the predefined regional risk coefficient table.

[0020] As a preferred, step S31 improved A* algorithm: its calculation formula is:

[0021]

[0022] f(n) is the comprehensive evaluation function value of node n;

[0023] h( n ) is the estimated cost from node n to the target node;

[0024] is the preliminary estimated cost calculated based on the spatial position relationship between node n and the target node;

[0025] is the risk estimation cost that node n may face on the path from node n to the target node, combined with current environmental perception and deep learning prediction;

[0026] g( n) represents the integrated actual cost from the start node to node n;

[0027] wherein r(n) is the risk factor of the grid cell where node n is located;

[0028] w r is the risk weight factor.

[0029] As preferred, the path smoothing processing in step S32 comprises:

[0030] fitting the path nodes by piecewise cubic spline curve equation;

[0031] applying continuity condition and free boundary condition to preliminarily generate the smoothed path;

[0032] verifying whether the curvature radius R(t) of all nodes of the smoothed path is greater than the minimum curvature radius Rmin of the UAV and detecting whether the smoothed path collides with the obstacles, and continuously optimizing the smoothed path.

[0033] As preferred, the environmental data in step S33 comprises:

[0034] dynamic and static obstacle information, weather data and crowd density, and the future trajectory of the dynamic obstacle is predicted by using Kalman filter.

[0035] As preferred, the LSTM model in step S35 comprises:

[0036] input layer: receiving multi-dimensional time series data;

[0037] hidden layer: using ReLU activation function to enhance the non-linear fitting ability; adding Dropout layer to avoid overfitting;

[0038] output layer: predicting the risk factor of the future K time steps and outputting, wherein the mean square error loss function is used to measure the error between the predicted risk factor and the real risk factor, and drive the model optimization.

[0039] On the other hand, the present application also provides a UAV safe flight path planning system in urban environment based on deep learning, which is used to realize the above-mentioned UAV safe flight path planning method in urban environment based on deep learning, comprising airspace initialization module, environment perception module, dynamic risk assessment module, dynamic risk assessment module and deep learning module.

[0040] airspace initialization module: used to rasterize the flight map, divide it into a series of adjacent three-dimensional square cells according to the size, flight accuracy and information of the UAV, assign it an initial risk factor according to the region where it is located, and construct the flight region boundary;

[0041] An environmental perception module: real-time collection of environmental information through sensors of the UAV, including population density below the UAV, static and dynamic obstacles, meteorological conditions and other data, combined with perception data to predict potential risks in the environment;

[0042] A dynamic risk assessment module: based on a deep learning model and real-time perception information, dynamically updating a risk coefficient table, and predicting and evaluating the risk of each grid unit of the flight area;

[0043] A path planning module: using an improved A* algorithm, combined with a dynamic risk coefficient table and aircraft kinematic constraints, planning a safe and efficient flight path, and adjusting in real time during flight;

[0044] A deep learning module: processing dynamic environmental perception data through an LSTM time series model, and predicting future risk changes using a deep learning algorithm to optimize path planning.

[0045] Compared with the prior art, the present application provides a UAV safe flight path planning method in urban environment based on deep learning, which has the following beneficial effects:

[0046] The airspace initialization module of the present application: for gridding the flight map, dividing it into a series of adjacent three-dimensional square cells according to the size of the UAV, flight accuracy and information, assigning it an initial risk coefficient according to its region, and constructing the flight area boundary; the environmental perception module: real-time collection of environmental information through sensors of the UAV, including population density below the UAV, static and dynamic obstacles, meteorological conditions and other data, combined with perception data to predict potential risks in the environment; the dynamic risk assessment module: based on a deep learning model and real-time perception information, dynamically updating a risk coefficient table, and predicting and evaluating the risk of each grid unit of the flight area; the path planning module: using an improved A* algorithm, combined with a dynamic risk coefficient table and aircraft kinematic constraints, planning a safe and efficient flight path, and adjusting in real time during flight; the deep learning module: processing dynamic environmental perception data through an LSTM time series model, and predicting future risk changes using a deep learning algorithm to optimize path planning. The present application improves the safety of the planned path in urban environment by interacting with the environment through the UAV and updating the flight route through deep learning.

