Multi-source data fusion unmanned aerial vehicle path planning method for emergency rescue

Abstract

The application discloses an emergency rescue-oriented multi-source data fusion unmanned plane path planning method, relates to the technical field of unmanned plane path planning and emergency rescue, and comprises the following steps: acquiring real-time terrain undulation and temporary obstacle distribution information from a mountain disaster area environment by a multi-source sensor array carried by an unmanned plane; adopting a grid division technology to perform spatial block processing on collected terrain data, integrating the terrain data into a structured terrain grid data set, and determining a preliminary unmanned plane flight environment spatial distribution; the emergency rescue-oriented multi-source data fusion unmanned plane path planning method can generate a safe, efficient and continuous unmanned plane rescue route by multi-target optimization and historical data reference, improves the response speed and overall efficiency of a rescue task, and provides a reliable unmanned plane cooperative passing solution for emergency rescue in remote mountain areas.

Description

Technical Field

[0001] This invention relates to the field of UAV path planning and emergency rescue technology, specifically to a multi-source data fusion UAV path planning method for emergency rescue. Background Technology

[0002] In the field of disaster relief and emergency management, especially in remote mountainous disaster areas, efficient drone path planning technology is crucial for the rapid deployment of rescue resources and the safety of personnel. When drones perform rescue missions in disaster areas, they need to quickly select feasible routes in complex terrain environments, while simultaneously dealing with constantly changing terrain and temporary obstacles such as landslides, rockfalls, or destroyed roads. These factors make path planning for drones during rescue missions extremely challenging.

[0003] While some existing UAV path planning methods can plan flight routes based on a single data source, they have significant shortcomings in multi-source data fusion, real-time environmental perception, and handling complex terrain in disaster areas. First, disaster area environmental data is diverse and voluminous, including terrain elevation, obstacle distribution, road damage, and disaster situation changes; a single data source cannot comprehensively reflect the actual environmental conditions. Second, existing methods often struggle to quickly generate safe and efficient flight routes in areas with complex terrain and frequent obstacles, and there is a lack of effective coordination between global flight route planning and local obstacle avoidance adjustments. Furthermore, the dynamic nature of the disaster area environment further increases the difficulty of path planning. For example, sudden rockfalls or terrain collapses may render pre-planned routes impassable, and existing methods have limited real-time monitoring and dynamic adjustment capabilities, failing to quickly correct paths and impacting UAV rescue efficiency. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-source data fusion UAV path planning method for emergency rescue, and to solve the problems existing in the prior art.

[0005] To achieve the above objectives, this invention provides the following technical solution: a multi-source data fusion UAV path planning method for emergency rescue, comprising: S1, acquiring real-time terrain undulation and temporary obstacle distribution information from the mountain disaster area environment using a multi-source sensor array mounted on the UAV, spatially dividing the collected terrain data using grid partitioning technology, integrating it into a structured terrain grid dataset, and determining the preliminary spatial distribution of the UAV flight environment; S2, based on the structured terrain grid dataset, selecting key flight nodes and obstacle-dense areas using a node priority ranking method, and extracting features from the grid data in conjunction with terrain complexity analysis to obtain a simplified terrain feature representation; S3, if the simplified terrain feature representation contains obstacle-dense areas... If the proportion of the clustered area exceeds a preset threshold, the newly added terrain change information is obtained from the sensor array on the UAV through a real-time data update mechanism. Combined with dynamic obstacle avoidance technology, the feature representation is locally corrected to determine the updated terrain feature distribution. S4. For the updated terrain feature distribution, the path cost calculation method is used to search for flight nodes in the global range. Combined with distance estimation method and obstacle weight allocation technology, the feasibility of potential routes is evaluated to obtain a preliminary global flight route. S5. Through the preliminary global flight route, a detailed terrain sub-map of the local risk area is obtained. The route feasibility score and time consumption prediction technology are used to prioritize the branches in the sub-map to determine the optimized local flight segments.

[0006] Preferably, step S1 includes acquiring raw point cloud data and multispectral images collected by a multi-source sensor array carried by the UAV; performing feature fusion on the raw point cloud data and multispectral images to extract real-time terrain undulation features and temporary obstacle distribution features; using grid partitioning to spatially divide the real-time terrain undulation features and temporary obstacle distribution features into discrete terrain grid units; logically associating the discrete terrain grid units according to a preset topological relationship to form a structured terrain grid dataset; and performing three-dimensional reconstruction based on the structured terrain grid dataset to determine the preliminary spatial distribution of the UAV flight environment.

