A method and system for trajectory planning of low-altitude unmanned aerial vehicles

Abstract

This invention discloses a method and system for trajectory planning of low-altitude unmanned aerial vehicles (UAVs), comprising: acquiring flight mission information, flight performance parameters, and real-time status information of the low-altitude UAV; constructing a hierarchical low-altitude state space based on low-altitude environmental data and extracting safety anchor points; using each safety anchor point as the source point, employing Hamilton-Jacobi reachability domain calculation to obtain candidate regions for reversible trajectories; generating multiple candidate four-dimensional elastic trajectory corridors based on flight mission information and performing comprehensive constraint verification; dividing the locking domain and deformation domain according to real-time status information, generating candidate deformation schemes and extracting ranking features; inputting the corridor ranking features into an improved LambdaMART model and outputting the ranking score of each candidate corridor deformation scheme. This invention achieves safe and stable trajectory planning for low-altitude UAVs through reversal analysis, elastic corridor construction, and ranking optimization.

Description

Technical Field

[0001] This invention relates to the field of intelligent control and trajectory planning technology for low-altitude unmanned aerial vehicles (UAVs), and in particular to a method and system for trajectory planning of low-altitude UAVs. Background Technology

[0002] With the development of the low-altitude economy and the application of drones, low-altitude drones are widely used in scenarios such as logistics delivery, urban inspection, emergency rescue, and surveying and data collection. The low-altitude flight environment usually includes buildings, trees, power facilities, temporary obstacles, dynamic flight targets, no-fly zones, and complex weather factors. Therefore, drones need to perform trajectory planning before performing missions to ensure that they can fly from the starting point to the target point according to the mission requirements.

[0003] Existing low-altitude UAV trajectory planning methods typically generate flight paths based on 3D maps, obstacle information, and UAV dynamic parameters, and obtain executable trajectories through path search, local obstacle avoidance, or trajectory smoothing. Some methods further incorporate dynamic obstacle prediction, communication quality assessment, or airspace restriction judgment to improve the safety and usability of trajectory planning.

[0004] However, existing methods mostly aim to generate a single trajectory, primarily focusing on whether the current trajectory avoids obstacles and meets flight dynamics constraints. They struggle to determine whether a UAV, after entering a low-altitude region, still possesses the ability to safely retreat under abnormal conditions. When dynamic obstacles change, communication links decline, navigation reliability decreases, or temporary airspace disturbances occur, existing methods also struggle to quickly select the optimal solution from multiple candidate trajectory adjustment schemes that balances safety, reversibility, energy consumption, and execution stability. This results in trajectory planning results that are currently flyable but difficult to handle safely in abnormal situations.

[0005] Therefore, how to provide a method and system for trajectory planning of low-altitude unmanned aerial vehicles is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a method and system for trajectory planning of low-altitude unmanned aerial vehicles (UAVs). This invention combines the construction of a reverse reversible domain with safety anchor points, the generation of a four-dimensional elastic trajectory corridor, and an improved LambdaMART model ranking and optimization method to achieve the generation of candidate reversible trajectories, the construction of candidate four-dimensional elastic trajectory corridors, the optimization of corridor deformation schemes, and the output of trajectory planning results for low-altitude UAVs in complex low-altitude environments. It has the advantages of strong safety reversal capability, high adaptability to dynamic environments, good trajectory execution stability, and reasonable selection of multiple candidate schemes.

[0007] A trajectory planning method for a low-altitude unmanned aerial vehicle according to an embodiment of the present invention includes:

[0008] Acquire flight mission information, flight performance parameters, and real-time status information of low-altitude UAVs;

[0009] A hierarchical low-altitude state space is constructed based on low-altitude environmental data, and safety anchor points that meet the conditions of airspace clearance, communication, landing, or return access are extracted.

[0010] Using each safety anchor point as the source point, and combining flight performance parameters and real-time status information, a layered retraction flow tube is generated by combining Hamilton-Jacobi reachability domain calculation with energy-time flow field. The layered retraction flow tube is then superimposed with the layered low-altitude state space to obtain a candidate area for retraction trajectory.

[0011] Based on flight mission information, multiple candidate four-dimensional elastic trajectory corridors are generated in the reversible trajectory candidate area. Each four-dimensional elastic trajectory corridor is composed of multiple elastic cells with three-dimensional spatial range and time passage window. Comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory corridor, and candidate four-dimensional elastic trajectory corridors that do not meet the conditions are eliminated.

[0012] Based on real-time status information, the selected four-dimensional elastic trajectory utility tunnel is divided into a locking domain and a deformation domain. Multiple candidate utility tunnel deformation schemes are generated for the deformation domain, and the utility tunnel ranking features corresponding to each candidate utility tunnel deformation scheme are extracted.

[0013] The ranking features of the utility tunnel are input into the improved LambdaMART model, and the ranking scores of each candidate utility tunnel deformation scheme are output. Combining flight mission information, flight performance parameters and real-time status information, the candidate utility tunnel deformation scheme that meets the conditions and has the highest ranking score is selected as the target trajectory utility tunnel. The target trajectory utility tunnel is updated in both the locking domain and the deformation domain, and the trajectory planning results are output.

[0014] Optionally, the flight mission information includes the starting point location, target point location, mission waypoint, flight altitude range, mission priority, and airspace restriction information; flight performance parameters include maximum flight speed, maximum acceleration, maximum rate of climb, maximum rate of descent, minimum turning radius, and hovering power consumption rate; and real-time status information includes the UAV's current position, current speed, current heading, remaining battery power, communication link quality, and navigation and positioning reliability.

[0015] Optionally, the construction of a hierarchical low-altitude state space based on low-altitude environmental data includes:

[0016] Acquire low-altitude environmental data and process it for coordinate unification, time unification, and altitude benchmark unification. The low-altitude environmental data includes low-altitude three-dimensional map data, building outline data, static obstacle data, dynamic obstacle data, airspace restriction data, communication link coverage data, navigation and positioning quality data, meteorological wind field data, and landable area data.

[0017] According to the preset spatial resolution and preset altitude interval, the low-altitude flight area is divided into multiple low-altitude spatial cells, and a cell attribute record is established for each low-altitude spatial cell. The cell attribute record includes spatial location, altitude range, time marker, obstacle occupancy status, dynamic occupancy status, airspace passage status, communication reliability status, navigation reliability status, and emergency response status.

[0018] Based on building outline data, static obstacle data, and terrain height data, each low-altitude space cell is assigned an obstacle occupancy status to form an obstacle occupancy layer; based on dynamic obstacle data, meteorological wind field data, and low-altitude traffic flow data, the dynamic occupancy status of each low-altitude space cell under different time markers is predicted to form a dynamic occupancy prediction layer.

[0019] Based on airspace restriction data, each low-altitude space cell is assigned an airspace access status, forming an airspace rule layer; based on communication link coverage data, each low-altitude space cell is assigned a communication trust status, forming a communication trust layer; based on navigation and positioning quality data, each low-altitude space cell is assigned a navigation trust status, forming a navigation trust layer.

[0020] Based on the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication trust layer, and navigation trust layer, emergency response status is marked for each low-altitude space cell to form an emergency response layer. Low-altitude space cells that meet the clearance conditions, communication conditions, landing conditions, or return access conditions are extracted from the emergency response layer as safety anchor points.

[0021] Optionally, obtaining the candidate region of the reversible trajectory includes:

[0022] Read the safety anchor point, flight performance parameters and real-time status information, and take each safety anchor point as the terminal state of the UAV’s abnormal retreat. Based on the UAV’s current position, current speed, current heading, current altitude, remaining battery power, maximum flight speed, maximum acceleration, maximum climb rate, maximum descent rate, minimum turning radius and hovering power consumption rate, determine the state transition range of the UAV when it retreats from the low-altitude space cell to the corresponding safety anchor point.

[0023] Using each safety anchor point as the starting point for reverse calculation, Hamilton-Jacobi reachability calculation is performed along the spatial adjacency direction, altitude layer switching direction and time marker reversal direction in the layered low-altitude state space to determine whether each low-altitude spatial cell marker can reach the corresponding safety anchor point within the state transition range limited by the UAV flight performance parameters.

[0024] For low-altitude space cells marked as reachable by Hamilton-Jacobi reachability domain calculation, the remaining power in real-time status information and hovering power consumption rate, climb power consumption characteristics, descent power consumption characteristics and level flight power consumption characteristics in flight performance parameters are combined to calculate the power consumption and time consumption required for the UAV to retreat from the low-altitude space cell to the corresponding safe anchor point. An energy-time flow field is formed based on the power consumption and time consumption results.

[0025] According to the retraction direction, retraction time sequence, retraction power consumption sequence, and altitude layer switching sequence in the energy-time flow field, the low-altitude space cells that continuously satisfy the retraction reachability relationship are connected into a layered retraction flow tube corresponding to each safety anchor point.

[0026] The layered retraction flow tubes are superimposed with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication reliability layer, navigation reliability layer, and emergency response layer in the layered low-altitude state space. Flow tube segments that are inconsistent with the obstacle occupancy state, dynamic occupancy state, airspace passage state, communication reliability state, navigation reliability state, and emergency response state are deleted. The low-altitude space cells covered by the remaining layered retraction flow tubes after deletion are determined as candidate areas for retraction trajectories.