[0047] The features and advantages of the present application will be described in detail in conjunction with the embodiments and the accompanying drawings. BRIEF DESCRIPTION OF DRAWINGS

[0048] Fig. 1 A flowchart of the UAV safe flight path planning method in urban environment based on deep learning of the present application;

[0049] Fig. 2A flowchart of a process for dynamically sensing and updating a risk coefficient table during flight of the unmanned aerial vehicle of the present application. DETAILED DESCRIPTION

[0050] To make the objectives, technical solutions and advantages of the present application clearer and more comprehensible, the present application is further described in detail below with reference to the drawings and examples. However, it should be understood that the specific examples described herein are merely used to explain the present application and are not used to limit the scope of the present application. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present application.

[0051] Referring to Figs. 1-2 A deep learning-based unmanned aerial vehicle safe flight path planning method in an urban environment includes the following steps:

[0052] S1, airspace initialization: the flight area is rasterized into a series of equidistant three-dimensional square cells, the flight area boundary is constructed, and the functional area is divided; based on the functional area division, an initial risk coefficient is assigned to each grid, a risk coefficient table is generated, and the flight boundary and start and end point coordinates are marked;

[0053] S2, input flight task, specify start point coordinates and end point coordinates;

[0054] S3, dynamic path generation and update:

[0055] S31, initial path planning: based on the risk coefficient table, an improved A* algorithm is used to generate an initial path;

[0056] S32, path smoothing processing: using a cubic spline curve to smooth the path nodes to generate a smooth static flight path, satisfying the kinematic constraints of the unmanned aerial vehicle;

[0057] S33, flight execution and sensing: the unmanned aerial vehicle flies according to the smoothed static flight path, while collecting environmental data in real time;

[0058] S34, dynamic risk update: fuse the sensing data to update the grid risk coefficient;

[0059] S35, deep learning prediction: organize historical risk data into a time series matrix and input it into an LSTM model to predict future risk coefficients and update the risk table; when the risk suddenly increases or the path is blocked, return to step S31 to regenerate the route.

[0060] Preferably, the improved A* algorithm of step S31: the calculation formula is:

[0061]

[0062] Formula meaning: f(n) is the comprehensive evaluation function value of node n, which is used to determine the priority of the node in path search. In the search process of A* algorithm, the node with the smallest value is selected for expansion each time. By adding (the actual cost from the starting node to the current node) and (the estimated cost from the current node to the target node), both the cost of the path already traveled and the estimated cost of reaching the target node are considered, enabling the algorithm to quickly and effectively explore towards the target node during the search process, while avoiding falling into local optimal solution, improving the efficiency and quality of path planning.

[0063]

[0064] Formula meaning: h(n) is the estimated cost from node n to the target node. Among them, is the preliminary estimated cost calculated based on the spatial location relationship between node n and the target node, using the Euclidean distance formula to estimate the distance cost from node n to the target node. is the risk estimation cost from node n to the target node path considering the current environment perception and deep learning prediction, which comprehensively considers the influence of risk coefficient distribution, dynamic obstacle movement trend, weather condition change and other factors on reaching the target node, so that can more comprehensively reflect the comprehensive estimated difficulty from node n to the target node, and further guide the algorithm to more reasonably select the search direction.

[0065]

[0066] Formula meaning: g(n) represents the comprehensive actual cost from the starting node to node n. It consists of two parts, is the cost from the starting node to node n calculated based on the Euclidean distance, is the risk weight factor, which is used to measure the importance of risk in the total cost, and r(n) is the risk coefficient of the grid cell where node n is located. Through this formula, the risk factor is quantified and integrated into the actual cost calculation of path search, so that the algorithm not only considers the length of the path, the difficulty of flight and other conventional factors when planning the path, but also gives corresponding cost penalty to the area with high risk, so as to guide the path planning to tend to low risk area.

[0067] Preferably, in the airspace initialization of step S1: each cell is defined by its left lower corner coordinates (x, y, z) and edge length L, where the edge length L is determined by the physical size of the UAV, flight accuracy and average flight height:

[0068]

[0069] where d is the diameter of the drone, p is the positioning system error range, and h is the average flight height.

[0070] Specifically, in the flight map rasterization process, the size of the drone, flight accuracy, and information are divided into a series of adjacent three-dimensional square cells.

[0071] Preferably, in the airspace initialization step S1:

[0072] The functional division includes densely populated areas, building-intensive areas, and traffic-intensive areas. To accurately assess the risks that may be encountered during drone flight, the flight area is first divided into regions, and the map is divided into different functional areas according to certain rules. Each region represents different flight environment characteristics, including but not limited to population density, building density, traffic flow, and other factors that can affect the safety of drone flight.