[0007] Preferably, step S2 includes obtaining the node elevation gradients and neighborhood connectivity of the structured terrain grid dataset to determine the terrain grid node feature matrix; based on the terrain grid node feature matrix, using a node priority ranking algorithm to evaluate the weights of the grid nodes and determine the set of key flight nodes; performing a difference operation on the set of key flight nodes and the terrain grid node feature matrix to obtain a spatial distribution map of densely obstructed areas; determining a terrain complexity distribution model based on a weighted fusion of the spatial distribution map of densely obstructed areas and the terrain grid node feature matrix; and using the terrain complexity distribution model to perform feature dimensionality reduction processing on the terrain grid nodes to obtain a simplified terrain feature representation.

[0008] Preferably, step S3 includes extracting the regional proportion of densely obstructed areas from the simplified terrain features; if the regional proportion exceeds a judgment threshold, acquiring new terrain change information through a sensor array; determining the adjustment range of the local topology based on the terrain change information, the adjustment range being used to guide the correction operator to locally correct the simplified terrain features; reconstructing the simplified terrain features using the correction operator to obtain a feature mapping model, and performing a global consistency check on the feature mapping model to determine the updated terrain feature distribution.

[0009] Preferably, step S4 includes obtaining a global flight node set based on terrain feature distribution, the global flight node set containing the spatial coordinates and connectivity data of each node; traversing the global flight node set using a path cost calculation method to determine the path cost values ​​between each node; generating a distance weight distribution for each potential flight route based on the path cost values ​​and a preset distance estimation method; conducting a feasibility assessment of the potential flight routes by combining the distance weight distribution and obstacle weight allocation techniques; and selecting the optimal path based on the feasibility assessment results to obtain a preliminary global flight route.

[0010] Preferably, step S5 includes extracting terrain feature data based on the geospatial extent of the global flight route, constructing a terrain sub-map set for local risk areas, determining the risk level value within each sub-map through spatial overlap analysis of terrain complexity and dynamic obstacles within the terrain sub-map set, and if the risk level value exceeds a preset threshold, then weighting the branches within the terrain sub-map set using route connectivity and meteorological interference to obtain the route feasibility score for each branch, predicting the time consumption of the branches based on the mapping relationship between the route feasibility score and the flight energy consumption ratio, and determining the flight time series for each branch, and prioritizing the branches through correlation analysis between the flight time series and airspace congestion to obtain the priority sequence for each branch, and reconstructing the local flight segments based on the priority sequence to obtain optimized local flight segments, thereby achieving precise planning of local flight segments.

[0011] Preferably, it also includes S6: If the optimized local flight segment detects falling rocks or terrain collapse during real-time monitoring, relevant change features are extracted from the terrain feature distribution, and the local flight segment is dynamically adjusted in combination with path deviation penalty and local optimization inspection technology to determine the corrected local flight path. Specifically, this includes acquiring real-time point cloud data of the local flight segment and spatially registering it with a preset terrain benchmark model; determining the geometric boundaries of terrain collapse areas in the point cloud data; extracting the elevation change gradient and surface roughness change from the terrain feature distribution based on the geometric boundaries; and mapping the change to a path cost matrix to determine the obstacle avoidance cost distribution of the local flight segment.

[0012] Preferably, step S6 further includes calculating the overlap between the current trajectory and the obstacle avoidance cost distribution using a path deviation penalty function; if the overlap exceeds a preset risk threshold, a candidate path set is determined; and the candidate path set is iteratively calculated using a cost function through local optimization inspection to determine the corrected local flight path, thereby realizing the dynamic adjustment of local flight segments.

[0013] Preferably, it also includes S7, which involves using path smoothing and multi-objective optimization techniques to fuse and adjust the two based on the spatial matching degree between the corrected local flight path and the global flight path, and refining the final route by combining historical data references to obtain a UAV collaborative passage plan for mountain rescue. Specifically, this includes acquiring elevation data and local flight path trajectories of mountainous areas, constructing a three-dimensional terrain grid model, projecting the local flight path onto the surface of the three-dimensional terrain grid model to obtain a spatial matching degree matrix, and using a cubic spline interpolation algorithm to smooth the local flight path based on the spatial matching degree matrix to generate corrected flight path trajectory data.

[0014] Preferably, step S7 further includes inputting the corrected flight path data into a multi-objective optimization function, determining the optimal cooperative travel path coordinate sequence through a genetic algorithm, associating and matching the optimal cooperative travel path coordinate sequence with historical data, and obtaining a UAV cooperative travel plan for mountain rescue by correcting the offset.