[0027] Optionally, generating multiple candidate four-dimensional elastic trajectory corridors within the reversible trajectory candidate region includes:

[0028] Based on the starting point, target point, mission waypoint, flight altitude range and mission time requirements in the flight mission information, low-altitude space cells connected to the starting point, target point and mission waypoint are selected from the candidate area of ​​the reversible trajectory to form the candidate corridor search range.

[0029] Based on the UAV's body size, safety interval, positioning error, and low-altitude flight control accuracy, the low-altitude spatial cells within the candidate pipe gallery search range are spatially expanded and time windows are assigned to generate elastic cells with three-dimensional spatial range and time passage windows.

[0030] Based on the starting point, target point, mission waypoint, and flight altitude range, adjacent elastic cells with continuous passage time are connected. Multiple candidate four-dimensional elastic trajectory corridors are generated according to different altitude layers, different lateral offset directions, different time passage windows, and different waypoint connection sequences.

[0031] A comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory tunnel. The comprehensive constraint verification includes matching the elastic cells in each candidate four-dimensional elastic trajectory tunnel with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication credibility layer, navigation credibility layer and layered backflow tube, respectively, to obtain the passage verification result of each candidate four-dimensional elastic trajectory tunnel.

[0032] Based on the verification results, candidate four-dimensional elastic trajectory utility tunnels that pass the comprehensive constraint verification are retained, while candidate four-dimensional elastic trajectory utility tunnels that fail the comprehensive constraint verification are removed.

[0033] Optionally, the step of generating multiple candidate utility tunnel deformation schemes for the deformation domain and extracting the utility tunnel ranking features corresponding to each candidate utility tunnel deformation scheme includes:

[0034] Read the four-dimensional elastic trajectory tunnel that has been saved after comprehensive constraint verification, and determine the current position, current speed, current heading, current altitude and current time of the UAV in the four-dimensional elastic trajectory tunnel based on the real-time status information;

[0035] Based on the spatial distance between the current position of the UAV and each elastic cell in the four-dimensional elastic trajectory tunnel, as well as the estimated time for the UAV to reach each elastic cell, the four-dimensional elastic trajectory tunnel is divided into a locking domain and a deformation domain.

[0036] The system acquires dynamic obstacle change information, communication link change information, navigation reliability change information, wind field disturbance change information, and low-altitude traffic flow change information. It then maps the change information to elastic cells in the deformation domain to obtain the dynamic occupancy change state, communication change state, navigation change state, wind field change state, and traffic flow change state of each elastic cell in the deformation domain.

[0037] Based on the dynamic occupancy, communication, navigation, wind field, and traffic flow changes of each elastic cell within the deformation domain, multiple candidate tunnel deformation schemes are generated for the deformation domain.

[0038] For each candidate utility tunnel deformation scheme, the utility tunnel ranking features are extracted. These features include path length features, expected energy consumption features, comprehensive risk features, obstacle safety distance features, backoff margin features, communication reliability features, navigation reliability features, utility tunnel elasticity margin features, time window width features, and locking stability features.

[0039] Optionally, the step of updating the locking domain and deformation domain of the target trajectory tunnel and outputting the trajectory planning result includes:

[0040] An improved LambdaMART model is constructed, which includes a shared input layer, a task attention channel, a gradient boosting tree channel, a fast incremental update layer, and an output sorting layer. The shared input layer is connected to the task attention channel, the task attention channel is connected to the gradient boosting tree channel, the gradient boosting tree channel is connected to the fast incremental update layer, and the fast incremental update layer is connected to the output sorting layer.

[0041] The candidate tunnel deformation schemes are input into the shared input layer along with the tunnel sorting features, flight mission information, flight performance parameters, and real-time status information. Field mapping and numerical normalization are performed on path length, estimated energy consumption, comprehensive risk value, obstacle safety distance, backoff margin, communication reliability, navigation reliability, tunnel elasticity margin, time window width, and locking stability. The results are then concatenated to form the input feature vector.

[0042] The input feature vector is fed into the task attention channel. Through a fully connected feature selection layer and a multi-head attention layer, the backoff margin feature, communication credibility feature and navigation credibility feature are weighted and enhanced, and the task-weighted feature vector is output.

[0043] The task-weighted feature vector is input into the gradient boosting tree channel. An improved LambdaMART backbone network, formed by serially stacking multiple gradient boosting decision trees, is used to rank and learn the candidate tube gallery deformation schemes. A feature resampling layer is inserted into the gradient boosting tree channel to resample the high-dimensional sparse features.

[0044] The ranking results output from the gradient boosting tree channel are input into the fast incremental update layer. Based on the changes in dynamic obstacles, communication links, and navigation confidence, local incremental updates are performed on the affected gradient boosting decision tree nodes. The updated ranking results are input into the output ranking layer to generate the final ranking score of the candidate tunnel deformation scheme.

[0045] The candidate pipe gallery deformation schemes are ranked according to the final ranking score. Combining flight mission information, flight performance parameters and real-time status information, the candidate pipe gallery deformation scheme with the highest ranking score is selected as the target trajectory pipe gallery. Locking domain update and deformation domain update are performed on the target trajectory pipe gallery, and the trajectory planning results for low-altitude UAV flight control are output.

[0046] A trajectory planning system for a low-altitude unmanned aerial vehicle according to an embodiment of the present invention includes:

[0047] The information acquisition module is used to acquire flight mission information, flight performance parameters, and real-time status information;

[0048] The state space construction module is used to construct a hierarchical low-altitude state space based on low-altitude environmental data and extract safety anchor points from it.

[0049] The retraction candidate area generation module is used to generate layered retraction flow tubes based on safety anchor points, flight performance parameters and real-time status information, and superimpose the layered low-altitude state space to obtain retraction trajectory candidate areas.

[0050] The utility tunnel generation and verification module is used to generate candidate four-dimensional elastic trajectory utility tunnels in the reversible trajectory candidate area based on flight mission information, and to perform comprehensive constraint verification.

[0051] The deformation feature extraction module is used to divide the verified four-dimensional elastic trajectory pipe gallery into a locking domain and a deformation domain, generate candidate pipe gallery deformation schemes, and extract pipe gallery ranking features.

[0052] The sorting output module is used to input the pipe gallery sorting features into the improved LambdaMART model, determine the target trajectory pipe gallery, update the locking domain and deformation domain, and output the trajectory planning results.

[0053] The beneficial effects of this invention are:

[0054] This invention enables low-altitude UAVs to identify safe retraction areas during the trajectory planning stage by constructing safety anchor points, layered low-altitude state spaces, and layered retraction flow tubes. This avoids UAVs entering low-altitude areas that, although passable for short periods, are difficult to evacuate in case of communication anomalies, navigation anomalies, or intrusion of dynamic obstacles, thereby improving the safety of trajectory planning and the ability to handle anomalies.

[0055] This invention generates a four-dimensional flexible trajectory corridor within a reversible trajectory candidate area, enabling trajectory planning results to move beyond a single rigid path and incorporate three-dimensional spatial adjustment margins and time passage windows. When dynamic obstacle changes, wind disturbances, temporary airspace changes, or low-altitude traffic flow changes occur in the low-altitude environment, the UAV can perform lateral offsets, altitude adjustments, time window adjustments, or local waiting within the deformation domain, reducing the need for frequent replanning of the entire trajectory.

[0056] This invention improves the LambdaMART model to rank and optimize multiple candidate pipe rack deformation schemes. It can use flight mission information, flight performance parameters, real-time status information, and pipe rack ranking characteristics together for trajectory scheme selection, so that the final output target trajectory pipe rack takes into account safety, reversibility, energy consumption, communication reliability, navigation reliability, and execution stability, thereby improving the trajectory planning efficiency and engineering feasibility of low-altitude UAVs in complex environments. Attached Figure Description

[0057] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0058] Figure 1 This is a flowchart of a trajectory planning method for low-altitude unmanned aerial vehicles proposed in this invention;

[0059] Figure 2 This is a schematic diagram of a trajectory planning system for low-altitude unmanned aerial vehicles proposed in this invention. Detailed Implementation

[0060] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0061] refer to Figure 1 A method for trajectory planning of low-altitude unmanned aerial vehicles, comprising:

[0062] Acquire flight mission information, flight performance parameters, and real-time status information of low-altitude UAVs;

[0063] A hierarchical low-altitude state space is constructed based on low-altitude environmental data, and safety anchor points that meet the conditions of airspace clearance, communication, landing, or return access are extracted.

[0064] Using each safety anchor point as the source point, and combining flight performance parameters and real-time status information, a layered retraction flow tube is generated by combining Hamilton-Jacobi reachability domain calculation with energy-time flow field. The layered retraction flow tube is then superimposed with the layered low-altitude state space to obtain a candidate area for retraction trajectory.

[0065] Based on flight mission information, multiple candidate four-dimensional elastic trajectory corridors are generated in the reversible trajectory candidate area. Each four-dimensional elastic trajectory corridor is composed of multiple elastic cells with three-dimensional spatial range and time passage window. Comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory corridor, and candidate four-dimensional elastic trajectory corridors that do not meet the conditions are eliminated.

