A drone intelligent path planning system

By generating a time-space aerodynamic tolerance corridor through state wind field estimation and aerodynamic disturbance assessment, the problem of wind field adaptation for UAVs in complex urban environments is solved, improving flight safety and efficiency while reducing hardware costs and power consumption.

CN123111279APending Publication Date: 2026-11-17MINIVISION
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611265025.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing UAV path planning technology cannot adapt to non-uniform wind fields in complex urban environments, resulting in insufficient resistance to crosswind disturbances, insufficient wind field information perception, and the inability to achieve coupling and adaptation between channel geometry and wind field characteristics, leading to low flight safety and efficiency.

Method used

The system employs a state wind field estimation module, an environmental map construction module, an aerodynamic disturbance assessment module, and a tolerance corridor generation module. By combining an extended Kalman filter and the A* algorithm, it generates a temporal and aerodynamic tolerance corridors. Dynamic adaptive flight path planning is achieved through local three-dimensional wind force vector assessment and convex polyhedral corridor modification.

Benefits of technology

It enables safe flight of drones in complex wind fields, improves crosswind resistance and flight efficiency, ensures the safety and passability of flight passage, and reduces hardware costs and power consumption.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123111279A_ABST
    Figure CN123111279A_ABST
Patent Text Reader

Abstract

This invention provides an intelligent path planning system for unmanned aerial vehicles (UAVs), comprising: a state wind field estimation module that calculates a local three-dimensional wind force vector based on the difference between the rotational speed commanded by the flight controller and the actual rotational speed fed back by the electronic speed controller; an environment map construction module that converts the raw point cloud collected by the depth camera into a collision-free voxel map of the environment and constructs a convex polyhedral corridor from the current position to the target point in the collision-free voxel map; an aerodynamic disturbance assessment module that assesses the offset of the UAV when subjected to aerodynamic disturbances based on the local three-dimensional wind force vector; a tolerance corridor generation module that modifies the convex polyhedral corridor into an asymmetric spatiotemporal aerodynamic tolerance corridor based on the offset; and a trajectory planning and control module that performs UAV path planning based on the spatiotemporal aerodynamic tolerance corridor. By estimating the state of the UAV in the local wind field, the system adjusts the safety margin in each direction of the basic convex polyhedral corridor, so that sufficient buffer can be provided in the windward direction during trajectory planning, and space occupation can be reduced in the safe direction.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to an intelligent path planning system for unmanned aerial vehicles (UAVs). Background Technology

[0002] Currently, in autonomous flight operations of drones in scenarios such as urban logistics delivery, building inspection, and emergency rescue, the mainstream path planning technology generally adopts a hierarchical planning architecture. First, it relies on 3D point cloud data to complete the 3D reconstruction of the operating environment. Based on the expansion of the free space of the environment, it constructs a safe flight corridor formed by continuous splicing of multiple convex polyhedra. Then, using this flight corridor as a constraint, it solves and generates the drone motion trajectory that meets the requirements of flight smoothness through optimization algorithms such as Model Predictive Control (MPC) and Minimum Jerk.

[0003] The existing methods for constructing flight corridors generally adopt a safety margin setting mechanism of uniform expansion in all directions. That is, the spatial expansion is completed only based on the radius of the UAV body plus a fixed safety buffer threshold. The core theoretical premise of this construction method is that the external airflow disturbances experienced by the UAV are symmetrical in all directions and the disturbance amplitude is in a negligible range.

[0004] However, in the complex operating environment of densely built-up urban areas, building clusters are prone to triggering funneling effects and wake vortex phenomena, resulting in significant spatial non-uniform distribution and abrupt changes in wind direction and speed in local wind fields. This leads to numerous technical defects in traditional uniform expansion flight corridor planning schemes, making them unsuitable for the safe flight requirements of UAVs in complex urban wind fields. The specific defects are as follows: First, the ability to resist crosswind disturbances is insufficient, and the fault tolerance rate for flight safety is low: When the drone flies at low altitude close to the building facade, a sudden strong crosswind will cause the drone to deviate instantaneously. Moreover, this deviation process occurs before the underlying flight control PID adjustment mechanism takes effect. The traditional uniform expansion flight corridor does not reserve redundant buffer space for lateral disturbances. The lateral deviation of the drone is very likely to break through the safety boundary of the flight corridor, causing the flight risk of colliding with obstacles.

[0005] Secondly, there is a lack of a mechanism for advance perception and adaptation of wind field information: the existing path planning architecture does not establish an advance feedback link from wind field data to the flight corridor construction layer. The planning system cannot obtain dynamic parameters such as wind field direction and wind speed intensity of the flight passage ahead in advance, and can only rely on fixed empirical values ​​to set safety margins. If the safety margin is set too conservatively, it will greatly compress the passable space, causing a large number of narrow passages that meet the flight conditions to be misjudged as no-fly zones, reducing the efficiency of drone flight passage and scene adaptability. If the safety margin is set too aggressively, it will be unable to cope with sudden wind field disturbances, greatly reducing flight fault tolerance.

[0006] Third, it is impossible to achieve a coupling and adaptation between the geometric characteristics of the flight path and the wind field characteristics. The geometric scales of different flight paths in a city vary significantly. The same incoming flow from a narrow passage between buildings and an open urban area will generate wind field disturbances with extremely different amplitudes after being amplified by the funnel effect. However, traditional planning schemes use uniform fixed safety margin parameters, without taking into account the different wind field disturbance characteristics such as the width and opening / closing of the flight path. This makes it impossible to achieve dynamic adaptive adjustment of the safety margin and to balance the safety of UAV flight with the utilization rate of environmental traffic. Summary of the Invention

[0007] In view of this, the present invention proposes an intelligent path planning system for unmanned aerial vehicles (UAVs) that can generate a time- and aerodynamically tolerant corridors with a safety margin, so as to plan a flight path that is fully connected and will not cause collisions.

[0008] The technical solution of this invention is implemented as follows: An intelligent path planning system for unmanned aerial vehicles (UAVs) includes: The state wind field estimation module is used to calculate the local three-dimensional wind force vector based on the difference between the commanded rotational speed issued by the flight controller and the actual rotational speed fed back by the electronic speed controller; The environment map building module is used to convert the raw point cloud collected by the depth camera into a collision-free voxel map of the environment, and to build a convex polyhedron corridor from the current position to the target point in the collision-free voxel map of the environment. The aerodynamic disturbance assessment module is used to assess the body offset of the UAV when subjected to aerodynamic disturbances based on local three-dimensional wind vectors. The Tolerant Corridor Generation Module is used to modify convex polyhedral corridors into asymmetric spatiotemporal tolerant corridors based on offsets. The trajectory planning and control module is used for UAV path planning based on a time-space dynamic tolerance corridor. The state wind field estimation module is connected to the aerodynamic disturbance assessment module, and the tolerance corridor generation module is connected to the environmental map construction module, the aerodynamic disturbance assessment module, and the trajectory planning and control module, respectively.