[0073] The initial risk coefficient is assigned by matching the grid position with the predefined regional risk coefficient table; each region has a specific risk coefficient, for example, a high population density area has a higher risk coefficient, and a traffic-intensive area may have a higher flight risk. These risk coefficients are stored in a table called the regional risk table. This table provides a safety assessment value for each region, which quantifies the risk based on different flight environment conditions. For each grid, its initial risk coefficient is obtained according to its region and the regional risk table.

[0074] Preferably, the path smoothing process in step S32 includes: to more realistically simulate the flight path, a three-dimensional spline curve method is used to smooth the path, perform curve turning, and ensure that the path is executable in the real environment; extract the node data of the path to be smoothed: obtain the three-dimensional coordinate information of all nodes from the preliminary flight path generated by the improved A* algorithm, and use a cubic spline curve to smooth the path. The cubic spline curve has good smoothness and local controllability, which can better meet the requirements of the drone flight path.

[0075] The path nodes are fitted by a cubic spline curve equation in sections;

[0076] A continuity condition and a free boundary condition are applied to preliminarily generate a smoothed path;

[0077] Verify whether the curvature radius R(t) of all points of the smoothed path is greater than the minimum curvature radius Rmin of the drone and detect whether the smoothed path collides with obstacles, continuously optimize the smoothed path until the curvature radius of all points is greater than the minimum curvature radius of the drone, which meets the flight characteristic requirements of the drone, i.e., the minimum curvature radius limit.

[0078] Through the implementation steps of the above three-dimensional spline curve method, the preliminary flight path of the unmanned aerial vehicle can be effectively smoothed, the stability and safety of the unmanned aerial vehicle flight are improved, and the flight requirements of the unmanned aerial vehicle in the complex urban environment are met.

[0079] Preferably, the environmental data in step S33 comprises:

[0080] Dynamic obstacle information, weather data, and crowd density are used to predict the future trajectory of dynamic obstacles using Kalman filtering.

[0081] Based on the described dynamic risk coefficient calculation of the unmanned aerial vehicle, the risk coefficient formula is designed as follows, combining the influence of dynamic obstacles, crowd density, and weather conditions:

[0082] Let represent the risk coefficient of the unmanned aerial vehicle in the coordinate at time t:

[0083]

[0084] Wherein:

[0085] : The weight coefficients of static risk, dynamic risk, weather risk, and crowd risk, respectively, are used to adjust the contribution of different factors to the overall risk.

[0086] Static risk coefficient: ; Dynamic risk coefficient: ;

[0087] Weather risk coefficient: Combined with real-time weather information (wind speed and precipitation);

[0088] Crowd density risk coefficient: .

[0089] Preferably, the LSTM model in step S35 comprises:

[0090] Input layer: receives multi-dimensional time series data;

[0091] Hidden layer: uses ReLU activation function to enhance non-linear fitting ability; adds Dropout layer to avoid overfitting;

[0092] Output layer: predicts the risk coefficient of the next K time steps and outputs, wherein the mean square error loss function is used to measure the error between the predicted risk coefficient and the true risk coefficient to drive model optimization.

[0093] The data processing steps of the LSTM model include:

[0094] Data preprocessing: collect real-time acquired risk data,

[0095] Model input: Organize time series into a multi-dimensional matrix, with each row corresponding to a time step of risk features, into a time series feature matrix;

[0096] Input real-time perception data, historical risk coefficient table, output current risk value and short-term prediction value of each grid unit;

[0097] Based on the output of the LSTM model, dynamically update the risk coefficient table.

[0098] The application also provides a deep learning-based unmanned aerial vehicle safety flight path planning system in an urban environment, which is used to implement the above-mentioned deep learning-based unmanned aerial vehicle safety flight path planning method in an urban environment, and comprises an airspace initialization module, an environment perception module, a dynamic risk assessment module, a dynamic risk assessment module and a deep learning module.

[0099] The airspace initialization module is used to rasterize the flight map, divide a series of adjacent three-dimensional square cells according to the size of the unmanned aerial vehicle, flight accuracy and information, and assign an initial risk coefficient to each cell according to the region where the cell is located, and construct the flight region boundary.

[0100] The environment perception module: real-time collection of environmental information by sensors of the unmanned aerial vehicle, including population density below the unmanned aerial vehicle, static and dynamic obstacles, meteorological conditions and other data, and prediction of potential risks of the environment in combination with the perception data.

[0101] The dynamic risk assessment module: based on the deep learning model and real-time perception information, dynamically updating the risk coefficient table, and predicting and evaluating the risk of each grid unit in the flight region;

[0102] The path planning module: using an improved A* algorithm, combining the dynamic risk coefficient table and the kinematics constraints of the aircraft, planning a safe and efficient flight path, and adjusting in real time during flight;

[0103] The deep learning module: processing dynamic environment perception data through an LSTM time series model, and predicting future risk changes using a deep learning algorithm to optimize path planning.