[0066] Based on real-time status information, the selected four-dimensional elastic trajectory utility tunnel is divided into a locking domain and a deformation domain. Multiple candidate utility tunnel deformation schemes are generated for the deformation domain, and the utility tunnel ranking features corresponding to each candidate utility tunnel deformation scheme are extracted.

[0067] The ranking features of the utility tunnel are input into the improved LambdaMART model, and the ranking scores of each candidate utility tunnel deformation scheme are output. Combining flight mission information, flight performance parameters and real-time status information, the candidate utility tunnel deformation scheme that meets the conditions and has the highest ranking score is selected as the target trajectory utility tunnel. The target trajectory utility tunnel is updated in both the locking domain and the deformation domain, and the trajectory planning results are output.

[0068] In this embodiment, the flight mission information includes the starting point location, target point location, mission waypoint, flight altitude range, mission priority, and airspace restriction information. The flight performance parameters include maximum flight speed, maximum acceleration, maximum climb rate, maximum descent rate, minimum turning radius, and hovering power consumption rate. The real-time status information includes the UAV's current position, current speed, current heading, remaining battery power, communication link quality, and navigation and positioning reliability.

[0069] In this embodiment, the construction of a hierarchical low-altitude state space based on low-altitude environmental data includes:

[0070] Acquire low-altitude environmental data and process it for coordinate unification, time unification, and altitude benchmark unification. The low-altitude environmental data includes low-altitude three-dimensional map data, building outline data, static obstacle data, dynamic obstacle data, airspace restriction data, communication link coverage data, navigation and positioning quality data, meteorological wind field data, and landable area data.

[0071] According to a preset spatial resolution and preset altitude interval, the low-altitude flight area is divided into multiple low-altitude spatial cells. A cell attribute record is established for each low-altitude spatial cell. The cell attribute record includes spatial location, altitude range, time stamp, obstacle occupancy status, dynamic occupancy status, airspace clearance status, communication reliability status, navigation reliability status, and emergency response status. Specifically, the division of the low-altitude flight area into multiple low-altitude spatial cells according to the preset spatial resolution and preset altitude interval is as follows:

[0072] Horizontal grid generation: Read the coordinates of the starting point and the target point, add a 200m buffer to the outer edge according to the maximum yaw angle of the mission segment to form the minimum bounding rectangle, and divide the rectangle into equal intervals with a preset horizontal resolution of 20m in the east-west direction and 20m in the south-north direction to obtain several two-dimensional horizontal grid units, each unit corresponding to a horizontal block ID;

[0073] Vertical layer slicing: Based on the minimum safe height and the maximum allowable height of the task, the system is divided upwards from the maximum allowable height of the task at preset height intervals of 10m to obtain nine height layers; the resulting two-dimensional mesh is copied in each height layer and numbered according to the layer, and stacked on top of each other to form a cubic mesh;

[0074] Safety redundancy expansion: For each cubic mesh, expand outward by a 5m safety boundary. The side length after expansion is uniformly recorded as 30m horizontal × 10m vertical, and the final low-altitude space cell is obtained.

[0075] Unique Index and Basic Label: Generate a unique index for each cell by layer number, row number, and column number, and write the cell center latitude, longitude and height coordinates, horizontal side length, vertical height range and time period to the spatial location, height range and time identifier.

[0076] Attribute writing:

[0077] Based on the 3D model of the building and the fixed facility model, cells that overlap with the building volume within 1m are marked as obstacles with an occupancy status of 1, and the rest are marked as 0.

[0078] Using a three-second time window for dynamic target prediction, cells that will be passed by drones or helicopters within the next 15 seconds are written with a dynamic occupancy status of 1, and the rest are written with 0.

[0079] Based on the no-fly, restricted-fly, and passage periods issued by air traffic control, write the prohibited or restricted cell to airspace passage status = 0, and write 1 to allow passage.

[0080] Using a simulated cellular network coverage map, cells with signal strength below -110dBm are written with a communication trust status of 0, and the rest are written with 1.

[0081] Call the GNSS multipath model, write navigation confidence state = 0 for cells with estimated positioning error exceeding 5m, and write 1 for the rest;

[0082] Read the coordinates of the landing platform and the entrance to the backup altitude layer, write the cell containing the center point to the emergency response status = 1, and write 0 to the rest;

[0083] Based on building outline data, static obstacle data, and terrain height data, each low-altitude space cell is assigned an obstacle occupancy status, forming an obstacle occupancy layer. Based on dynamic obstacle data, meteorological wind field data, and low-altitude traffic flow data, the dynamic occupancy status of each low-altitude space cell at different time markers is predicted, forming a dynamic occupancy prediction layer, wherein:

[0084] The obstacle occupancy layer is formed as follows:

[0085] Read the building outline file and unfold the 3D building surface into a set of facets; take the latitude, longitude, and elevation of the center point of each low-altitude space cell, compare it with the horizontal projection range of the building and the height of the building top, if the horizontal projection of the center point falls within the range and the height of the cell bottom is one meter lower than the height of the building top, then mark the cell's obstacle occupancy status as one; read the 3D bounding box of the fixed facility model, if the cell center point is located within the bounding box, then mark it as one; read the terrain digital elevation data, if the height of the cell bottom is lower than the local ground elevation plus two meters, then mark it as one; otherwise, mark it as zero, completing the obstacle occupancy layer;

[0086] A dynamic occupancy prediction layer is formed, specifically as follows:

[0087] Read the broadcast trajectory of low-altitude traffic aircraft and perform cubic spline interpolation smoothing at one-second intervals; extrapolate each trajectory to the next 15 seconds in one-second steps to obtain the predicted track point; for each time step, map the cylindrical range of the predicted track point with a radius of 10 meters to the corresponding cell and mark the dynamic occupancy status as 1; d) use airborne lidar to cluster the real-time point cloud to obtain the bird flock center point, map the radius of 5 meters to the corresponding cell and mark it as 1; call the micro-meteorological model to calculate the wind field change area in the next 15 seconds, mark the cell where the wind speed change exceeds 12 meters per second as 1; unmarked cells are marked as zero at the corresponding time step, and the dynamic occupancy prediction layer is output;

[0088] Based on airspace restriction data, each low-altitude space cell is assigned an airspace accessibility status, forming an airspace rule layer; based on communication link coverage data, each low-altitude space cell is assigned a communication trustworthiness status, forming a communication trustworthiness layer; based on navigation and positioning quality data, each low-altitude space cell is assigned a navigation trustworthiness status, forming a navigation trustworthiness layer, wherein:

[0089] The airspace rule layer is formed as follows:

[0090] Read the published no-fly and restricted-fly zone files and filter the currently valid rules; intersect the three-dimensional volume of the no-fly zone with the cell. If the intersection volume ratio exceeds one percent, mark the airspace passage status of the cell as zero; check the speed and altitude requirements of the restricted-fly zone. If the cell center altitude exceeds the limit or the flight speed conflicts with the requirements, mark it as zero; mark other cells as one, and obtain the airspace rule layer.

[0091] The communication trust layer is formed as follows:

[0092] Import the list of fifth-generation mobile communication base stations and their transmission power, and use a ray tracing model to calculate the received signal strength at the center point of each cell. If the strength is lower than -110 dB / mW or the signal-to-noise ratio is lower than -5 dB, the cell's communication trust status is marked as zero. For long-distance low-power wireless or satellite backup links in cellular blind spots, if the backup link's signal-to-noise ratio is higher than 10 dB, the communication trust status is marked as one; otherwise, it is marked as zero. Output the communication trust layer.

[0093] The navigation trust layer is formed as follows:

[0094] Import satellite navigation ephemeris and city 3D model, perform multipath simulation for each cell center point, and calculate the 3D root mean square positioning error; if the positioning error is greater than five meters or the number of visible satellites is less than eight, mark the cell navigation confidence state as zero; check the availability of visual-inertial fusion, if the ambient illumination is higher than fifty lux and the number of feature points exceeds two hundred, mark the navigation confidence state as one; output navigation confidence layer.

[0095] Based on the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication trust layer, and navigation trust layer, emergency response status is marked for each low-altitude space cell to form an emergency response layer. Low-altitude space cells that meet the airspace clearance conditions, communication conditions, landing conditions, or return-to-base access conditions are extracted from this emergency response layer as safety anchor points. Specifically, the emergency response layer is formed as follows:

[0096] Read the list of landable platforms, including rooftop helipads, temporary take-off and landing sites, and open plazas. Calculate the horizontal projection boundary and allowable landing height for each platform. Overlay the boundary onto low-altitude space cells. If the cell center point is located within the boundary and the cell bottom height falls within the allowable landing height range of the platform, mark the cell's emergency response status as one.

[0097] Read the list of backup altitude layer entrances, including multi-level airway take-off and landing nodes, and for each entrance, take the center coordinates and a cylindrical area with a radius of 10 meters; if the cell center point is located within the cylindrical area, mark the emergency response status as one;

[0098] Check the status of the marked cells in the obstacle occupancy layer and the dynamic occupancy prediction layer. If the obstacle occupancy status or the dynamic occupancy status is one, roll back the marking and reset the emergency response status to zero.

[0099] Check the passage status of the marked cells in the airspace rule layer. If the airspace passage status is zero, roll back the marking and reset the emergency response status to zero.