[0009] Preferably, the execution steps of the state wind field estimation module are as follows: An extended Kalman filter is constructed, which takes the triaxial acceleration and triaxial angular velocity output by the IMU as process inputs and the position and velocity of GPS / visual positioning as measurement updates. A rotation matrix is ​​constructed based on the attitude quaternions output by the extended Kalman filter. The theoretical acceleration and the actual acceleration are then calculated based on the rotation matrix.

[0010] in For theoretical acceleration, Let be a rotation matrix. For thrust coefficient, For the quality of drones, The first flight control system issued The commanded speed of each motor, The vector of gravitational acceleration. For actual acceleration, For the first ESC feedback The actual speed of each motor; The control residuals are calculated based on theoretical and actual accelerations. Local three-dimensional wind vectors are then calculated based on the control residuals and the aerodynamic drag experienced by the UAV during its motion.

[0011] in This is a local three-dimensional wind force vector. To control residuals, This is a diagonal drag coefficient matrix. To extend the world coordinate system linear velocity output by the Kalman filter.

[0012] Preferably, the state update formula for the extended Kalman filter is:

[0013] in For the posterior state estimate at the current time, It is a nonlinear state transition function. This is an estimate of the posterior state from the previous time step. and These are the triaxial acceleration and triaxial angular velocity data acquired by the IMU, respectively. For Kalman gain, For measurement vectors, For observation functions.

[0014] Preferably, the execution steps of the environment map construction module are as follows: The raw point cloud acquired by the depth camera is converted to the world coordinate system of the three-dimensional space where the UAV is located, and the point coordinates in the converted world coordinate system are obtained. The three-dimensional space is divided into several cubic voxels. Each cubic voxel maintains a Log-odds occupancy probability. Based on the occupancy probability, obstacle voxels and free space voxels are labeled, and finally a binary environment collision-free voxel map is formed. Starting from the UAV's current position and ending at the next target in the global waypoint, the A* algorithm is used to search for a discrete path skeleton in the free space voxels of the collision-free voxel map of the environment. The node evaluation cost is:

[0015] in The first extension of the A* algorithm search 1 path node For nodes The overall cost of the comprehensive assessment From the starting point to the node The known cost, As weight, For nodes The corresponding point coordinates in the world coordinate system of three-dimensional space. Here are the coordinates of the endpoint of this segmented flight in the world coordinate system. It is the vector norm 2; Expanding segmentally along the path skeleton along the XYZ axes until touching the obstacle object, generating multiple sets of cuboid bounding boxes, which are then converted into convex polyhedral corridors formed by the intersection linear inequality constraints of the half-space:

[0016] in For a set of convex polyhedral corridors, The total number of convex polyhedra. The sequence number of the convex polyhedron is used for segmentation. For the first Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix is ​​formed by the normal vectors of the hyperplane of the segmental convex polyhedron, where each row of the matrix corresponds to an external normal vector of the constraining hyperplane. For the first The distance intercept corresponding to the hyperplane of the segmental convex polyhedron.

[0017] Preferably, the transformation formula for converting the raw point cloud acquired by the depth camera to point coordinates in the world coordinate system is:

[0018] in To transform to the world coordinate system The coordinates of the points Let be the rotation matrix from the body coordinate system to the world coordinate system. Let be the extrinsic parameter matrix of the depth camera relative to the machine body. In the camera coordinate system, the first The coordinates of the points Let be the extrinsic translation vector of the depth camera relative to the object. The world coordinates of the drone body.

[0019] Preferably, the execution steps of the aerodynamic disturbance assessment module are as follows: Traverse the hyperplane of each convex polyhedron in the corridor and find the nearest obstacle surface based on the spatial position of the hyperplane in the collision-free voxel map of the environment. Principal component analysis is used to extract the normal vectors of the adjacent obstacle surfaces, and the channel width of the corresponding convex polyhedron at the current cross-section is calculated.

[0020] in The sequence number of the segmented convex polyhedron. For the first The channel width of the cross-section of the segmental convex polyhedron. , The first The coordinates of the feature sampling points of the obstacles on the left and right sides of the cross section of the convex polyhedron. Let be the normal vector of the obstacle surface. For the first The local surface gradient of an obstacle at the cross section of a segmental convex polyhedron. Let L be the vector L2 norm of the local surface gradient; The local three-dimensional wind force vector is projected from the normal vector of the obstacle surface and divided into normal wind force and tangential wind force:

[0021] in For the first The cross-section of the convex polyhedron is perpendicular to the normal wind force on the wall surface. This is a local three-dimensional wind force vector. For the first The cross section of the segmental convex polyhedron is parallel to the tangential wind force along the corridor direction of the convex polyhedron; The offset at the cross section is calculated based on the channel width, normal wind force, and the slit effect:

[0022] in For the first drone The displacement of the body at the cross section of the segmental convex polyhedron when subjected to aerodynamic disturbance. This is the empirical coefficient for amplification of the narrowing effect. For reference width, For the quality of drones, This refers to the delay time from when the flight control system detects a deviation to when it outputs effective compensating thrust.

[0023] Preferably, the execution steps of the tolerance corridor generation module are as follows: Calculate the maximum dynamic shrinkage margin and predict the unit vector of the drift direction based on the offset:

[0024] in For the first The maximum dynamic shrinkage allowance of the cross section of a segmental convex polyhedron. For the first The unit vector of the predicted drift direction of the cross section of a segmental convex polyhedron. Represents the 2-norm of a vector; Differential shrinkage update of the hyperplane intercept based on the maximum dynamic shrinkage margin and the predicted drift direction unit vector:

[0025] in For the first The hyperplane number in a segmented convex polyhedron and The first The first segment of a convex polyhedron The original intercept and the intercept updated by differential shrinkage of each hyperplane This is the static foundation safety margin. For the first The first segment of a convex polyhedron Transpose of the normal vectors of the hyperplane; Based on the hyperplane intercept updated by differential shrinkage, a half-space intersection operation is performed. After removing redundant constraint surfaces, a new set of convex polyhedra is obtained, and the output is a spatiotemporal dynamic tolerance corridor.

[0026] Preferably, when the wind field is spatially uniform, the expression for the spatiotemporal dynamic tolerance corridor is:

[0027] in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. For the first after differential shrinkage Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where .