[0104] In one embodiment, the airspace division module specifically comprises: rasterizing the flight region, dividing the map into a plurality of equidistant three-dimensional square grid cells (grids), each grid cell being defined by a three-dimensional plane lower left corner position and a side length, and the side length of the grid cell being determined by the physical size of the unmanned aerial vehicle, the flight accuracy and the average flight height;

[0105] Marking the boundary of the flight region on the rasterized map, and marking the starting point coordinates and the ending point coordinates, the grid center points of the starting point and the ending point being represented by and respectively.

[0106] According to the regional division result, a three-dimensional regional map is generated, in which the coordinates of each grid cell represent its position in the map, for unmanned aerial vehicle flight strategy planning and path optimization;

[0107] According to the region where each grid is located, its initial risk coefficient is assigned.

[0108] In an embodiment, the initial risk coefficient of each grid is assigned according to the region where it is located, which specifically includes:

[0109] The flight area is divided into different functional areas, including but not limited to densely populated areas, building concentrated areas and traffic flow concentrated areas;

[0110] A regional risk coefficient table corresponding to each functional area is established, and the corresponding risk coefficient is obtained by matching the current position with the regional division map;

[0111] In an embodiment, the environment perception module specifically includes: dynamic and static obstacle perception: through the laser radar and camera carried by the unmanned aerial vehicle, the surrounding environment is scanned in real time to identify dynamic and static obstacles in the flight area. For static obstacles, the position and shape of the obstacle are obtained based on the ground map or environment modeling; for dynamic obstacles, the position, speed, direction and other information of the dynamic obstacles are obtained in real time, and the motion trajectory of the dynamic obstacles is predicted by comparing the sensor data at different time points; trajectory prediction: based on the real-time obtained dynamic obstacle information, the future trajectory of the dynamic obstacle is predicted by using Kalman filter. Through the trajectory prediction of multiple target dynamic obstacles, the potential collision risk can be calculated in advance and included in the dynamic risk coefficient table update;

[0112] Weather condition perception: through the weather sensor, real-time wind speed, precipitation and cloud distribution data are collected to provide environmental constraint information for path planning, and are included in the dynamic risk coefficient table update;

[0113] Crowd density perception: the visual sensor carried is used to identify the number and distribution of crowds under the unmanned aerial vehicle through target detection algorithm; the crowd density is converted into corresponding risk value as part of the risk assessment.

[0114] In an embodiment, the dynamic risk assessment module includes:

[0115] According to the perception data and deep learning model, the risk change of each grid cell is predicted, and the risk coefficient table is dynamically updated;

[0116] The short-term risk trend is predicted by integrating obstacle trajectory, weather condition and historical flight data;

[0117] Utilize the environmental perception results to determine the safety of the path and trigger path re-planning if necessary; in one embodiment, the path planning module includes:

[0118] Initial path planning: using improved A* algorithm, based on risk coefficient table and flight targets, calculate the static optimal path;

[0119] Path smoothing: combined with three-dimensional spline curve method, generate smooth path that meets the requirements of unmanned aerial vehicle flight; dynamically adjust the path:

[0120] When an emergency occurs during flight, dynamically adjust the flight route to ensure flight safety;

[0121] In one embodiment, the dynamic path adjustment: when an emergency occurs during flight, dynamically adjust the flight route to ensure flight safety includes:

[0122] Through real-time monitoring of environmental changes in the flight area through perception data, when a significant increase in risk or path obstruction is detected, trigger route re-planning;

[0123] In the case of dynamic obstacle approach, trigger the automatic obstacle avoidance module to ensure flight safety, and then re-plan the route;

[0124] In one embodiment, the deep learning module includes:

[0125] Use LSTM time series model to predict dynamic risk changes;

[0126] Input real-time perception data, historical risk coefficient table into the model, output the current risk value and short-term prediction value of each grid unit;

[0127] Based on the prediction results of the deep learning model, dynamically update the risk coefficient table and feed back to the path planning module.