[0100] Check the trust status of the marked cell in the communication trust layer and the navigation trust layer. If either trust status is zero, roll back the marking and reset the emergency response status to zero.

[0101] For cells that still retain the emergency response status as one, perform connectivity clustering, calculate the one-to-one mapping relationship between the center point of each connected component and the landing platform or backup altitude layer entrance, extract the center points of connected components that meet the clearance conditions, communication conditions, landing conditions or return access conditions as safety anchor points, and output the set of safety anchor points.

[0102] In this embodiment, obtaining the candidate region of the reversible trajectory includes:

[0103] Read the safety anchor point, flight performance parameters and real-time status information, and take each safety anchor point as the terminal state of the UAV’s abnormal retreat. Based on the UAV’s current position, current speed, current heading, current altitude, remaining battery power, maximum flight speed, maximum acceleration, maximum climb rate, maximum descent rate, minimum turning radius and hovering power consumption rate, determine the state transition range of the UAV when it retreats from the low-altitude space cell to the corresponding safety anchor point.

[0104] Using each safety anchor point as the starting point for reverse calculation, Hamilton-Jacobi reachability domain calculation is performed along the spatial adjacency direction, altitude layer switching direction, and time stamp reversal direction in the layered low-altitude state space. This verifies whether each low-altitude spatial cell marker can reach the corresponding safety anchor point within the state transition range defined by the UAV's flight performance parameters. Specifically, the Hamilton-Jacobi reachability domain calculation along the spatial adjacency direction, altitude layer switching direction, and time stamp reversal direction in the layered low-altitude state space is as follows:

[0105] Add the cell containing the safety anchor point to the queue to be expanded, and record that the energy required to reach the cell is zero and the arrival time is zero;

[0106] Take the first cell from the queue to be expanded, read its spatial location, height layer number and time identifier, search for adjacent cells in the horizontal plane according to the four adjacent directions of up, down, left and right, and search for adjacent cells above and below in the vertical direction according to the upper height layer and the lower height layer.

[0107] For each candidate adjacent cell, the hierarchical low-altitude state spatial attributes are checked in reverse. If the obstacle occupancy state, dynamic occupancy state, airspace passage state, communication reliability state, and navigation reliability state all meet the passage requirements, the energy and time required for the UAV to fly from the current cell to the candidate adjacent cell are calculated, and the energy and time are added to the current recorded value.

[0108] If the accumulated energy does not exceed the remaining power and the accumulated time does not exceed the task's allowed rollback time, then the candidate adjacent cell is marked as reachable, and an arrival record is written in the frame with the time marker decremented by one, and pushed into the queue to be expanded.

[0109] Repeat the process until the queue to be expanded is empty or the accumulated energy and time of all remaining queue nodes exceed the limit.

[0110] Write the corresponding security anchor number to all cells that have been marked as reachable, forming a hierarchical reachable domain obtained by backward expansion from the security anchor;

[0111] For low-altitude space cells marked as reachable by Hamilton-Jacobi reachability domain calculation, the remaining battery power in real-time status information and the hovering power consumption rate, climb power consumption characteristics, descent power consumption characteristics, and level flight power consumption characteristics in flight performance parameters are combined to calculate the power consumption and time consumption required for the UAV to retreat from the low-altitude space cell to the corresponding safe anchor point. An energy-time flow field is then formed based on the power consumption and time consumption results. Specifically, the calculation of the power consumption and time consumption required for the UAV to retreat from the low-altitude space cell to the corresponding safe anchor point is as follows:

[0112] For each low-altitude space cell that has been marked as reachable, read the coordinates of the center point and the corresponding safety anchor point, and decompose it into three types of flight segments: horizontal straight segment, climbing segment, and descent segment;

[0113] For horizontal straight sections, calculate the horizontal flight distance and power consumption based on the cruise speed and power consumption per unit distance in the flight performance parameters, and record the horizontal flight time;

[0114] For the climbing section, calculate the power required for climbing based on the height difference and the power consumption rate per unit height, and calculate the climbing time based on the maximum climbing rate;

[0115] For the descent phase, the required power consumption for descent is calculated based on the height difference and the power consumption rate per unit height descent, and the descent time is calculated based on the maximum descent rate;

[0116] Adding the energy consumption and time for the three flight segments of horizontal, climb, and descent, we obtain the total energy consumption and total time consumption required to retreat from the cell to the corresponding safe anchor point.

[0117] An energy-time flow field is formed based on the results of power consumption and time consumption, specifically as follows:

[0118] Record the total power consumption and total time consumption for each reachable cell, and construct a rollback cost grid with power consumption and time consumption as two dimensions;

[0119] Write a directional marker in each reachable cell that points to the simultaneous decrease of both the charge and time cost in adjacent cells, forming a pullback direction field;

[0120] All reachable cells are numbered in ascending order of energy-time cost to form a pullback time sequence table;

[0121] By storing the orientation field, cost grid, and sequence list in a unified manner, the energy-time flow field is obtained.

[0122] Based on the retraction direction, retraction time sequence, retraction power consumption sequence, and altitude layer switching sequence in the energy-time flow field, low-altitude space cells that continuously satisfy the retraction reachability relationship are connected into layered retraction flow tubes corresponding to each safety anchor point. Specifically, the connection of low-altitude space cells that continuously satisfy the retraction reachability relationship into layered retraction flow tubes corresponding to each safety anchor point is as follows:

[0123] Starting from the cell where each safety anchor point is located, follow the energy-time flow field to the adjacent cells where the cost decreases simultaneously, trace sequentially according to the direction marker, and record the sequence of cells traversed by the path;

[0124] When a path crosses a horizontal adjacent boundary, it is connected in the horizontal plane; when a path crosses an upper or lower layer boundary, it is connected in the vertical direction. Layer numbers are added to paths at different height levels.

[0125] If the cost stops decreasing or a recorded path is encountered during the tracking process, the tracking will stop and the current path will be merged with the recorded path.

[0126] Perform connectivity checks on multiple paths obtained from the same safety anchor point, merge path segments with a distance less than twice the cell side length to form a backflow tube with a continuous spatial corridor;

[0127] For each backtracking flow, record the corresponding safety anchor point identifier, the included height layer number, the cumulative power consumption along the path, and the cumulative time along the path, and output a set of layered backtracking flow.

[0128] The stratified retraction flow tube includes a low-altitude spatial cell sequence, corresponding safety anchor point markers, retraction direction markers, altitude layer switching markers, and time passage markers;

[0129] The layered retraction flow tubes are superimposed with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication reliability layer, navigation reliability layer, and emergency response layer in the layered low-altitude state space. Flow tube segments that are inconsistent with the obstacle occupancy state, dynamic occupancy state, airspace passage state, communication reliability state, navigation reliability state, and emergency response state are deleted. The low-altitude space cells covered by the remaining layered retraction flow tubes after deletion are determined as candidate areas for retraction trajectories.

[0130] In this embodiment, generating multiple candidate four-dimensional elastic trajectory corridors within the reversible trajectory candidate region includes:

[0131] Based on the starting point, target point, mission waypoint, flight altitude range and mission time requirements in the flight mission information, low-altitude space cells connected to the starting point, target point and mission waypoint are selected from the candidate area of ​​the reversible trajectory to form the candidate corridor search range.

[0132] Based on the UAV's airframe size, safety interval, positioning error, and low-altitude flight control accuracy, the low-altitude spatial cells within the candidate pipe gallery search range are spatially expanded and time windows are assigned, generating elastic cells with a three-dimensional spatial range and a time passage window. Specifically, generating elastic cells with a three-dimensional spatial range and a time passage window involves:

[0133] Read the length, width and height of the drone body, and add a two-meter safety interval and a fifty-centimeter positioning error redundancy to the outer edge of the drone body to obtain the minimum safety envelope size;

[0134] Based on the minimum safe envelope size, each low-altitude space cell is simultaneously expanded by half the safe envelope size in the three directions of east-west, north-south, and up-down within the candidate tube gallery search range to obtain the expanded three-dimensional space block.

[0135] The spatial latitude and longitude range, minimum height and maximum height corresponding to the three-dimensional spatial block are recorded as the three-dimensional spatial range of the elastic cell;

[0136] Read the maximum allowable lateral deviation and maximum allowable longitudinal deviation of the UAV control system, and write them into the lateral offset range and longitudinal offset range of the elastic cell, respectively;

[0137] Based on the expected time of the task and the tolerable amount of advance or delay, write the start time and end time into the elastic cell to form a time passage window;

[0138] Write the three-dimensional spatial extent and time together through a window into the elastic cell record, and output the elastic cell set with the three-dimensional spatial extent and time through a window;

[0139] Based on the starting point, target point, mission waypoint, and flight altitude range, adjacent elastic cells with consecutive passage times are connected. Multiple candidate four-dimensional elastic trajectory corridors are generated according to different altitude layers, different lateral offset directions, different time passage windows, and different waypoint connection sequences. Specifically, the generation of multiple candidate four-dimensional elastic trajectory corridors is as follows:

[0140] In the order of mission start point, target point and waypoints, select the first batch of elastic cells in the set of elastic cells that coincide with the waypoints and whose altitude layer meets the flight altitude range.