[0028] Preferably, when the asymmetric contraction directions of two adjacent convex polyhedra differ significantly, the step of obtaining the spatiotemporal dynamic tolerance corridor is as follows: Calculate the projected area of ​​the interface between two adjacent convex polyhedra:

[0029] in For the first Segmented convex polyhedron and the first +1 Projected area of ​​the interface of a convex polyhedron This is a function for calculating the area of ​​polygons within the projection plane. This is a projection operator that projects a 3D feasible region onto a plane orthogonal to the normal vectors of the interface between two convex polyhedra. For the intersection operation of sets, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where , For the first +1 coefficient matrix formed by the normal vectors of the hyperplane of the convex polyhedron For the first +1 hyperplane intercept set updated by hyperplane differential shrinkage of convex polyhedron; Will With preset threshold In comparison, if When the interface is marked as having degraded connectivity, for each interface marked as having degraded connectivity, a hyperplane within the nearest influence domain is found, the intercept is extrapolated slightly, and the relaxation is iteratively performed until the projected area of ​​the interface is not less than a preset threshold. The relaxation iteration formula is as follows:

[0030] in For the first The first segment of a convex polyhedron The intercept after each hyperplane relaxation iteration For the first The first segment of a convex polyhedron The relaxation direction coefficients of each hyperplane. The relaxation step size for a single iteration. For the first The set of intercepts after hyperplane relaxation iteration of a segmental convex polyhedron. The interface projected area is recalculated using the intercept set after relaxation iteration; Integrate all updated convex polyhedra to output a fully connected, anisotropic spatiotemporally tolerant corridor:

[0031] in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. The first after differential shrinkage and relaxation correction Segmentally convex polyhedron, To satisfy the area constraint for all adjacent convex polyhedra.

[0032] Preferably, the execution steps of the trajectory planning and control module are as follows: Constructing spatial linear constraints for the corridor: Each set of linear inequalities in the spatiotemporally dynamic tolerance corridor is transformed into trajectory control point constraints. The entire trajectory does not exceed the bounds only if all piecewise polynomial control points fall within the convex polyhedron. The formula for the spatial linear constraint is:

[0033] in For the first Coordinates of control points To be constrained in the first The set of control point indices within a segmented convex polyhedron; Construct thrust acceleration constraints and jerk constraints, and build a cost function based on jerk integral and flight time weighted penalty: Thrust acceleration constraints:

[0034] jerk constraints:

[0035] Cost function:

[0036] in Let be the second derivative of the trajectory position at time t. The vector of gravitational acceleration. This represents the maximum total thrust of the drone when all its motors are fully engaged. For the quality of drones, The third derivative of the trajectory position at time t. To increase the upper limit of jerk, For the total cost, For the first The total flight time of the trajectory of the segmental convex polyhedron. This refers to the flight time penalty weighting coefficient. The spatial linear constraints, thrust acceleration constraints, jerk constraints, and cost function are combined to form a large-scale sparse quadratic programming problem. The interior-point method is used to solve the sparse quadratic programming problem, and the flight trajectory is obtained. The flight trajectory is packaged and sent to the flight control system. The formula for evaluating the flight trajectory in the time domain is:

[0037] in Let be the optimal trajectory position vector at time t. For the order of the polynomial, To optimize the solution of the th The first segment of the convex polyhedron trajectory Optimal coefficients of the order polynomial For the first The effective time interval of the trajectory of a segmental convex polyhedron.

[0038] Compared with the prior art, the beneficial effects of the present invention are: ① Only the original onboard sensors of the UAV are reused to complete the three-dimensional local wind field calculation, without the need to add new wind speed and wind direction detection hardware, effectively controlling the UAV's payload, hardware cost and operating power consumption; ② The real-time wind field direction and intensity are coupled with the geometric features of the forward channel to quantify the wind amplification effect caused by the narrow tube effect and accurately predict the body offset caused by wind disturbance. ③ Differentiated adaptive allocation of safety margins in each direction of the flight corridor, with sufficient buffer space reserved in the windward and other high collision risk directions to resist sudden crosswinds, and the redundancy margin reduced on the leeward safety side, to maximize the release of flight-feasible airspace in narrow building passages. This solves the pain points of traditional uniform expansion corridors being prone to collisions or being overly conservative in prohibiting flights, while taking into account both the safety of urban low-altitude flights and the efficiency of route traffic. ④ It provides two corridor generation implementation methods. In complex scenarios, it can automatically detect and eliminate corridor connection bottlenecks to ensure full airspace connectivity. Combined with rolling time domain MPC trajectory optimization, it can achieve dynamic adaptive flight and adapt to various UAV autonomous operation scenarios such as urban logistics, building inspection, and emergency rescue. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only preferred embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 This is a schematic diagram of an intelligent path planning system for unmanned aerial vehicles (UAVs) according to the present invention. In the diagram, 1 is the state wind field estimation module; 2 is the environmental map construction module; 3 is the aerodynamic disturbance assessment module; 4 is the tolerance corridor generation module; and 5 is the trajectory planning and control module. Detailed Implementation

[0041] To better understand the technical content of this invention, a specific embodiment is provided below, and the invention will be further described in conjunction with the accompanying drawings.

[0042] See Figure 1 The present invention provides an intelligent path planning system for unmanned aerial vehicles (UAVs), comprising: State wind field estimation module 1 is used to calculate the local three-dimensional wind force vector based on the difference between the commanded rotational speed issued by the flight controller and the actual rotational speed fed back by the electronic speed controller. The execution steps are as follows: An extended Kalman filter is constructed, which uses the triaxial acceleration and triaxial angular velocity output from the IMU as process inputs and the position and velocity from GPS / visual positioning as measurement updates. The state transition of the extended Kalman filter uses standard kinematic equations. Each cycle first performs time-recursive prediction, then corrects using measurement data. The state update formula is as follows:

[0043] in For the posterior state estimate at the current time, It is a nonlinear state transition function. This is an estimate of the posterior state from the previous time step. and These are the triaxial acceleration and triaxial angular velocity data acquired by the IMU, respectively. For Kalman gain, For measurement vectors, The state vector of the filtered output is the observation function, which contains three-dimensional position, three-dimensional linear velocity, and attitude quaternions. A rotation matrix is ​​constructed based on the attitude quaternions output by the extended Kalman filter. The theoretical acceleration and the actual acceleration are then calculated based on the rotation matrix.

[0044] in For theoretical acceleration, Let be a rotation matrix. The thrust coefficient is obtained from calibration on a ground test bench; a 5-inch propeller typically has a thrust coefficient of... arrive At this level, For the quality of drones, The first flight control system issued The commanded speed of each motor, The vector of gravitational acceleration. m / s², For actual acceleration, For the first ESC feedback The actual speed of each motor; The control residuals are calculated based on theoretical and actual accelerations. Local three-dimensional wind vectors are then calculated based on the control residuals and the aerodynamic drag experienced by the UAV during its motion.

[0045] in This is a local three-dimensional wind force vector. To control residuals, The diagonal drag coefficient matrix (a pre-calibrated constant, typically 0.1~0.3 N·s / m for each axis of a quadcopter). To extend the world coordinate system linear velocity output by the Kalman filter.