[0015] As can be seen from the above technical solution, the present invention has the following beneficial effects: This multi-source data fusion UAV path planning method for emergency rescue utilizes a multi-source sensor array to acquire real-time terrain and temporary obstacle information in disaster areas. Combined with structured terrain grids, global and local flight path planning, and dynamic adjustment techniques, this method enables efficient UAV path planning in remote mountainous disaster areas. This method not only effectively handles complex and dynamically changing environmental information, improving the real-time performance and accuracy of UAV flight path generation, but also achieves coordinated optimization between global flight paths and local obstacle avoidance, significantly reducing the risk of flight path infeasibility due to sudden terrain changes or obstacles. Furthermore, through multi-objective optimization and historical data reference, this method can generate safe, efficient, and continuous UAV rescue flight paths, improving the response speed and overall efficiency of rescue missions, and providing a reliable UAV collaborative access solution for emergency rescue in remote mountainous areas. Attached Figure Description

[0016] Figure 1 This is a flowchart of the multi-source data fusion UAV path planning method of the present invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] like Figure 1 As shown, this invention provides a technical solution: a multi-source data fusion UAV path planning method for emergency rescue, including: S1, acquiring real-time terrain undulation and temporary obstacle distribution information from the mountain disaster area environment through a multi-source sensor array on the UAV, spatially dividing the collected terrain data into blocks using grid partitioning technology, integrating them into a structured terrain grid dataset, and determining the preliminary spatial distribution of the UAV flight environment; S2, based on the structured terrain grid dataset, selecting key flight nodes and obstacle-dense areas using a node priority ranking method, and extracting features from the grid data in conjunction with terrain complexity analysis to obtain a simplified terrain feature representation; S3, if the proportion of obstacle-dense areas in the simplified terrain feature representation exceeds a preset threshold, acquiring new terrain change information from the sensor array on the UAV through a real-time data update mechanism, and locally correcting the feature representation in conjunction with dynamic obstacle avoidance technology to determine the updated terrain feature distribution; S4, for the updated terrain... The system first identifies the characteristic distribution and then uses path cost calculation to search for flight nodes globally. It then combines distance estimation and obstacle weight allocation techniques to assess the feasibility of potential flight routes, resulting in a preliminary global flight path. Next, based on the preliminary global flight path, detailed terrain sub-maps of local risk areas are obtained. Route feasibility scoring and time consumption prediction techniques are used to prioritize branches within the sub-maps, determining optimized local flight segments. If, during real-time monitoring, the optimized local flight segment detects falling rocks or terrain collapse, relevant change features are extracted from the terrain feature distribution. Combined with path deviation penalties and local optimization inspection techniques, the local flight segment is dynamically adjusted to determine the corrected local flight path. Finally, based on the spatial matching degree between the corrected local flight path and the global flight path, path smoothing and multi-objective optimization techniques are used to fuse and adjust the two. Historical data is then used to refine the final route, resulting in a UAV collaborative passage plan for mountain rescue operations.

[0019] In this embodiment, the method is used for UAV path planning in emergency rescue scenarios in mountainous disaster areas. Mountainous disaster areas present problems such as undulating terrain, road blockages, rockfalls, landslides, fallen trees, temporary rescue facilities occupying airspace, and communication obstruction. The UAV acquires real-time environmental data through a multi-source sensor array, converting the continuous mountainous space into a structured terrain grid dataset. Then, through key node filtering, obstacle-dense area identification, real-time terrain updates, global route search, local flight segment optimization, and route fusion processing, a UAV collaborative passage plan for rescue missions is generated.

[0020] The UAV's multi-source sensor array consists of a 32-line LiDAR, an RGB visible light camera, an infrared thermal imaging camera, a millimeter-wave radar, an inertial measurement unit (IMU), a barometric altimeter, and an RTK satellite positioning module. The 32-line LiDAR acquires point cloud data of the mountain surface and obstacle contour data at a sampling frequency of 10Hz; the RGB visible light camera acquires surface images at 30 frames per second to identify fallen rocks, fallen trees, rescue vehicles, and temporary facilities; the infrared thermal imaging camera acquires heat source images at 25 frames per second to identify heat sources from people in nighttime, smoky, and low-light environments; the millimeter-wave radar detects obstacles in rain, fog, and dusty environments at a sampling frequency of 20Hz; the IMU acquires the UAV's attitude angles, angular velocities, and linear acceleration; the barometric altimeter acquires the flight altitude; and the RTK satellite positioning module acquires the UAV's centimeter-level spatial coordinates. All sensor data are timestamped and uniformly converted to the same local northeast-northeast coordinate system.

[0021] In S1, a grid-based approach is used to spatially divide the mountainous disaster area environment into blocks. The area to be planned is divided into a three-dimensional voxel grid, with each voxel having a length, width, and height of 2 meters. Each grid cell records elevation, slope, terrain roughness, obstacle occupancy probability, obstacle height, sensor confidence level, and update timestamp. The elevation is determined by the mean Z-axis coordinate of the LiDAR point cloud; the slope is determined by the ratio of the elevation difference between adjacent grids to the horizontal distance; the terrain roughness is determined by the elevation variance of the point cloud within the grid; and the obstacle occupancy probability is obtained by fusing LiDAR point cloud density, image recognition results, and millimeter-wave radar echo intensity. When the obstacle height within a grid is greater than 0.5 meters and the obstacle occupancy probability is not less than 0.65, the grid is marked as an obstacle grid. Through this gridding process, the system obtains a structured terrain grid dataset and forms a preliminary spatial distribution of the UAV flight environment.