[0141] Within the same height layer, search for elastic cells that are adjacent to the selected elastic cells and whose time passes through the window in a horizontal octet. Arrange the search results in chronological order to form an initial horizontal channel.

[0142] In sections that require crossing multiple height layers, search for elastic cells with consecutive time windows in adjacent height layers in chronological order, connect them with the horizontal channel to form a vertical channel;

[0143] For multiple lateral offset branches existing in the same path direction, sub-channels are constructed using three lateral connection strategies: left offset, centerline, and right offset, to obtain path combinations with different lateral offset directions.

[0144] Connect the horizontal passage, vertical passage and lateral branch road in the order from the starting point to the destination to form a complete travel path;

[0145] For each combination of height layer arrangement, lateral offset direction and time window, a four-dimensional elastic trajectory utility tunnel is generated. Multiple combinations can yield multiple candidate four-dimensional elastic trajectory utility tunnels, and a set of candidate utility tunnels is output.

[0146] A comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory tunnel. This verification includes matching the elastic cells in each candidate four-dimensional elastic trajectory tunnel with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication credibility layer, navigation credibility layer, and layered backtracking flow tube, respectively, to obtain the passage verification results for each candidate four-dimensional elastic trajectory tunnel. Specifically, obtaining the passage verification results for each candidate four-dimensional elastic trajectory tunnel involves:

[0147] Read all elastic cells in each candidate four-dimensional elastic trajectory tunnel in sequence, and perform a one-to-one lookup between the unique index of the elastic cell and the obstacle occupancy layer index table. If any elastic cell is marked as occupied in the obstacle occupancy layer, immediately set the passage verification result of the trajectory tunnel to obstacle not passing.

[0148] For the trajectory tunnel that is still in the pass state, the elastic cell index is matched with the dynamic occupancy prediction layer according to the time identifier. If there is a dynamic occupancy mark at any time step, the pass verification result is set to dynamic occupancy failure.

[0149] For the trajectory tunnel that is still in the passable state, match the elastic cell index with the spatial rule layer. If the spatial passability state is impassable, set the passability verification result to spatial rule failure.

[0150] For the trajectory tunnel that is still in the pass state, the elastic cell index is matched with the communication trust layer. If the communication trust state of any elastic cell is zero, the pass verification result is set to communication failure.

[0151] For the trajectory tunnel that is still in the pass state, the elastic cell index is matched with the navigation confidence layer. If the navigation confidence state of any elastic cell is zero, the pass verification result is set to navigation failure.

[0152] For the trajectory tube gallery that is still in the pass state, the elastic cells are matched with the layered retreat flow tubes in the path order. If the continuous path segment exceeds the coverage of the retreat flow tube by more than three elastic cells, the pass verification result is set to retreat failure; otherwise, the pass verification result is set to full pass.

[0153] Output the passage verification result for each candidate four-dimensional elastic trajectory utility tunnel;

[0154] Based on the verification results, candidate four-dimensional elastic trajectory utility tunnels that pass the comprehensive constraint verification are retained, while candidate four-dimensional elastic trajectory utility tunnels that fail the comprehensive constraint verification are removed.

[0155] In this embodiment, the step of generating multiple candidate pipe gallery deformation schemes for the deformation domain and extracting the pipe gallery ranking features corresponding to each candidate pipe gallery deformation scheme includes:

[0156] Read the four-dimensional elastic trajectory tunnel that has been saved after comprehensive constraint verification, and determine the current position, current speed, current heading, current altitude and current time of the UAV in the four-dimensional elastic trajectory tunnel based on the real-time status information;

[0157] Based on the spatial distance between the current position of the UAV and each elastic cell in the four-dimensional elastic trajectory tunnel, and the estimated time for the UAV to reach each elastic cell, the four-dimensional elastic trajectory tunnel is divided into a locking domain and a deformation domain. Specifically, the division of the four-dimensional elastic trajectory tunnel into the locking domain and the deformation domain is as follows:

[0158] Read the current position coordinates, current speed and current time of the UAV, calculate the Euclidean space distance from the current position of the UAV to the center point of each elastic cell in the four-dimensional elastic trajectory gallery, and sort them in ascending order of distance;

[0159] Based on the current velocity and the start time of the elastic cell time window, estimate the expected time for the UAV to arrive at each elastic cell;

[0160] Set a spatial threshold D and a time threshold T. For elastic cells whose distance does not exceed D or whose expected arrival time does not exceed T, mark them as lock candidates.

[0161] For all elastic cells marked as lock candidates, combine them consecutively in the order of the path; if the combination length exceeds three elastic cells, mark the consecutive combination as a lock domain; if the combination length is less than three elastic cells, retain the original mark but do not include it in the lock domain.

[0162] Remove the elastic cells that have been assigned to the locking domain in the four-dimensional elastic trajectory tube gallery from the tube gallery path set, and mark the remaining set of elastic cells that have not been removed as the deformation domain.

[0163] Record the start and end indices of the locked field and the start and end indices of the deformable field;

[0164] Information on dynamic obstacle changes, communication link changes, navigation reliability changes, wind field disturbance changes, and low-altitude traffic flow changes is acquired. This information is then mapped to elastic cells within the deformation domain, yielding the dynamic occupancy, communication, navigation, wind field, and traffic flow states of each elastic cell within the deformation domain. Specifically, the dynamic occupancy, communication, navigation, wind field, and traffic flow states of each elastic cell within the deformation domain are obtained as follows:

[0165] Read the latest airborne radar spot and ground ADS-B trajectory increment within one minute, and project the newly added or displaced target trajectory of more than ten meters onto the deformation domain three-dimensional grid. If the target cylinder with a diameter of ten meters overlaps with the elastic cell, write the dynamic occupancy change state = 1 in the elastic cell, otherwise write zero.

[0166] Get the base station signal strength change table for the last five seconds. If the received power at the center of the elastic cell drops by more than 15 dB / mW or the signal-to-noise ratio drops by more than 8 dB, write the communication change status = 1; otherwise, write zero.

[0167] Read the positioning error fitting table for the last five seconds. If the three-dimensional positioning error threshold of the elastic cell center point changes from less than five meters to greater than five meters, write navigation change status = 1; otherwise, write zero.

[0168] Read the latest wind field update from the micro-meteorological model. If the wind speed increment at the center point of the elastic cell exceeds eight meters per second or the wind direction change angle exceeds thirty degrees, write the wind field change status = 1; otherwise, write zero.

[0169] Read the change in low-altitude traffic flow density grid. If the number density of drones in the horizontal grid where the elastic cell is located jumps from below the low density threshold to above the high density threshold, write the traffic flow change state = 1; otherwise, write zero.

[0170] Based on the dynamic occupancy, communication, navigation, wind field, and traffic flow changes of each elastic cell within the deformation domain, multiple candidate utility tunnel deformation schemes are generated for the deformation domain. Specifically, the generation of multiple candidate utility tunnel deformation schemes for the deformation domain is as follows:

[0171] State clustering is performed on the deformation domain, and elastic cells with continuous labels of changing states are aggregated into changing segments, while unlabeled segments are aggregated into stable segments.

[0172] For each change segment, the following three types of basic deformation actions are generated: lateral migration of five to fifteen meters, vertical ascent or descent of one to two height levels, and time window shifting backward by three to ten seconds;

[0173] The basic deformation actions are combined according to the type of change: if the type of change is dynamic occupancy or traffic flow, lateral migration is given priority; if the type of change is communication or navigation, vertical ascent is given priority; if the type of change is wind field, both lateral migration and time window shifting are considered.

[0174] For each change segment, three threshold levels are enumerated to generate corresponding deformation action amplitudes, forming multiple action combinations;

[0175] The action combinations of each change section are spliced ​​together in the path order and spliced ​​together with the original path segment of the stable section to form a complete deformation scheme. Three types of candidate tunnel deformation schemes are generated: horizontal priority scheme, vertical priority scheme and time delay scheme. Each candidate tunnel deformation scheme is written into the scheme list.

[0176] For each candidate utility tunnel deformation scheme, the utility tunnel ranking features are extracted. These features include path length features, expected energy consumption features, comprehensive risk features, obstacle safety distance features, backoff margin features, communication reliability features, navigation reliability features, utility tunnel elasticity margin features, time window width features, and locking stability features.

[0177] In this embodiment, the step of updating the locking domain and deformation domain of the target trajectory tunnel and outputting the trajectory planning result includes:

[0178] An improved LambdaMART model is constructed, which includes a shared input layer, a task attention channel, a gradient boosting tree channel, a fast incremental update layer, and an output sorting layer. The shared input layer is connected to the task attention channel, the task attention channel is connected to the gradient boosting tree channel, the gradient boosting tree channel is connected to the fast incremental update layer, and the fast incremental update layer is connected to the output sorting layer.

[0179] The candidate tunnel deformation schemes are input into the shared input layer along with the tunnel sorting features, flight mission information, flight performance parameters, and real-time status information. Field mapping and numerical normalization are performed on path length, estimated energy consumption, comprehensive risk value, obstacle safety distance, backoff margin, communication reliability, navigation reliability, tunnel elasticity margin, time window width, and locking stability. The results are then concatenated to form the input feature vector.