[0046] The core problem that the state wind field estimation module 1 needs to solve is: to simultaneously obtain the precise motion state of the UAV and the three-dimensional wind information of its current location without installing an onboard anemometer. The general idea is as follows: the flight controller issues a commanded rotational speed, and the electronic speed controller (ESC) also provides the actual rotational speed. If there is no wind, the thrust acceleration generated by both should be the same. However, in actual flight, there is a difference. After deducting the aerodynamic drag generated by the UAV's own motion, the remaining difference is caused by the wind. Specifically, this involves four steps: the first step is to perform state calculation using an extended Kalman filter; the second step is to calculate the theoretical acceleration; after constructing the rotation matrix using the attitude quaternions output by the extended Kalman filter, the theoretical rotational speed commands of the four motors issued by the flight controller are converted into the theoretical resultant speed in the Z-axis direction of the UAV using thrust coefficients. The first step is to calculate the theoretical acceleration by rotating to the world coordinate system and subtracting gravity. The second step is to calculate the control residual. The actual acceleration is calculated in the same way as the theoretical acceleration. The actual acceleration can be calculated by using the actual rotation speed of the electronic speed controller feedback as input. The control residual is obtained by subtracting the theoretical acceleration from the actual acceleration. The control residual is the part of acceleration that the flight controller outputs to counteract external disturbances. If the actual rotation speed is greater than the commanded rotation speed, it means that the flight controller is trying to compensate for the external force in a certain direction. Finally, there is the wind inversion process. Multiplying the control residual by the mass can obtain the equivalent external force. However, the equivalent external force is mixed with the aerodynamic drag that the UAV experiences during its own motion. Therefore, after deducting this part of the aerodynamic drag, the local three-dimensional wind vector can be finally output.

[0047] Environment map construction module 2 is used to convert the raw point cloud acquired by the depth camera into a collision-free voxel map of the environment, and to construct a convex polyhedron corridor from the current location to the target point in the collision-free voxel map of the environment. The execution steps are as follows: The raw point cloud acquired by the depth camera is transformed into the world coordinate system of the 3D space where the UAV is located, and the point coordinates in the transformed world coordinate system are obtained. The transformation formula is:

[0048] in To transform to the world coordinate system The coordinates of the points Let be the rotation matrix from the body coordinate system to the world coordinate system. Let be the extrinsic parameter matrix of the depth camera relative to the machine body. In the camera coordinate system, the first The coordinates of the points Let be the extrinsic translation vector of the depth camera relative to the object. The world coordinates of the drone body.

[0049] After coordinate transformation, voxelization is performed using an octree, dividing the 3D space into several cubic voxels (voxel side length is generally 0.1~0.3m). Each cubic voxel maintains the Log-odds occupancy probability. When a cloud falls in, the probability is updated. If the probability is greater than 0.7, it is marked as an obstacle voxel. If it is less than 0.3, it is marked as a free space voxel. Finally, a binary environment collision-free voxel map is formed. Then, a front-end path search is performed, starting from the UAV's current position and ending at the next target in the global waypoint. The A* algorithm is used to search for a discrete path skeleton in the free space voxels of the collision-free voxel map, where the node evaluation cost is:

[0050] in The first extension of the A* algorithm search 1 path node For nodes The overall cost of the comprehensive assessment From the starting point to the node The known cost, The weights are set to 1.0 for standard A* to guarantee optimality, while 1.5 can speed up the search but the path may be slightly off. For nodes The corresponding point coordinates in the world coordinate system of three-dimensional space. Here are the coordinates of the endpoint of this segmented flight in the world coordinate system. Given the vector norm 2, the final search result is a path skeleton composed of a series of discrete waypoints; Expanding segmentally along the path skeleton along the XYZ axes until touching the obstacle object, generating multiple sets of cuboid bounding boxes, which are then converted into convex polyhedral corridors formed by the intersection linear inequality constraints of the half-space:

[0051] in For a set of convex polyhedral corridors, The total number of convex polyhedra. The sequence number of the convex polyhedron is used for segmentation. For the first Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix is ​​formed by the normal vectors of the hyperplane of the segmental convex polyhedron, where each row of the matrix corresponds to an external normal vector of the constraining hyperplane. For the first The distance intercept corresponding to the hyperplane of the segmental convex polyhedron.

[0052] The environment map construction module 2 is responsible for converting the raw point cloud of the depth camera into a structured 3D map, and finding a safe passage from the current position to the target point in the map, namely the convex polyhedron corridor. Subsequent trajectory optimization requires the corridor to be given constraints in the form of existing inequalities.

[0053] Aerodynamic disturbance assessment module 3 is used to assess the displacement of the UAV when subjected to aerodynamic disturbances based on local three-dimensional wind vectors. The execution steps are as follows: Traverse the hyperplane of each convex polyhedron in the corridor and find the nearest obstacle surface based on the spatial position of the hyperplane in the collision-free voxel map of the environment. Principal component analysis is used to extract the normal vectors of the surfaces of nearby obstacles (pointing towards the free space inside the corridor), and the channel width of the corresponding convex polyhedron at the current cross-section is calculated.

[0054] in The sequence number of the segmented convex polyhedron. For the first The channel width of the cross-section of the segmental convex polyhedron. , The first The coordinates of the feature sampling points of the obstacles on the left and right sides of the cross section of the convex polyhedron. Let be the normal vector of the obstacle surface. For the first The local surface gradient of an obstacle at the cross section of a segmental convex polyhedron. Given the vector L2 norm of the local surface gradient, pack the normal vectors of all cross sections and the channel width into a corridor geometric feature matrix; The local three-dimensional wind force vector is projected from the normal vector of the obstacle surface and divided into normal wind force and tangential wind force:

[0055] in For the first The normal wind force perpendicular to the wall surface of the cross-section of the convex polyhedron, i.e., the component that pushes the drone against the wall, is the main threat. This is a local three-dimensional wind force vector. For the first The tangential wind force parallel to the direction of the convex polyhedron's cross section affects the speed along the corridor direction, thus the risk of collision is relatively small. The wind speed amplification factor is estimated using the ratio of the channel width to the reference width (the narrower the channel, the greater the amplification). Then, the amplified wind speed is integraled twice over the flight control response delay time window to obtain the offset.

[0056] in For the first drone The displacement of the body at the cross section of the segmental convex polyhedron when subjected to aerodynamic disturbance. For the amplification of the slit effect, the empirical coefficient is 1.5 for regular square building passages and 1.8~2.0 for irregular narrow slits; The width is a reference value, representing the equivalent channel size when the wind field is unconstrained in an open area, and is generally set to 20~30m. For the quality of drones, To account for the delay time from the detection of deviation to the output of effective compensation thrust in flight control, PID controllers typically require 0.2 to 0.5 seconds, while nonlinear controllers such as INDI can achieve 0.1 to 0.2 seconds.