[0022] The node priority ranking method in S2 is used to select critical flight nodes from the structured terrain grid dataset. Critical flight nodes include rescue starting points, target points, communication relay points, safe hovering points, valley passage entrances, obstacle boundary points, and risk area connection points. The system calculates a priority score P for each candidate node, which is a weighted average of the target distance score, obstacle density score, flight altitude margin score, communication quality score, and node connectivity score. The target distance score has a weight of 0.30, obstacle density a weight of 0.25, flight altitude margin a weight of 0.20, communication quality a weight of 0.15, and node connectivity a weight of 0.10. The system ranks candidate nodes from highest to lowest P value and selects the top 30% of candidate nodes as critical flight nodes.

[0023] Terrain complexity analysis is used to extract key features from grid data. The system calculates the terrain complexity T for each grid cell, where T is composed of the elevation change rate, slope change rate, obstacle occupancy probability, and traversable space continuity. The elevation change rate represents the intensity of mountain undulation; the slope change rate represents the degree of terrain abruptness; the obstacle occupancy probability represents the risk of a grid cell being occupied by obstacles; and traversable space continuity represents whether adjacent grid cells can form a continuous flight path. When the number of obstacle grid cells in a certain area reaches 40% of the total number of grid cells in that area, the area is marked as an obstacle-dense area. Through node priority ranking and terrain complexity analysis, the original terrain grid data is transformed into a simplified terrain feature representation that includes key nodes, obstacle-dense areas, traversable boundaries, and risk levels.

[0024] In S3, the preset threshold is set to 35%. The system calculates the area ratio R of densely obstructed regions in the simplified terrain feature representation, where R is equal to the ratio of the number of grid cells in densely obstructed regions to the total number of grid cells. When R is greater than 35%, the system triggers a real-time data update mechanism. The UAV reacquires new terrain change information through a multi-source sensor array, with an update cycle set to 0.5 seconds. New terrain change information includes newly appearing rocks, newly formed collapse boundaries, areas occupied by temporary rescue equipment, locations of fallen trees, and areas obscured by smoke and dust. The system maps the new change information to the original terrain grid dataset and updates the obstacle occupancy probability, obstacle height, terrain roughness, and update timestamp of the affected grid cells.

[0025] Dynamic obstacle avoidance technology is used to locally correct the simplified terrain feature representation. The system establishes a local correction zone with a radius of 8 meters centered on the center point of the newly added obstacle, and recalculates the grid occupancy state and passage cost within this zone. For rockfall obstacles, the system extends the rockfall boundary by 2 meters as a safety buffer; for terrain collapse areas, the system extends the collapse boundary by 5 meters as a no-fly buffer; for fallen tree areas, the system extends the outer edge of the tree canopy by 3 meters as an obstacle avoidance buffer. After local correction, the system obtains the updated terrain feature distribution.

[0026] The path cost calculation method in S4 is used to search for flight nodes globally. The system treats key flight nodes as nodes in a graph structure, candidate flight segments between adjacent passable nodes as edges, and calculates a comprehensive path cost C for each candidate flight segment. The comprehensive path cost C consists of distance cost Cd, altitude change cost Ch, obstacle risk cost Co, time cost Ct, and communication cost Cc, where Cd has a weight of 0.30, Ch has a weight of 0.20, Co has a weight of 0.30, Ct has a weight of 0.10, and Cc has a weight of 0.10. The distance estimation method uses three-dimensional Euclidean distance, calculated from the three-dimensional coordinates of the candidate flight segment's start and end points. The obstacle weight allocation technique assigns a risk weight of 1.0 to rockfall areas, 1.0 to collapse areas, 0.8 to dense forest areas, 0.6 to communication blind spots, and 0.1 to flat passable areas. The system uses the A-star search method to search for the node sequence with the minimum comprehensive path cost C within the global node range, and obtains a preliminary global flight path.

[0027] In S5, the system extracts detailed terrain sub-maps of local risk areas based on the initial global flight path. Local risk areas include rockfall zones, collapse zones, dense forests, narrow valley passages, and communication blind spots located less than 10 meters from the global flight path. The grid side length and altitude resolution of the detailed terrain sub-map are set to 0.5 meters. The system generates multiple candidate routes from the detailed terrain sub-map and calculates a route feasibility score F and a predicted time consumption value E for each candidate route. The route feasibility score F is composed of safety clearance, turning radius, flight altitude margin, communication quality, and sensor observability, with safety clearance having a weight of 0.35, turning radius 0.20, flight altitude margin 0.20, communication quality 0.15, and sensor observability 0.10. The predicted time consumption value E is calculated from the segment length, UAV cruise speed, climb time, descent time, and turn time. The system sorts the candidate routes in descending order of F value and ascending order of E value, and selects the candidate route ranked first as the optimized local flight segment.