[0180] The input feature vector is fed into the task attention channel. A fully connected feature selection layer and a multi-head attention layer are used to enhance the weights of the pullback margin feature, communication reliability feature, and navigation reliability feature, outputting a task-weighted feature vector. Specifically, the task-weighted feature vector is as follows:

[0181] The input feature vector is fed into a fully connected feature selection layer, and features with a sparsity of 30% or less that have an impact of less than one on the ranking are filtered out.

[0182] The filtered feature vectors are fed into a multi-head attention layer with eight attention heads. Each attention head calculates attention weights for three key features: pullback margin, communication reliability, and navigation reliability.

[0183] The weighted results of the eight attention heads are concatenated along the channel dimension, and the concatenated results are normalized by the layer to obtain the task-weighted feature vector.

[0184] The task-weighted feature vector is input into the gradient boosting tree channel. An improved LambdaMART backbone network, formed by serially stacking multiple gradient boosting decision trees, is used to rank and learn the candidate tunnel deformation schemes. A feature resampling layer is inserted into the gradient boosting tree channel to resample the high-dimensional sparse features. Specifically, the improved LambdaMART backbone network, formed by serially stacking multiple gradient boosting decision trees, is used to rank and learn the candidate tunnel deformation schemes.

[0185] Initialize the first set of base decision trees in the gradient boosting tree channel. The first set of base decision trees includes twenty-five gradient boosting decision trees. The maximum depth of each gradient boosting decision tree is set to four layers. Each tree randomly selects 20% of the training samples from all candidate tunnel deformation schemes as training sub-samples.

[0186] The task-weighted feature vector of each candidate tunnel deformation scheme is input into the first set of base decision trees. Each base decision tree outputs the initial ranking score of the corresponding candidate tunnel deformation scheme. The outputs of the twenty-five base decision trees are summed to obtain the initial ranking result.

[0187] Based on the difference between the initial sorting result and the preset sorting labels in the training samples, a sorting residual is generated, and the sorting residual is used as the training target for the next set of gradient boosting decision trees.

[0188] Gradient boosting decision trees are added sequentially with a learning rate of 0.04. After every 25 gradient boosting decision trees are added, the sorting residual is recalculated, and the output of the newly added gradient boosting decision trees is added to the sorting result of the previous round.

[0189] When the total number of gradient boosting decision trees reaches 150, or the residual decrease rate is less than 0.5% for two consecutive rounds of sorting, stop adding gradient boosting decision trees;

[0190] The outputs of all gradient boosting decision trees are weighted and summed to obtain the original score vector for ranking the candidate tunnel deformation schemes;

[0191] The high-dimensional sparse features are resampled, specifically as follows:

[0192] The first feature resampling layer is inserted after the 75th gradient boosting decision tree, and the second feature resampling layer is inserted after the 125th gradient boosting decision tree.

[0193] Read the features of the candidate tube gallery deformation schemes that are currently entering the feature resampling layer, and count the occurrence frequency and distribution ratio of each feature in the training samples;

[0194] Sparse features with an occurrence rate of less than one percent are categorized into corresponding feature groups according to their source. These feature groups include dynamic obstacle change group, communication link change group, navigation reliability change group, wind field disturbance change group, and traffic flow change group.

[0195] Sparse features within the same feature group are merged, and the merged features are used as sparse bucket features to input the gradient boosting decision tree.

[0196] For sparse features that cannot be classified into feature groups, a fixed-length hash mapping is performed, mapping them to the corresponding bucket position in one thousand hash buckets. The sparse bucket features and hash bucket features are concatenated with the original non-sparse features to form a resampled feature vector.

[0197] The ranking results output from the gradient boosting tree channel are input into the fast incremental update layer. Based on changes in dynamic obstacles, communication links, and navigation reliability, local incremental updates are performed on the affected gradient boosting decision tree nodes. The updated ranking results are then input into the output ranking layer to generate the final ranking score for the candidate tunnel deformation schemes, where:

[0198] Perform local incremental updates on the affected gradient boosting decision tree nodes, specifically:

[0199] Read the real-time updated dynamic obstacle change status, communication change status, and navigation change status, and compare them with the corresponding status in the previous round of sorting;

[0200] When the dynamic obstacle state, communication state, or navigation state changes, determine the affected utility tunnel sequencing characteristics;

[0201] In all gradient boosting decision trees, find the tree node whose splitting condition contains the sorting characteristics of the affected utility tunnel, and mark the tree node and its subordinate child nodes as nodes to be updated.

[0202] Reread the training sample statistics and current candidate tunnel deformation scheme feature values ​​corresponding to the node to be updated, and update the splitting direction and leaf node output value of the node to be updated.

[0203] If the output value of the leaf node of the node to be updated changes by more than 2%, then rebuild the subtree structure of no more than three levels below the node to be updated.

[0204] Keep the tree nodes that are not marked as nodes to be updated and their output values ​​unchanged to complete the local incremental update.

[0205] The final ranking score for generating candidate utility tunnel deformation schemes is as follows:

[0206] The gradient boosting tree output after local incremental updates is input into the output sorting layer;

[0207] The original scores of each candidate utility tunnel deformation scheme are normalized to obtain the standard ranking score.

[0208] A power correction factor is generated based on the remaining power in the real-time status information, and a mission correction factor is generated based on the mission priority in the flight mission information.

[0209] The standard ranking score, power correction factor, and task correction factor are weighted and fused to obtain the final ranking score for each candidate tunnel deformation scheme.

[0210] Candidate utility tunnel deformation schemes are ranked from highest to lowest according to their final scores.

[0211] The candidate pipe gallery deformation schemes are ranked according to the final ranking score. Combining flight mission information, flight performance parameters and real-time status information, the candidate pipe gallery deformation scheme with the highest ranking score is selected as the target trajectory pipe gallery. Locking domain update and deformation domain update are performed on the target trajectory pipe gallery, and the trajectory planning results for low-altitude UAV flight control are output.

[0212] refer to Figure 2 A trajectory planning system for low-altitude unmanned aerial vehicles, comprising:

[0213] The information acquisition module is used to acquire flight mission information, flight performance parameters, and real-time status information;

[0214] The state space construction module is used to construct a hierarchical low-altitude state space based on low-altitude environmental data and extract safety anchor points from it.

[0215] The retraction candidate area generation module is used to generate layered retraction flow tubes based on safety anchor points, flight performance parameters and real-time status information, and superimpose the layered low-altitude state space to obtain retraction trajectory candidate areas.

[0216] The utility tunnel generation and verification module is used to generate candidate four-dimensional elastic trajectory utility tunnels in the reversible trajectory candidate area based on flight mission information, and to perform comprehensive constraint verification.

[0217] The deformation feature extraction module is used to divide the verified four-dimensional elastic trajectory pipe gallery into a locking domain and a deformation domain, generate candidate pipe gallery deformation schemes, and extract pipe gallery ranking features.

[0218] The sorting output module is used to input the pipe gallery sorting features into the improved LambdaMART model, determine the target trajectory pipe gallery, update the locking domain and deformation domain, and output the trajectory planning results.

[0219] Example 1: In an anonymized urban low-altitude delivery simulation scenario, a drone needs to fly from a takeoff point to a target point. The horizontal flight range is 1600m × 1200m, with a flight altitude of 60m to 150m. There are two waypoints. The area includes 86 buildings, 3 fixed tower cranes, 12 power transmission towers, 2 temporary restricted flight zones, 14 dynamic low-altitude targets, and a communication obstruction corridor. The drone has 88% battery power at takeoff, a maximum speed of 12m / s, a maximum climb rate of 3m / s, a maximum descent rate of 2.5m / s, and a minimum turning radius of 12m. The mission requires completing the flight within 160s and maintaining a safety margin of no less than 20% battery power in abnormal situations.

[0220] After acquiring flight mission information, flight performance parameters, and real-time status information, the system divides the low-altitude flight area into 43,200 low-altitude spatial cells with a horizontal resolution of 20m×20m and an altitude interval of 10m. For each cell, the system writes its spatial location, altitude range, time stamp, obstacle occupancy status, dynamic occupancy status, airspace clearance status, communication reliability status, navigation reliability status, and emergency response status. After matching building outline data with the cells, 9,240 cells were marked as obstacle occupancy; 628 additional occupancy cells were added for fixed tower cranes and power transmission towers; terrain elevation verification excluded 310 low-altitude cells. During dynamic occupancy prediction, the system reads the trajectory points of 14 dynamic low-altitude targets, extrapolates to the next 15 seconds with a 1s step size, and sets the occupancy radius to 10m, marking a total of 1,365 dynamic occupancy cells. The wind field model detected a strong disturbance area with a sudden increase in wind speed to 13.2m / s; 482 cells corresponding to this area were written into the dynamic occupancy status.

[0221] In the airspace rules layer, after mapping two temporary restricted flight zones to cells, 2870 cells were marked as airspace impassable. In the communication trust layer, ray tracing was used to calculate the received signal strength at the center point of each cell. 4216 cells with signal strength below -110dBm or a signal-to-noise ratio below -5dB were marked as communication untrusted. In the navigation trust layer, the system determined availability based on satellite navigation multipath simulation and visual-inertial fusion. 3518 cells with a 3D positioning error exceeding 5m and fewer than 200 visual feature points were marked as navigation untrusted.