[0128] The above only describes the preferred embodiments of the present application, and is not intended to limit the present application, any modification, equivalent replacement or improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A method for safe route planning of unmanned aerial vehicles (UAVs) in urban environments based on deep learning, characterized in that: The method comprises the following steps: S1, airspace initialization: the flight area is rasterized into a series of equidistant three-dimensional cubic cells, the flight area boundary is constructed, and the functional area is divided; based on the functional area division, an initial risk coefficient is assigned to each grid, a risk coefficient table is generated, and the flight boundary and the start and end point coordinates are marked; S2, input flight task, specify start point coordinates and end point coordinates; S3, dynamic route generation and update: S31, initial path planning: based on the risk coefficient table, an improved A* algorithm is used to generate an initial path; S32, path smoothing processing: using a cubic spline curve to smooth the path nodes to generate a smooth static route, meeting the kinematic constraints of the unmanned aerial vehicle; S33, flight execution and perception: the unmanned aerial vehicle flies according to the smoothed static route, and simultaneously collects environmental data in real time; S34, dynamic risk update: the risk coefficient of the grid is updated by fusing the perception data; S35, deep learning prediction: the historical risk data is organized into a time series matrix and input into an LSTM model to predict future risk coefficients and update the risk coefficient table. When the risk suddenly increases or the path is blocked, return to step S31 to regenerate the route.

2. The method of claim 1, wherein the method is based on deep learning. The step S1 airspace initialization process: Each cell is defined by its lower left corner coordinates (x, y, z) and edge length L, wherein the edge length L is determined by the physical size of the unmanned aerial vehicle, the flight accuracy and the average flight height: Wherein: d is the diameter of the unmanned aerial vehicle, p is the positioning system error range, and h is the average flight height. 3.The method of claim 1, wherein: The step S1 airspace initialization process: The functional area division includes densely populated areas, building concentrated areas and traffic dense areas; The initial risk coefficient is assigned by matching the grid position with the pre-defined regional risk coefficient table. 4.The method of claim 1, wherein: The path smoothing processing in step S32 includes: Segment fitting of path nodes by cubic spline curve equation; Apply continuity condition and free boundary condition to preliminarily generate a smooth path; Verify whether the curvature radius R(t) of all nodes of the smooth path is greater than the specified minimum curvature radius Rmin of the unmanned aerial vehicle, and detect whether the smooth path collides with obstacles, and continuously optimize the smooth path.

5. The method of claim 1, wherein: The improved A* algorithm of step S31: the calculation formula is: f(n) is the comprehensive evaluation function value of node n; h( n ) is an estimated cost from node n to the target node; is a preliminary estimated cost calculated based on the spatial location relationship between the node n and the target node; h risk ( n ) is the prediction cost of the risk that node n can face on the path to the target node combining current environment perception and deep learning prediction; g( n ) denotes the integrated actual cost from the start node to node n; Wherein r(n) is the risk coefficient of the grid cell where node n is located; w r is the risk weight factor.

6. The method of claim 1, wherein the method is based on deep learning. The environmental data in step S33 includes: Dynamic and static obstacle information, weather data and crowd density, and Kalman filter is used to predict the future trajectory of dynamic obstacles.

7. The method of claim 1, wherein the method is based on deep learning. The LSTM model in step S35 includes: Input layer: receiving multi-dimensional time series data; Hidden layer: using ReLU activation function to enhance non-linear fitting ability; adding Dropout layer to avoid overfitting; Output layer: predicting the risk coefficient of the next K time steps and outputting, wherein the mean square error loss function is used to measure the error between the predicted risk coefficient and the true risk coefficient, and drive the model optimization.

8. A deep learning-based unmanned aerial vehicle (UAV) safe flight path planning system in urban environment, configured to implement the deep learning-based unmanned aerial vehicle (UAV) safe flight path planning method in urban environment according to any one of claims 1-7. It comprises an airspace initialization module, an environment perception module, a dynamic risk assessment module, a dynamic risk assessment module and a deep learning module; Space initialization module: used to rasterize the flight map, divide it into a series of adjacent three-dimensional cubic cells according to the size of the UAV, flight accuracy and information, assign initial risk coefficients according to the area it is located in, and construct the boundaries of the flight area; Environmental perception module: real-time collection of environmental information through the sensors of the UAV, including population density below the UAV, static and dynamic obstacles, and meteorological condition data, combined with perception data to predict potential risks in the environment; Dynamic risk assessment module: based on deep learning models and real-time perception information, dynamically update the risk coefficient table, and predict and evaluate the risk of each grid cell in the flight area; Path planning module: using an improved A* algorithm, combined with the dynamic risk coefficient table and the kinematic constraints of the aircraft, to plan a safe and efficient flight path and adjust it in real time during flight; Deep learning module: processing dynamic environmental perception data through an LSTM time series model and predicting future risk changes using deep learning algorithms to optimize path planning.

Images