[0028] In the S6, the UAV continuously monitors for falling rocks and terrain collapses while executing optimized local flight segments. The system reads lidar point cloud data, millimeter-wave radar echoes, and camera recognition results every 0.2 seconds. When a new falling rock or collapse boundary is detected, the system extracts change features from the terrain feature distribution. These change features include the center coordinates, boundary outline, altitude, velocity direction, expansion range, and minimum distance to the current flight path of the new obstacle. When the minimum distance between the new obstacle and the current flight path is less than 5 meters, the system triggers dynamic adjustments to the local flight segment.

[0029] Path deviation penalties are used to constrain the degree of deviation of the local flight path from the global flight path. The system limits the maximum lateral deviation distance between the local flight path and the global flight path to 15 meters, the maximum altitude deviation to 8 meters, and the maximum heading angle change to 35 degrees. Local optimization inspection technology generates candidate obstacle avoidance segments within a spatial range of 30 meters in front of the UAV, 15 meters to the left and right, and 8 meters vertically above and below its current position. The system calculates the safe distance, turning radius, energy consumption, time consumption, and path deviation penalty value for each candidate obstacle avoidance segment, and selects the candidate obstacle avoidance segment with the lowest total cost as the corrected local flight path.

[0030] In S7, the system calculates the spatial matching degree M between the corrected local flight path and the global flight path. The spatial matching degree M consists of the deviation of the connection point position, heading angle deviation, flight altitude deviation, curvature deviation, and temporal continuity deviation. When M is not less than 0.85, the system determines that the local flight path and the global flight path meet the fusion conditions; when M is less than 0.85, the system regenerates the connecting segments and recalculates M until M reaches 0.85. Path smoothing uses cubic B-spline curves to fit the trajectory of the connecting segments, ensuring the continuity of the speed, acceleration, and heading angle of the connecting segments. Multi-objective optimization technology uses the minimum safety risk, shortest flight time, lowest energy consumption, and highest communication quality as optimization objectives to iteratively adjust the fused route. Historical data references include previous rescue routes, records of rockfall-prone areas, collapse areas, communication blind spots, and verified valley passages. The system maps risk points from historical data to the current terrain feature distribution and adds risk costs to the corresponding segments in the final route, completing the final route refinement and obtaining a UAV collaborative passage plan for mountain rescue.

[0031] S1 includes acquiring raw point cloud data and multispectral imagery from a multi-source sensor array carried by the UAV; performing feature fusion on the raw point cloud data and multispectral imagery to extract real-time terrain undulation features and temporary obstacle distribution features; using grid partitioning to spatially divide the real-time terrain undulation features and temporary obstacle distribution features into discrete terrain grid units; logically associating the discrete terrain grid units according to a preset topological relationship to form a structured terrain grid dataset; and performing 3D reconstruction based on the structured terrain grid dataset to determine the preliminary spatial distribution of the UAV flight environment.

[0032] In this embodiment, S1 is used to convert the raw point cloud data and multispectral imagery collected by the UAV in the mountainous disaster area into 3D environmental data that can be directly used for path planning. After the UAV enters the disaster area, the 32-line lidar continuously scans the mountain surface, road fracture areas, rockfall piles, collapse boundaries, fallen tree areas, and temporary rescue facilities, forming raw point cloud data with information on collection time, distance, azimuth, and altitude. The multispectral camera simultaneously acquires images in five bands: blue, green, red, red-edge, and near-infrared. Each pixel in the image records the reflection intensity of the corresponding ground object in different bands. The UAV's inertial measurement unit records the aircraft's attitude, the real-time dynamic differential satellite positioning module records the UAV's spatial position, and the barometric altimeter records the flight altitude. The above data are used together to determine the spatial correspondence between the point cloud and the imagery at the same time.

[0033] After the raw point cloud data enters the processing module, the system first sorts the point cloud frames according to the acquisition time, and then uses the UAV's position and attitude data to perform coordinate transformation on each frame of point cloud. During the transformation, the system uses the UAV's takeoff point as the local coordinate origin, transforming the point cloud in the lidar coordinate system to the local spatial coordinate system, so that point clouds obtained at different acquisition times fall into the same 3D space. For points less than 1 meter away from the UAV, the system determines them as noise reflected by the aircraft or propellers and deletes them; for points more than 150 meters away from the UAV and whose echo intensity is lower than the effective echo threshold, the system determines them as long-distance weak echo noise and deletes them; for discrete points with no other adjacent points within 0.5 meters of a single point, the system determines them as isolated noise and deletes them. After the above processing, the raw point cloud data is organized into effective point cloud data that can reflect the true spatial contour of the mountain.