[0222] The system forms an emergency response layer. Within the test area, 18 landable platforms, 6 open landing areas, 8 backup altitude level entrances, and 5 return access points were pre-positioned, initially corresponding to 147 candidate cells for safe anchor points. The system checked the status of these candidate cells one by one in the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication trust layer, and navigation trust layer, eliminating 31 obstacle occupancy cells, 18 dynamic occupancy cells, 22 airspace impassable cells, 17 communication untrusted cells, and 9 navigation untrusted cells, ultimately retaining 50 emergency response cells, which, after connectivity clustering, formed 11 safe anchor points.

[0223] The system performs Hamilton-Jacobi reachability calculations using 11 safety anchors as the backpropagation source. Each cell containing a safety anchor is initially written with a pullback time of 0 seconds and a pullback power consumption of 0%, and then added to the backpropagation queue. During queue expansion, the system searches layer by layer along the horizontal eight-adjacency direction, the vertical altitude layer direction, and the reverse time direction. If a candidate cell simultaneously satisfies the conditions for passage, communication, navigation, and dynamics, a reachability marker is written. After the first expansion, 2680 reachable cells were obtained; after the fifth expansion, 12430 were obtained; and after the tenth expansion, 21960 were obtained. The system then calculates the pullback power consumption and pullback time based on the power consumption rates of level flight (0.014% / m), climb (0.19% / m), descent (0.08% / m), and hovering (0.11% / s). The calculation results show that 4870 reachable cells, although spatially reachable, had less than 20% remaining power after retraction and were therefore deleted; another 2135 cells had a retraction time exceeding 45 seconds and were also deleted. Ultimately, 14955 cells were retained to form an energy-time flow field, which was then connected into 23 layered retraction flow tubes according to the retraction direction, retraction time sequence, and altitude layer switching sequence. After superimposing the layered retraction flow tubes with the layered low-altitude state space, 11842 candidate cells for retraction trajectories were obtained.

[0224] Within the reversible trajectory candidate area, the system generates candidate four-dimensional flexible trajectory corridors based on the takeoff point, target point, and two waypoints. Each low-altitude spatial cell is expanded according to the UAV's body size, a safety interval of 2m, a positioning error of 0.5m, and a control error of 0.8m, generating a flexible cell with a three-dimensional spatial range. Each flexible cell has a time window that can be entered 3s in advance or 8s in later, based on the expected passage time. The system generates 56 candidate four-dimensional flexible trajectory corridors according to different altitude layers, lateral offset directions, time window combinations, and waypoint connection sequences. During the comprehensive constraint verification, 13 corridors failed due to obstacle occupancy, 9 failed due to dynamic occupancy, 6 failed due to unreliable communication, 5 failed due to unreliable navigation, and 7 failed due to more than 4 consecutive flexible cells not matching the layered reversible flow tube. 16 candidate four-dimensional flexible trajectory corridors are retained.

[0225] At the 67th second of the flight simulation, the UAV's current position was approximately 610m from the target point. A dynamic flying target appeared within the forward square variable domain, crossing the original pipe corridor, and was predicted to enter the originally planned passage area in 13 seconds. Simultaneously, the communication strength in this area decreased from -96dBm to -112dBm, and the navigation error increased from 3.4m to 6.1m. The system designated 17 elastic cells within 120m of the UAV's current position or expected to arrive within 20 seconds as the locking domain; the remaining 92 elastic cells were designated as the deformation domain. Within the deformation domain, the system identified 11 cells with dynamic occupancy changes, 19 cells with communication changes, 14 cells with navigation changes, 7 cells with wind field changes, and 16 cells with traffic flow changes. It also generated 21 candidate tunnel deformation schemes, including combinations such as lateral migration of 5m, lateral migration of 10m, rising one height level, rising two height levels, shifting the time window backward by 6s, lateral migration of 10m and shifting backward by 6s, and rising one height level and shifting backward by 6s.

[0226] When training the improved LambdaMART model, the system used 6200 sets of anonymized simulation samples, including 4960 sets of training samples, 620 sets of validation samples, and 620 sets of test samples. Each set of samples contained at least 6 candidate tunnel deformation schemes, and the input features included path length, expected energy consumption, overall risk, obstacle safety distance, backoff margin, communication reliability, navigation reliability, tunnel elasticity margin, time window width, and locking stability. In the typical training samples, Scheme A involves a lateral migration of 10m, a path length of 956m, an estimated power consumption of 13.8%, a nearest obstacle distance of 18.6m, a retreat time of 34s, a communication strength of -108dBm, a navigation error of 4.8m, and a manual sorting label of 3; Scheme B involves ascending one height level and moving backwards for 6s, a path length of 982m, an estimated power consumption of 14.4%, a nearest obstacle distance of 27.5m, a retreat time of 26s, a communication strength of -101dBm, a navigation error of 4.1m, and a manual sorting label of 4; Scheme C involves maintaining the original utility tunnel, a path length of 914m, an estimated power consumption of 12.9%, a nearest obstacle distance of 8.2m, a retreat time of 48s, a communication strength of -114dBm, a navigation error of 6.3m, and a manual sorting label of 0. After model inference, Scheme B scored 0.91, Scheme A scored 0.78, and Scheme C scored 0.22. The system selected Scheme B as the target trajectory tunnel and wrote the updated 19 elastic cells at the proximal end into the locking domain, while the remaining cells were written into the deformation domain.

[0227] Under the same simulation conditions, the method of this invention is compared with the traditional 3D path search plus local obstacle avoidance method. The traditional method takes 2.6s for initial planning, while this invention takes 3.5s; the adjustment time after dynamic disturbance is 5.8s for the traditional method, and 1.2s for this invention; the traditional method replans 4 times throughout the process, while this invention replans only once; the traditional method enters 23 irreversible cells, while this invention enters 0; the traditional method has a minimum obstacle distance of 7.6m, while this invention has 24.3m; the traditional method has a minimum communication strength of -114dBm, while this invention has -101dBm; the traditional method has a maximum navigation error of 6.7m, while this invention has 4.1m; the traditional method's time to retreat to the nearest safe anchor point is 48s, while this invention's is 26s; the traditional method's power reserve after retreat is 9.1%, while this invention's is 18.4%; the traditional method's task completion time is 119s, while this invention's is 124s; the traditional method's total power consumption is 13.2%, while this invention's is 14.4%.

[0228] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A trajectory planning method for low-altitude unmanned aerial vehicles (UAVs), characterized in that, include: Acquire flight mission information, flight performance parameters, and real-time status information of low-altitude UAVs; A hierarchical low-altitude state space is constructed based on low-altitude environmental data, and safety anchor points that meet the conditions of airspace clearance, communication, landing, or return access are extracted. Using each safety anchor point as the source point, and combining flight performance parameters and real-time status information, a layered retraction flow tube is generated by combining Hamilton-Jacobi reachability domain calculation with energy-time flow field. The layered retraction flow tube is then superimposed with the layered low-altitude state space to obtain a candidate area for retraction trajectory. Based on flight mission information, multiple candidate four-dimensional elastic trajectory corridors are generated in the reversible trajectory candidate area. Each four-dimensional elastic trajectory corridor is composed of multiple elastic cells with three-dimensional spatial range and time passage window. Comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory corridor, and candidate four-dimensional elastic trajectory corridors that do not meet the conditions are eliminated. Based on real-time status information, the selected four-dimensional elastic trajectory utility tunnel is divided into a locking domain and a deformation domain. Multiple candidate utility tunnel deformation schemes are generated for the deformation domain, and the utility tunnel ranking features corresponding to each candidate utility tunnel deformation scheme are extracted. The ranking features of the utility tunnel are input into the improved LambdaMART model, and the ranking scores of each candidate utility tunnel deformation scheme are output. Combining flight mission information, flight performance parameters and real-time status information, the candidate utility tunnel deformation scheme that meets the conditions and has the highest ranking score is selected as the target trajectory utility tunnel. The target trajectory utility tunnel is updated in both the locking domain and the deformation domain, and the trajectory planning results are output.

2. The trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The flight mission information includes the starting point location, target point location, mission waypoint, flight altitude range, mission priority, and airspace restriction information. Flight performance parameters include maximum flight speed, maximum acceleration, maximum rate of climb, maximum rate of descent, minimum turning radius, and hovering power consumption rate. Real-time status information includes the UAV's current position, current speed, current heading, remaining battery power, communication link quality, and navigation and positioning reliability.