[0057] The aerodynamic disturbance assessment module 3 is used to calculate the offset of the UAV before the underlying flight control system can react when the UAV is blown by the wind at a certain section of the corridor. This offset is the basis for the subsequent internal measurement of the corridor. Finally, the offset vectors of all sections are organized into a high-dimensional tensor structure according to the spatial index, which is called the corridor aerodynamic disturbance boundary tensor. This tensor records the maximum distance that the UAV may be blown away by the wind at each position and in each direction of the corridor.

[0058] The tolerance corridor generation module 4 is used to modify convex polyhedral corridors into asymmetric spatiotemporal dynamic tolerance corridors based on offsets. The execution steps are as follows: The offset information of each section is extracted from the tensor. The maximum dynamic contraction margin is obtained by taking the modulus of the maximum sideslip yaw prediction displacement vector of each section, and then normalized to obtain the unit vector of the predicted drift direction.

[0059] in For the first The maximum dynamic shrinkage allowance of the cross section of a segmental convex polyhedron. For the first The unit vector of the predicted drift direction of the cross section of a segmental convex polyhedron. To represent the vector L2 norm, a mapping table is built that associates the maximum dynamic shrinkage margin, the predicted drift direction unit vector, and the convex polyhedron. Then, each hyperplane in the corridor is judged and translated. The judgment logic is: calculate the hyperplane normal vector. With drift direction The dot product is calculated as follows: a dot product greater than zero indicates that the surface is in the drift direction (the side the drone will collide with after being blown by the wind), requiring more shrinkage; a dot product less than or equal to zero indicates that it is on the leeward side or in an orthogonal direction, requiring only the basic margin to be shrunk. The differential shrinkage update formula for the intercept is:

[0060] in For the first The hyperplane number in a segmented convex polyhedron and The first The first segment of a convex polyhedron The original intercept and the intercept updated by differential shrinkage of each hyperplane This is a static safety margin (mainly covering the body radius and positioning error; for a 450mm wheelbase UAV, it is generally taken as about 0.4m). For the first The first segment of a convex polyhedron Transpose of the normal vectors of the hyperplane; After updating the intercepts of all hyperplanes, a half-space intersection operation is performed based on the hyperplane intercepts updated by differential shrinkage. After removing redundant constraint surfaces, a new set of convex polyhedra is obtained, and the output is a spatiotemporal dynamic tolerance corridor. This process can be implemented in two ways: Method 1: Directly generate independent convex polyhedra by intersecting the updated sets of inequalities. Adjacent polyhedra are directly stitched together in the spatial overlap region. This method is computationally fast and suitable for scenarios with relatively uniform wind field spatial distribution. The expression for the spatiotemporal dynamic tolerance corridor is:

[0061] in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. For the first after differential shrinkage Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where .

[0062] If the asymmetric contraction directions of two adjacent convex polyhedra differ greatly (for example, the first segment contracts to the left and the second segment contracts to the right), the overlapping area of ​​their interface will degenerate into a very narrow region, which may even be so small that the drone cannot fly over it. The subsequent trajectory optimizer will not be able to find a feasible solution at this bottleneck position.

[0063] Method 2: Based on Method 1, two additional processing steps are added to address the junction bottleneck problem. The first step is junction degradation detection: calculating the projected area of ​​the interface between two adjacent convex polyhedra.

[0064] in For the first Segmented convex polyhedron and the first +1 Projected area of ​​the interface of a convex polyhedron This is a function for calculating the area of ​​polygons within the projection plane. This is a projection operator that projects a 3D feasible region onto a plane orthogonal to the normal vectors of the interface between two convex polyhedra. For the intersection operation of sets, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where , For the first +1 coefficient matrix formed by the normal vectors of the hyperplane of the convex polyhedron For the first +1 hyperplane intercept set updated by hyperplane differential shrinkage of convex polyhedron; Will With preset threshold (For a 450mm wheelbase drone, take 0.8~1.2 m²) In comparison, if... At that time, the interface is marked as having degraded connectivity.

[0065] The second step is adaptive relaxation: For interfaces marked as having degraded connectivity, find hyperplanes within the nearby influence domains, extrapolate the intercepts slightly, and iterate until the projected area of ​​the interface is not less than a preset threshold. The relaxation iteration formula is as follows:

[0066] in For the first The first segment of a convex polyhedron The intercept after each hyperplane relaxation iteration For the first The first segment of a convex polyhedron The relaxation direction coefficients of each hyperplane are positive only for hyperplanes within the influence domain, and remain zero for hyperplanes outside the domain. The relaxation step size for a single iteration. For the first The set of intercepts after hyperplane relaxation iteration of a segmental convex polyhedron. The interface projected area is recalculated using the intercept set after relaxation iteration. The relaxation iteration ensures that the minimum concession is made only at the most vulnerable interface, without affecting the wind deflection resistance of other areas.

[0067] Integrate all updated convex polyhedra to output a fully connected, anisotropic spatiotemporally tolerant corridor:

[0068] in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. The first after differential shrinkage and relaxation correction Segmentally convex polyhedron, To satisfy the area constraint for all adjacent convex polyhedra.

[0069] Since urban building passageways are inherently narrow, if the basic convex polyhedral corridor contracts equally in all directions to ensure a safety margin on the windward side, not only will the leeward side waste space, but the corridor may also become so narrow that drones cannot fly through. The tolerance corridor generation module 4, based on the offset, contracts the basic convex polyhedral corridor more in the dangerous direction and less in the safe direction, thereby forming an asymmetrical safe flight corridor, i.e., an aerodynamic tolerance corridor. Through asymmetrical contraction, it can maximize the use of space while ensuring safety.

[0070] The trajectory planning and control module 5 is used for UAV path planning based on a time-space aerodynamic tolerance corridor. The calculated trajectory must simultaneously meet three conditions: it cannot fly out of the corridor boundary, it cannot exceed the physical limits of motor thrust and angular acceleration, and the higher-order derivatives must be continuous (otherwise the flight control cannot keep up). The method used is piecewise polynomial trajectory optimization under the model predictive control (MPC) framework. The execution steps are as follows: Constructing linear constraints for corridor space: Spatiotemporal and aerodynamic tolerance. Each convex polyhedron of the corridor is a set of linear inequalities. The trajectory is represented by a piecewise polynomial, which has a useful property: the polynomial curve lies within the convex hull. Therefore, as long as all control points are within the convex polyhedron, the entire curve will not exceed the boundary. Each set of linear inequalities in the spatiotemporal dynamic tolerance corridor is transformed into trajectory control point constraints. The entire trajectory will not exceed the boundary only if all piecewise polynomial control points fall within the convex polyhedron. The formula for the spatial linear constraint is:

[0071] in For the first Coordinates of control points To be constrained in the first The set of control point indices within a segmented convex polyhedron; Construct thrust acceleration constraints and jerk constraints, and build a cost function based on jerk integral and flight time weighted penalty: The quadrotor exhibits the differential flatness property, meaning all states and control inputs can be uniquely determined by the position trajectory and its derivatives. Utilizing this property, the thrust limit can be equivalently transformed into an acceleration amplitude constraint. The thrust-acceleration constraint is as follows:

[0072] The angular acceleration limit is equivalent to a jerk constraint:

[0073] The cost function is designed as a weighted sum of the squared integral of jerk (control energy) plus a time penalty:

[0074] in Let be the second derivative of the trajectory position at time t. The vector of gravitational acceleration. This is the maximum total thrust of the drone when all four motors are running (approximately 25-30N for a 1.5kg quadcopter). For the required total thrust acceleration, For the quality of drones, The third derivative of the trajectory position at time t. To set the upper limit for jerk, a typical quadcopter uses 20-40 m / s³. Exceeding this value would mean the motor would need to drastically change its speed instantaneously, which is physically impossible. The total cost is calculated as follows: the first term in the cost function aims to make the trajectory as smooth as possible (saving power and minimizing motor wear), and the second term aims to minimize flight time. For the first The total flight time of the trajectory of the segmental convex polyhedron is a questionable optimization variable. The flight time penalty weighting coefficient is set to around 200 for better balance (too high a value results in fast but aggressive flight paths, while too low a value results in smooth but sluggish flight paths). The initial conditions are taken from the UAV dynamic state data output by the extended Kalman filter: current position ,speed acceleration Boundary conditions serving as the starting point of the trajectory.

[0075] The spatial linear constraints, thrust acceleration constraints, jerk constraints, and cost function constitute a large-scale sparse quadratic programming problem (QP). The interior-point method is used to solve the sparse quadratic programming problem, and the flight trajectory is obtained. The flight trajectory is packaged and sent to the flight control system. The formula for evaluating the flight trajectory in the time domain is:

[0076] in Let be the optimal trajectory position vector at time t. The polynomial order is determined by a fifth-order polynomial, ensuring continuity from position to the third derivative, thus satisfying the requirements of quadrotor dynamics for continuity of acceleration and jerk. To optimize the solution of the th The first segment of the convex polyhedron trajectory Optimal coefficients of the order polynomial For the first The effective time interval of the trajectory of a segmental convex polyhedron.

[0077] After the solution is completed, the trajectory is packaged and sent to the flight controller. The flight controller interpolates the position, velocity, and acceleration from the trajectory at a frequency of 50Hz as feedforward commands for tracking. Because the trajectory itself is constrained within the aerodynamic tolerance corridor, and the windward side of the corridor has reserved sufficient buffer distance, even if it is blown off course by crosswinds during flight, it will not collide with anything. At the same time, the trajectory satisfies thrust and jerk constraints, and there will be no motor saturation during flight control tracking. The entire MPC is executed in a rolling manner. Each control cycle is reinitialized and resolved with the latest extended Kalman filter state. In this way, the trajectory can be adjusted accordingly when the wind field changes or the corridor is updated.

[0078] The effectiveness of this invention is illustrated by the following examples: A 1.5kg quadcopter, Calibrated as The drag coefficient is set to 0.15 N·s / m for each axis. At a certain moment, the extended Kalman filter calculates an X-axis velocity of 2.0 m / s, with an approximately horizontal attitude. The flight control command speeds for all four motors are 850 rad / s, while the actual speeds fed back by the electronic speed controllers are 862, 858, 865, and 860 rad / s, respectively. Calculated using the formula, the theoretical thrust acceleration on the Z-axis is approximately 9.81 m / s² (just enough to balance gravity), while the actual thrust acceleration on the Z-axis is approximately 10.05 m / s². Due to attitude fine-tuning, the X-axis thrust acceleration is approximately 0.3 m / s² higher. The control residual on the X-axis is 0.3 m / s², multiplied by the mass of 1.5 to obtain 0.45 N. Subtracting the X-axis drag of 0.15 × 2.0 = 0.30 N, the final X-axis wind force is approximately 0.15 N. This value is then passed to the subsequent aerodynamic disturbance assessment module 3.

[0079] The drone then flies over a street approximately 8 meters wide, with voxel sides of 0.2m. The depth camera captures about 30,000 points per frame, which are then transformed and updated using voxelization. Voxel grids on the walls on both sides of the street are marked as obstacles, while the central airspace is free space. The current position is [10.0, 4.0, 15.0]m, and the target point is [50.0, 4.0, 15.0]m. An A* search is used to find a path along the street's central axis. Expansion centers are taken at approximately 5m intervals, each expanding until it touches the walls on both sides, resulting in a bounding box 3-4m wide and approximately 3m high. This ultimately forms about 8 convex polyhedra, overlapping approximately 1m between adjacent ones, covering a flight distance of 40m.

[0080] The distance between the two buildings is 6m. ),set up Wind inversion results After decomposition, the normal wind force of this section The value is 1.8N (pointing to the building on the left), and the drone weighs 1.5kg. Substituting into the formula: magnification factor equivalent acceleration Predicted displacement In other words, if the flight controller fails to correct the error within 0.3 seconds, the drone will deviate approximately 0.34m to the left of the building. This value is written into the tensor, so the tolerance corridor generation module 4 knows that the left side of this section needs to be shrunk by an additional 0.34m.

[0081] Suppose there is a corridor with 5 segments. The offset vector of the 3rd segment is [0.34, 0.12, 0.0]m, modulo approximately 0.36m, and after normalization, the drift direction is approximately [0.94, 0.33, 0.0]. The normal vector of a certain hyperplane of this segment is [1,0,0] (the wall in the positive X direction). The dot product of this vector and the drift direction is 0.94 > 0, indicating it is on the windward side. =0.4m, total shrinkage The dot product of the surface normal vector in the opposite direction [-1,0,0] is -0.94 < 0, resulting in a shrinkage of only 0.4m. This causes an additional 0.34m shrinkage on the windward side, which just covers the wind deflection.

[0082] Five corridors cover a distance of 40m. The drone is currently located at [10.0, 4.0, 15.0]m, with a velocity of [2.0, 0.0, 0.0]m / s and zero acceleration. Its mass is... The interior-point method yields a solution in approximately 15ms, with time intervals of 1.2s, 1.5s, 1.8s, 1.4s, and 1.1s, for a total flight time of about 7 seconds. In the third segment (the narrow passage between two buildings), because the windward side of the corridor has been reduced by 0.74m, the optimized trajectory automatically shifts towards the leeward side, maintaining a distance of approximately 1.1m from obstacles on the windward side. Even if it is blown off course by 0.34m, there is still a 0.76m margin, ensuring safety.