[0034] After the multispectral imagery enters the processing module, the system first performs radiometric correction on the five band images to ensure consistent reflectance intensity for the same ground feature in adjacent images. Subsequently, the system corrects lens distortion using camera calibration parameters and performs spatial registration of the multispectral imagery based on UAV attitude and spatial location data. After spatial registration, each pixel in the multispectral imagery corresponds to a ground projection location in the local spatial coordinate system. The system identifies vegetation-covered areas based on the difference between near-infrared and red light reflectance intensity, exposed rock areas based on abrupt changes in red and green light reflectance intensity, waterlogged or slippery mud surfaces based on areas with decreased near-infrared reflectance and increased blue light reflectance, and landslide or collapse traces based on continuous dark textures and irregular boundaries. Through this processing, the multispectral imagery is converted into image feature data containing ground feature categories, anomalous areas, and surface material characteristics.

[0035] When fusing point cloud data and multispectral imagery, the 3D spatial location of the point cloud is used as the fusion benchmark. The system projects each valid point cloud point onto the multispectral imagery, reading the five-band reflectance intensity and land cover category of the pixel corresponding to the projection location. When the projection location falls between two pixels, the system takes the average reflectance intensity of the four pixels closest to that projection location as the spectral attribute of that point. When the projection location exceeds the effective range of the imagery, the system only retains the spatial attribute of the point and does not write the spectral attribute. After fusion, each point cloud point simultaneously possesses height, horizontal location, echo intensity, land cover category, and spectral attribute, and the system forms fused point cloud data accordingly.

[0036] Real-time terrain undulation features are extracted from ground points in the fused point cloud data. Within each horizontal sampling region, the system selects points with continuous height and slope, and stable connections to adjacent regions, as ground points. Points significantly higher than the surrounding ground and with independent contours are designated as non-ground points. For each ground sampling region, the system takes the median height of the ground points within the region as the elevation value, the maximum elevation difference between the region and its adjacent regions as the elevation change, and the ratio of the elevation change to the horizontal distance between the center of the region as the slope value. Terrain roughness is determined by the dispersion of the height of each ground point within the same region relative to the median elevation; the greater the dispersion, the more uneven the terrain surface. The system combines elevation values, slope values, aspect changes, and terrain roughness to form real-time terrain undulation features.

[0037] The distribution characteristics of temporary obstacles are extracted from non-ground points in the fused point cloud data and image recognition results. The system aggregates spatially continuous non-ground points that are close together into obstacle clusters, with a maximum spacing of 0.6 meters between adjacent points within a cluster. After each obstacle cluster is formed, the system determines the obstacle height based on the height difference between the highest point of the cluster and the ground elevation below it, the obstacle's occupied area based on the outer boundary of the cluster on the horizontal plane, and the obstacle's coverage area based on the area of ​​the outer boundary of the cluster. For clusters with a height exceeding 0.5 meters and a coverage area exceeding 1 square meter, the system marks them as temporary obstacles. The system then combines the recognition results of multispectral imagery and visible light images to distinguish obstacle categories; obstacles with high surface reflectivity, rigid boundaries, and abrupt changes in point cloud height are classified as falling rocks or rock deposits; obstacles with a strip-like distribution and accompanying vegetation spectral characteristics are classified as fallen trees; and obstacles with regular rectangular boundaries and artificial material spectral characteristics are classified as rescue vehicles or temporary facilities. The system records the obstacle's center location, outer boundary, height, category, and coverage area as the temporary obstacle distribution characteristics.

[0038] After feature fusion, the system employs grid partitioning technology to spatially divide real-time terrain undulation features and temporary obstacle distribution features. The area to be planned is divided into 3D voxel grids, with the length, width, and height of each grid cell set to 2 meters. The system writes each fused point's spatial coordinates into its corresponding grid cell; points falling into the same grid cell participate in the attribute calculation for that cell. The elevation value of each grid cell is determined by the median elevation of ground points within that cell; the slope value is determined by the elevation variation between that cell and its neighboring cells; the terrain roughness is determined by the elevation dispersion of ground points within that cell; and the surface category is determined by the most frequently occurring image feature category within that cell. For grid cells containing temporary obstacle clusters, the system records the obstacle category, obstacle height, obstacle occupancy range, and obstacle outer boundary position.

[0039] In determining the occupancy status of grid cells, the system uses a tiered approach based on ground condition, obstacle status, and flight safety status. A grid cell is marked as passable if it contains no obstacle clusters and has no obstacles within three consecutive altitude layers above it. A grid cell is marked as an obstacle grid if it contains temporary obstacles exceeding 0.5 meters in height or if the outer boundary of an obstacle crosses the grid cell. When a grid cell is located near a collapse boundary, rockfall boundary, the outer edge of a fallen tree, or the outer edge of a temporary rescue facility, the system establishes a safety buffer zone based on the corresponding obstacle category: 2 meters for rockfall boundaries, 5 meters for collapse boundaries, 3 meters for the outer edge of a fallen tree crown, and 2 meters for the outer edge of a rescue facility. Grid cells falling into these buffer zones are marked as buffer grids. Through this process, continuous terrain undulations and obstacle distributions are converted into discretized terrain grid cells.