3. The trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The construction of a hierarchical low-altitude state space based on low-altitude environmental data includes: Acquire low-altitude environmental data and process it for coordinate unification, time unification, and altitude benchmark unification. The low-altitude environmental data includes low-altitude three-dimensional map data, building outline data, static obstacle data, dynamic obstacle data, airspace restriction data, communication link coverage data, navigation and positioning quality data, meteorological wind field data, and landable area data. According to the preset spatial resolution and preset altitude interval, the low-altitude flight area is divided into multiple low-altitude spatial cells, and a cell attribute record is established for each low-altitude spatial cell. The cell attribute record includes spatial location, altitude range, time marker, obstacle occupancy status, dynamic occupancy status, airspace passage status, communication reliability status, navigation reliability status, and emergency response status. Based on building outline data, static obstacle data, and terrain height data, each low-altitude space cell is assigned an obstacle occupancy status to form an obstacle occupancy layer; based on dynamic obstacle data, meteorological wind field data, and low-altitude traffic flow data, the dynamic occupancy status of each low-altitude space cell under different time markers is predicted to form a dynamic occupancy prediction layer. Based on airspace restriction data, each low-altitude space cell is assigned an airspace access status, forming an airspace rule layer; based on communication link coverage data, each low-altitude space cell is assigned a communication trust status, forming a communication trust layer; based on navigation and positioning quality data, each low-altitude space cell is assigned a navigation trust status, forming a navigation trust layer. Based on the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication trust layer, and navigation trust layer, emergency response status is marked for each low-altitude space cell to form an emergency response layer. Low-altitude space cells that meet the clearance conditions, communication conditions, landing conditions, or return access conditions are extracted from the emergency response layer as safety anchor points.

4. The trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The process of obtaining the candidate region for the reversible trajectory includes: Read the safety anchor point, flight performance parameters and real-time status information, and take each safety anchor point as the terminal state of the UAV’s abnormal retreat. Based on the UAV’s current position, current speed, current heading, current altitude, remaining battery power, maximum flight speed, maximum acceleration, maximum climb rate, maximum descent rate, minimum turning radius and hovering power consumption rate, determine the state transition range of the UAV when it retreats from the low-altitude space cell to the corresponding safety anchor point. Using each safety anchor point as the starting point for reverse calculation, Hamilton-Jacobi reachability calculation is performed along the spatial adjacency direction, altitude layer switching direction and time marker reversal direction in the layered low-altitude state space to determine whether each low-altitude spatial cell marker can reach the corresponding safety anchor point within the state transition range limited by the UAV flight performance parameters. For low-altitude space cells marked as reachable by Hamilton-Jacobi reachability domain calculation, the remaining power in real-time status information and hovering power consumption rate, climb power consumption characteristics, descent power consumption characteristics and level flight power consumption characteristics in flight performance parameters are combined to calculate the power consumption and time consumption required for the UAV to retreat from the low-altitude space cell to the corresponding safe anchor point. An energy-time flow field is formed based on the power consumption and time consumption results. According to the retraction direction, retraction time sequence, retraction power consumption sequence, and altitude layer switching sequence in the energy-time flow field, the low-altitude space cells that continuously satisfy the retraction reachability relationship are connected into a layered retraction flow tube corresponding to each safety anchor point. The layered retraction flow tubes are superimposed with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication reliability layer, navigation reliability layer, and emergency response layer in the layered low-altitude state space. Flow tube segments that are inconsistent with the obstacle occupancy state, dynamic occupancy state, airspace passage state, communication reliability state, navigation reliability state, and emergency response state are deleted. The low-altitude space cells covered by the remaining layered retraction flow tubes after deletion are determined as candidate areas for retraction trajectories.

5. A trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The process of generating multiple candidate four-dimensional elastic trajectory corridors within the reversible trajectory candidate region includes: Based on the starting point, target point, mission waypoint, flight altitude range and mission time requirements in the flight mission information, low-altitude space cells connected to the starting point, target point and mission waypoint are selected from the candidate area of ​​the reversible trajectory to form the candidate corridor search range. Based on the UAV's body size, safety interval, positioning error, and low-altitude flight control accuracy, the low-altitude spatial cells within the candidate pipe gallery search range are spatially expanded and time windows are assigned to generate elastic cells with three-dimensional spatial range and time passage windows. Based on the starting point, target point, mission waypoint, and flight altitude range, adjacent elastic cells with continuous passage time are connected. Multiple candidate four-dimensional elastic trajectory corridors are generated according to different altitude layers, different lateral offset directions, different time passage windows, and different waypoint connection sequences. A comprehensive constraint verification is performed on each candidate four-dimensional elastic trajectory tunnel. The comprehensive constraint verification includes matching the elastic cells in each candidate four-dimensional elastic trajectory tunnel with the obstacle occupancy layer, dynamic occupancy prediction layer, airspace rule layer, communication credibility layer, navigation credibility layer and layered backflow tube, respectively, to obtain the passage verification result of each candidate four-dimensional elastic trajectory tunnel. Based on the verification results, candidate four-dimensional elastic trajectory utility tunnels that pass the comprehensive constraint verification are retained, while candidate four-dimensional elastic trajectory utility tunnels that fail the comprehensive constraint verification are removed.

6. The trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The process involves generating multiple candidate utility tunnel deformation schemes in the deformation domain and extracting the utility tunnel ranking features corresponding to each candidate deformation scheme, including: Read the four-dimensional elastic trajectory tunnel that has been saved after comprehensive constraint verification, and determine the current position, current speed, current heading, current altitude and current time of the UAV in the four-dimensional elastic trajectory tunnel based on the real-time status information; Based on the spatial distance between the current position of the UAV and each elastic cell in the four-dimensional elastic trajectory tunnel, as well as the estimated time for the UAV to reach each elastic cell, the four-dimensional elastic trajectory tunnel is divided into a locking domain and a deformation domain. The system acquires dynamic obstacle change information, communication link change information, navigation reliability change information, wind field disturbance change information, and low-altitude traffic flow change information. It then maps the change information to elastic cells in the deformation domain to obtain the dynamic occupancy change state, communication change state, navigation change state, wind field change state, and traffic flow change state of each elastic cell in the deformation domain. Based on the dynamic occupancy, communication, navigation, wind field, and traffic flow changes of each elastic cell within the deformation domain, multiple candidate tunnel deformation schemes are generated for the deformation domain. For each candidate utility tunnel deformation scheme, the utility tunnel ranking features are extracted. These features include path length features, expected energy consumption features, comprehensive risk features, obstacle safety distance features, backoff margin features, communication reliability features, navigation reliability features, utility tunnel elasticity margin features, time window width features, and locking stability features.

7. The trajectory planning method for low-altitude unmanned aerial vehicles according to claim 1, characterized in that, The process of updating the locking domain and deformation domain of the target trajectory tunnel and outputting the trajectory planning results includes: An improved LambdaMART model is constructed, which includes a shared input layer, a task attention channel, a gradient boosting tree channel, a fast incremental update layer, and an output sorting layer. The shared input layer is connected to the task attention channel, the task attention channel is connected to the gradient boosting tree channel, the gradient boosting tree channel is connected to the fast incremental update layer, and the fast incremental update layer is connected to the output sorting layer. The candidate tunnel deformation schemes are input into the shared input layer along with the tunnel sorting features, flight mission information, flight performance parameters, and real-time status information. Field mapping and numerical normalization are performed on path length, estimated energy consumption, comprehensive risk value, obstacle safety distance, backoff margin, communication reliability, navigation reliability, tunnel elasticity margin, time window width, and locking stability. The results are then concatenated to form the input feature vector. The input feature vector is fed into the task attention channel. Through a fully connected feature selection layer and a multi-head attention layer, the backoff margin feature, communication credibility feature and navigation credibility feature are weighted and enhanced, and the task-weighted feature vector is output. The task-weighted feature vector is input into the gradient boosting tree channel. An improved LambdaMART backbone network, formed by serially stacking multiple gradient boosting decision trees, is used to rank and learn the candidate tube gallery deformation schemes. A feature resampling layer is inserted into the gradient boosting tree channel to resample the high-dimensional sparse features. The ranking results output from the gradient boosting tree channel are input into the fast incremental update layer. Based on the changes in dynamic obstacles, communication links, and navigation confidence, local incremental updates are performed on the affected gradient boosting decision tree nodes. The updated ranking results are input into the output ranking layer to generate the final ranking score of the candidate tunnel deformation scheme. The candidate pipe gallery deformation schemes are ranked according to the final ranking score. Combining flight mission information, flight performance parameters and real-time status information, the candidate pipe gallery deformation scheme with the highest ranking score is selected as the target trajectory pipe gallery. Locking domain update and deformation domain update are performed on the target trajectory pipe gallery, and the trajectory planning results for low-altitude UAV flight control are output.

8. A trajectory planning system for low-altitude unmanned aerial vehicles (UAVs), comprising executing the trajectory planning method for low-altitude UAVs as described in any one of claims 1 to 7, characterized in that, include: The information acquisition module is used to acquire flight mission information, flight performance parameters, and real-time status information; The state space construction module is used to construct a hierarchical low-altitude state space based on low-altitude environmental data and extract safety anchor points from it. The retraction candidate area generation module is used to generate layered retraction flow tubes based on safety anchor points, flight performance parameters and real-time status information, and superimpose the layered low-altitude state space to obtain retraction trajectory candidate areas. The utility tunnel generation and verification module is used to generate candidate four-dimensional elastic trajectory utility tunnels in the reversible trajectory candidate area based on flight mission information, and to perform comprehensive constraint verification. The deformation feature extraction module is used to divide the verified four-dimensional elastic trajectory pipe gallery into a locking domain and a deformation domain, generate candidate pipe gallery deformation schemes, and extract pipe gallery ranking features. The sorting output module is used to input the pipe gallery sorting features into the improved LambdaMART model, determine the target trajectory pipe gallery, update the locking domain and deformation domain, and output the trajectory planning results.

Images