[0083] Regarding the execution frequency and data flow timing of each module: the state wind field estimation module 1 runs on the IMU sampling frequency (usually 200~400Hz), and each time the extended Kalman filter is updated, the latest state and wind force estimation are output; the environment map construction module 2 is synchronized with the depth camera frame rate (usually 15~30Hz), and each time a point cloud frame is received, the voxel map is updated and the corridor is re-evaluated to see if reconstruction is needed; the aerodynamic disturbance assessment module 3 and the tolerance corridor generation module 4 are triggered when the corridor is updated or the wind force change exceeds the threshold; the trajectory planning and control module 5 is executed in a rolling manner with an MPC cycle (usually 20~50Hz).

[0084] Regarding the system's input and output interfaces: The total inputs include IMU data stream (three-axis acceleration + three-axis angular velocity), motor speed data (commanded speed + actual speed), depth camera point cloud frames, GPS / visual positioning data, and preset global waypoint sequences; the total output is the position-velocity-acceleration feedforward trajectory command stream sent to the flight controller.

[0085] Regarding parameter calibration: (Thrust coefficient) and (Drag coefficient matrix) needs to be pre-calibrated on a ground test bench; and (Camera external parameters) are calibrated at the factory or using a calibration plate; (Vascular effect coefficient) (Reference width) (Control delay) (Basic safety margin) and other empirical parameters are adjusted and determined based on actual flight tests.

[0086] Regarding hardware platform compatibility: All parameter values ​​in the above solution are based on a 450mm wheelbase and a 1.5kg quadcopter. If a larger or smaller platform is used, These parameters need to be recalibrated or set according to the actual platform. The overall algorithm architecture remains the same; only the parameter values ​​are different.

[0087] Regarding computational resource requirements: The extended Kalman filter and wind inversion computations are relatively small and can be performed by a flight controller of ARM Cortex-M7 level. Voxel map updates and A* search require an ARM Cortex-A level or an accompanying computing board (such as the Jetson series). MPC solving is the most computationally intensive part. The QP problem of 5 corridors can be solved in about 10~15ms on Jetson Orin NX using the OSQP solver, meeting the 50Hz rolling cycle requirement.

[0088] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent path planning system for unmanned aerial vehicles (UAVs), characterized in that, include: The state wind field estimation module is used to calculate the local three-dimensional wind force vector based on the difference between the commanded rotational speed issued by the flight controller and the actual rotational speed fed back by the electronic speed controller; The environment map building module is used to convert the raw point cloud collected by the depth camera into a collision-free voxel map of the environment, and to build a convex polyhedron corridor from the current position to the target point in the collision-free voxel map of the environment. The aerodynamic disturbance assessment module is used to assess the body offset of the UAV when subjected to aerodynamic disturbances based on local three-dimensional wind vectors. The Tolerant Corridor Generation Module is used to modify convex polyhedral corridors into asymmetric spatiotemporal tolerant corridors based on offsets. The trajectory planning and control module is used for UAV path planning based on a time-space dynamic tolerance corridor. The state wind field estimation module is connected to the aerodynamic disturbance assessment module, and the tolerance corridor generation module is connected to the environmental map construction module, the aerodynamic disturbance assessment module, and the trajectory planning and control module, respectively.

2. The intelligent path planning system for unmanned aerial vehicles according to claim 1, characterized in that, The execution steps of the state wind field estimation module are as follows: An extended Kalman filter is constructed, which takes the triaxial acceleration and triaxial angular velocity output by the IMU as process inputs and the position and velocity of GPS / visual positioning as measurement updates. A rotation matrix is ​​constructed based on the attitude quaternions output by the extended Kalman filter. The theoretical acceleration and the actual acceleration are then calculated based on the rotation matrix. in For theoretical acceleration, Let be a rotation matrix. For thrust coefficient, For the quality of drones, The first flight control system issued The commanded speed of each motor, The vector of gravitational acceleration. For actual acceleration, For the first ESC feedback The actual speed of each motor; The control residuals are calculated based on theoretical and actual accelerations. Local three-dimensional wind vectors are then calculated based on the control residuals and the aerodynamic drag experienced by the UAV during its motion. in This is a local three-dimensional wind force vector. To control residuals, This is a diagonal drag coefficient matrix. To extend the world coordinate system linear velocity output by the Kalman filter.

3. The intelligent path planning system for unmanned aerial vehicles according to claim 2, characterized in that, The formula for updating the state of the extended Kalman filter is: in For the posterior state estimate at the current time, It is a nonlinear state transition function. This is an estimate of the posterior state from the previous time step. and These are the triaxial acceleration and triaxial angular velocity data acquired by the IMU, respectively. For Kalman gain, For measurement vectors, For observation functions.

4. The intelligent path planning system for unmanned aerial vehicles according to claim 1, characterized in that, The execution steps of the environment map construction module are as follows: The raw point cloud acquired by the depth camera is converted to the world coordinate system of the three-dimensional space where the UAV is located, and the point coordinates in the converted world coordinate system are obtained. The three-dimensional space is divided into several cubic voxels. Each cubic voxel maintains a Log-odds occupancy probability. Based on the occupancy probability, obstacle voxels and free space voxels are labeled, and finally a binary environment collision-free voxel map is formed. Starting from the UAV's current position and ending at the next target in the global waypoint, the A* algorithm is used to search for a discrete path skeleton in the free space voxels of the collision-free voxel map of the environment. The node evaluation cost is: in The first extension of the A* algorithm search 1 path node For nodes The overall cost of the comprehensive assessment From the starting point to the node The known cost, As weight, For nodes The corresponding point coordinates in the world coordinate system of three-dimensional space. Here are the coordinates of the endpoint of this segmented flight in the world coordinate system. It is the vector norm 2; Expanding segmentally along the path skeleton along the XYZ axes until touching the obstacle object, generating multiple sets of cuboid bounding boxes, which are then converted into convex polyhedral corridors formed by the intersection linear inequality constraints of the half-space: in For a set of convex polyhedral corridors, The total number of convex polyhedra. The sequence number of the convex polyhedron is used for segmentation. For the first Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix is ​​formed by the normal vectors of the hyperplane of the segmental convex polyhedron, where each row of the matrix corresponds to an external normal vector of the constraining hyperplane. For the first The distance intercept corresponding to the hyperplane of the segmental convex polyhedron.

5. The intelligent path planning system for unmanned aerial vehicles according to claim 4, characterized in that, The transformation formula for converting the raw point cloud acquired by the depth camera to point coordinates in the world coordinate system is as follows: in To transform to the world coordinate system The coordinates of the points Let be the rotation matrix from the body coordinate system to the world coordinate system. Let be the extrinsic parameter matrix of the depth camera relative to the machine body. In the camera coordinate system, the first The coordinates of the points Let be the extrinsic translation vector of the depth camera relative to the object. The world coordinates of the drone body.