[0040] After the discretized terrain grid cells are formed, the system establishes logical connections between the grid cells according to preset topological relationships. Each grid cell has a unique spatial number, which is determined by the grid cell's lateral, longitudinal, and altitude layer positions. The system first establishes adjacency relationships for each grid cell within the same altitude layer in the front, back, left, right, and four diagonal directions. Then, it establishes inter-layer relationships between upper and lower grid cells in the vertical direction. If a UAV can fly from the current grid cell into an adjacent grid cell, the system marks the two grid cells as passable connections. If there are obstacle grids, no-fly buffer grids, or altitude changes exceeding the UAV's single climb limit between adjacent grid cells, the system marks the two grid cells as impassable connections. The single climb altitude limit for a UAV between adjacent grid cells is set to 2 meters, and the single lateral movement distance is determined by the grid side length and is also 2 meters.

[0041] Topological relationships are also used to express the transmission of risk between grids. Once the grid containing a rockfall obstacle is determined, the system marks two consecutive layers of adjacent grids in the downslope direction as rockfall-affected grids; the downslope direction is determined by the direction in which the elevations of adjacent grids decrease sequentially. Once the grid containing a collapse boundary is determined, the system marks the grids within a 5-meter radius on both sides of the boundary as collapse-affected grids. Once the grid containing a fallen tree is determined, the system marks the grids within a 3-meter radius extending outward from the outer edge of the tree crown as tree-affected grids. These affected grids are not directly equivalent to obstacle grids, but they are assigned a higher passage risk in subsequent path planning. Through adjacency relationships, inter-layer relationships, passage relationships, and risk transmission relationships, the system organizes discrete grids into a structured terrain grid dataset.

[0042] The structured terrain mesh dataset includes spatial location, terrain attributes, obstacle attributes, accessibility attributes, and risk attributes. Spatial location records the center coordinates and height layer number of the mesh cell; terrain attributes record elevation, slope, aspect, terrain roughness, and surface type; obstacle attributes record obstacle type, height, outer boundary, and occupancy status; accessibility attributes record the accessibility relationships between the mesh and adjacent meshes; and risk attributes record rockfall risk, collapse risk, collision risk, and communication obstruction risk. The system synchronously writes the data update time when writing each attribute, enabling subsequent real-time updates to replace only the changed meshes without reconstructing the entire mesh data.

[0043] When performing 3D reconstruction based on a structured terrain grid dataset, the system first reads the center coordinates, elevation values, and height layer numbers of each grid cell to generate a basic spatial model of the mountain surface. When the elevations of adjacent grids are continuous, the system connects the ground elevations of adjacent grids using continuous patches; when the elevation abrupt change between adjacent grids exceeds 2 meters, the system creates a steep slope or collapse boundary between them. Obstacle outlines are generated based on the outer boundaries of the obstacle grids. When multiple adjacent obstacle grids belong to the same obstacle category, the system merges them into a single complete obstacle. No-fly zones consist of obstacle grids and no-fly buffer grids. Buffer zones consist of grids extending beyond rockfall, collapse, fallen trees, and rescue facilities. Passable airspace consists of grids that meet height, obstacle, and passability connection conditions.

[0044] After 3D reconstruction, the system obtains a preliminary spatial distribution of the UAV flight environment. This spatial distribution includes the mountain surface, slope change areas, collapse boundaries, rockfall accumulation areas, fallen tree areas, areas occupied by rescue facilities, no-fly zones, buffer zones, and passable airspace. When the path planning module reads this spatial distribution, it can directly identify which grids are accessible, which are prohibited, which grids require increased passage risk, and which grids can serve as candidate locations for subsequent key flight nodes.

[0045] S2 includes obtaining the node elevation gradients and neighborhood connectivity of the structured terrain grid dataset to determine the terrain grid node feature matrix; based on the terrain grid node feature matrix, using a node priority ranking algorithm to evaluate the weights of the grid nodes and determine the set of key flight nodes; performing a difference operation on the set of key flight nodes and the terrain grid node feature matrix to obtain a spatial distribution map of obstacle-dense areas; determining the terrain complexity distribution model by weighted fusion of the obstacle-dense area spatial distribution map and the terrain grid node feature matrix; and using the terrain complexity distribution model to perform feature dimensionality reduction processing on the terrain grid nodes to obtain a simplified terrain feature representation.

[0046] In this embodiment, S2 is used to extract core terrain features affecting UAV flight safety and route connectivity based on the structured terrain grid dataset. In the structured terrain grid dataset, each grid node records its center coordinates, elevation value, obstacle occupancy status, buffer status, adjacent node numbers, and passage connection status. After reading the above data, the system does not directly input all grid nodes into the path search. Instead, it first judges the degree of terrain abrupt change and surrounding passage conditions for each node, and then filters out nodes that have a controlling effect on path turns, obstacle avoidance, and passage connections.