6. The intelligent path planning system for unmanned aerial vehicles according to claim 1, characterized in that, The execution steps of the aerodynamic disturbance assessment module are as follows: Traverse the hyperplane of each convex polyhedron in the corridor and find the nearest obstacle surface based on the spatial position of the hyperplane in the collision-free voxel map of the environment. Principal component analysis is used to extract the normal vectors of the adjacent obstacle surfaces, and the channel width of the corresponding convex polyhedron at the current cross-section is calculated. in The sequence number of the segmented convex polyhedron. For the first The channel width of the cross-section of the segmental convex polyhedron. , The first The coordinates of the feature sampling points of the obstacles on the left and right sides of the cross section of the convex polyhedron. Let be the normal vector of the obstacle surface. For the first The local surface gradient of an obstacle at the cross section of a segmental convex polyhedron. Let L be the vector norm of the local surface gradient; The local three-dimensional wind force vector is projected from the normal vector of the obstacle surface and divided into normal wind force and tangential wind force: in For the first The cross-section of the convex polyhedron is perpendicular to the normal wind force on the wall surface. This is a local three-dimensional wind force vector. For the first The cross section of the segmental convex polyhedron is parallel to the tangential wind force along the corridor direction of the convex polyhedron; The offset at the cross section is calculated based on the channel width, normal wind force, and the slit effect: in For the first drone The displacement of the body at the cross section of the segmental convex polyhedron when subjected to aerodynamic disturbance. This is the empirical coefficient for amplification of the narrowing effect. For reference width, For the quality of drones, This refers to the delay time from when the flight control system detects a deviation to when it outputs effective compensating thrust.

7. The intelligent path planning system for unmanned aerial vehicles according to claim 6, characterized in that, The execution steps of the tolerance corridor generation module are as follows: Calculate the maximum dynamic shrinkage margin and predict the unit vector of the drift direction based on the offset: in For the first The maximum dynamic shrinkage allowance of the cross section of a segmental convex polyhedron. For the first The unit vector of the predicted drift direction of the cross section of a segmental convex polyhedron. Represents the 2-norm of a vector; Differential shrinkage update of the hyperplane intercept based on the maximum dynamic shrinkage margin and the predicted drift direction unit vector: in For the first The hyperplane number in a segmented convex polyhedron and The first The first segment of a convex polyhedron The original intercept and the intercept updated by differential shrinkage of each hyperplane This is the static foundation safety margin. For the first The first segment of a convex polyhedron Transpose of the normal vectors of the hyperplane; Based on the hyperplane intercept updated by differential shrinkage, a half-space intersection operation is performed. After removing redundant constraint surfaces, a new set of convex polyhedra is obtained, and the output is a spatiotemporal dynamic tolerance corridor.

8. The intelligent path planning system for unmanned aerial vehicles according to claim 7, characterized in that, When the wind field is spatially uniform, the expression for the spatiotemporal dynamic tolerance corridor is: in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. For the first after differential shrinkage Segmentally convex polyhedron, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where .

9. The intelligent path planning system for unmanned aerial vehicles according to claim 8, characterized in that, When the asymmetric contraction directions of two adjacent convex polyhedra differ significantly, the steps for obtaining the spatiotemporal dynamic tolerance corridor are as follows: Calculate the projected area of ​​the interface between two adjacent convex polyhedra: in For the first Segmented convex polyhedron and the first +1 Projected area of ​​the interface of a convex polyhedron This is a function for calculating the area of ​​polygons within the projection plane. This is a projection operator that projects a 3D feasible region onto a plane orthogonal to the normal vectors of the interface between two convex polyhedra. For the intersection operation of sets, Let be any position vector in three-dimensional space. For three-dimensional space, For the first The coefficient matrix formed by the normal vectors of the hyperplane of the segmental convex polyhedron. For the first The set of hyperplane intercepts updated by the differential shrinkage of the hyperplane of the segmental convex polyhedron, where , For the first +1 coefficient matrix formed by the normal vectors of the hyperplane of the convex polyhedron For the first +1 hyperplane intercept set updated by hyperplane differential shrinkage of convex polyhedron; Will With preset threshold In comparison, if When the interface is marked as having degraded connectivity, for each interface marked as having degraded connectivity, a hyperplane within the nearest influence domain is found, the intercept is extrapolated slightly, and the relaxation is iteratively performed until the projected area of ​​the interface is not less than a preset threshold. The relaxation iteration formula is as follows: in For the first The first segment of a convex polyhedron The intercept after each hyperplane relaxation iteration For the first The first segment of a convex polyhedron The relaxation direction coefficients of each hyperplane. The relaxation step size for a single iteration. For the first The set of intercepts after hyperplane relaxation iteration of a segmental convex polyhedron. The interface projected area is recalculated using the intercept set after relaxation iteration; Integrate all updated convex polyhedra to output a fully connected, anisotropic spatiotemporally tolerant corridor: in A collection of time-space dynamic tolerance corridors. The total number of convex polyhedra. The first after differential shrinkage and relaxation correction Segmentally convex polyhedron, To satisfy the area constraint for all adjacent convex polyhedra.

10. The intelligent path planning system for unmanned aerial vehicles according to claim 9, characterized in that, The execution steps of the trajectory planning and control module are as follows: Constructing spatial linear constraints for the corridor: Each set of linear inequalities in the spatiotemporally dynamic tolerance corridor is transformed into trajectory control point constraints. The entire trajectory does not exceed the bounds only if all piecewise polynomial control points fall within the convex polyhedron. The formula for the spatial linear constraint is: in For the first Coordinates of control points To be constrained in the first The set of control point indices within a segmented convex polyhedron; Construct thrust acceleration constraints and jerk constraints, and build a cost function based on jerk integral and flight time weighted penalty: Thrust acceleration constraints: jerk constraints: Cost function: in Let be the second derivative of the trajectory position at time t. The vector of gravitational acceleration. This represents the maximum total thrust of the drone when all its motors are fully engaged. For the quality of drones, The third derivative of the trajectory position at time t. To increase the upper limit of jerk, For the total cost, For the first The total flight time of the trajectory of the segmental convex polyhedron. This refers to the flight time penalty weighting coefficient. The spatial linear constraints, thrust acceleration constraints, jerk constraints, and cost function are combined to form a large-scale sparse quadratic programming problem. The interior-point method is used to solve the sparse quadratic programming problem, and the flight trajectory is obtained. The flight trajectory is packaged and sent to the flight control system. The formula for evaluating the flight trajectory in the time domain is: in Let be the optimal trajectory position vector at time t. For the order of the polynomial, To optimize the solution of the th The first segment of the convex polyhedron trajectory Optimal coefficients of the order polynomial For the first The effective time interval of the trajectory of a segmental convex polyhedron.