[0047] To obtain the node elevation gradient, the system takes the current grid node as the center, reads the elevation value of the node itself, and reads the elevation values ​​of its four diagonally adjacent grid nodes (front, back, left, right, and four diagonally adjacent). The system compares the elevation difference between the current grid node and each of its adjacent grid nodes, and combines this with the horizontal distance between the centers of two grid nodes to determine the intensity of terrain change in that direction. After determining all eight directions, the system selects the direction with the greatest intensity of terrain change as the main terrain abrupt change direction for that node, and uses the corresponding intensity of change as the elevation gradient of that node. Through this processing, ridges, steep slopes, collapse edges, and rockfall boundaries form obvious elevation abrupt change characteristics in the node data, while gentle slopes and open passageways form low elevation abrupt change characteristics.

[0048] For grid nodes at the boundary of the planning area, the system only reads adjacent grid nodes that actually exist within the planning area for elevation gradient determination, without introducing blank data from outside the planning area. If an adjacent grid node is occupied by an obstacle, the system still reads the elevation value of that adjacent grid node, because the height difference between the obstacle edge and the ground surface affects the drone's detour and climb determination. For adjacent grid nodes that have not yet been updated, the system uses the median elevation of the surrounding updated valid grid nodes as the supplementary elevation and writes a low-confidence marker on that adjacent grid node. Nodes with low-confidence markers are not used as priority passage nodes during subsequent node sorting.

[0049] Neighborhood connectivity is determined by the feasibility of a UAV moving from one grid node to surrounding grid nodes. The system, centered on the current grid node, sequentially checks eight neighboring grids at the same altitude level, nine neighboring grids at the upper altitude level, and nine neighboring grids at the lower altitude level. For each neighboring grid, the system first determines if the grid exists, then whether it is an obstacle grid, a no-fly zone grid, or a grid with severe communication obstruction. Next, it determines if there is a passable connection between the current grid node and the neighboring grid, and whether the altitude change between the two grids is within the UAV's single climb or descent capability. If all conditions are met, the neighboring grid is recorded as a connectable neighborhood; if any condition is not met, the neighboring grid is recorded as an unconnectable neighborhood.

[0050] After completing the neighborhood check, the system determines the neighborhood connectivity of the current grid node based on the number of connectable neighbors. A higher number of connectable neighbors indicates more options for the UAV to move from that node to the surrounding space, making the node more suitable as a flight path connection point, hovering observation point, or turning node. Conversely, a lower number of connectable neighbors indicates that the passage space around the node is obstructed, there are abrupt terrain changes, or the no-fly zone is compressed, making the node less of a priority passage node. For nodes located in narrow valley passages, outside collapse boundaries, or at the edge of rockfall piles, even if they are not obstruction grids, the system still considers neighborhood connectivity to determine their passage stability, avoiding misclassifying nodes with only a single exit as safe nodes.

[0051] After determining the elevation gradient and neighborhood connectivity, the system generates a terrain grid node feature matrix. This feature matrix uses grid nodes as the recording unit, with each node corresponding to one node record. Each node record contains the node number, node center coordinates, node elevation gradient, main terrain abrupt change direction, neighborhood connectivity, obstacle occupancy status, buffer zone status, communication obstruction status, and update time. The elevation gradient reflects whether the terrain undergoes abrupt changes, neighborhood connectivity reflects spatial continuity, obstacle occupancy status reflects whether the node is occupied by physical obstacles, buffer zone status reflects whether the node is affected by falling rocks, collapses, or fallen trees, and communication obstruction status reflects the risk of data transmission risks caused by mountain obstruction at the node. Thus, the spatial data, obstacle data, and access data, originally scattered throughout the terrain grid dataset, are organized into unified node feature data.

[0052] When forming the terrain grid node feature matrix, the system performs uniform processing on various node features. Elevation gradient is a continuous numerical value, neighborhood connectivity is a statistical result of the number of neighbors, and obstacle occupancy status and buffer status are status markers. Directly mixing these would cause the ranking results to be affected by a single data type. The system first determines the maximum and minimum values ​​of each type of continuous feature within the current planning area, and then transforms the corresponding features of each node to the same numerical range. For status features, the system assigns levels according to the risk order of obstacle grids, no-fly buffer grids, obstacle avoidance buffer grids, and passable grids. After this processing, each node has a comparable feature representation, and the subsequent node priority ranking has a unified data foundation.

[0053] During the node priority ranking process, the system first excludes obstacle grids and no-fly buffer grids from the terrain grid node feature matrix. The remaining passable grids and obstacle avoidance buffer grids then enter the candidate node queue. When evaluating the weights of candidate nodes, the system considers target direction, terrain abrupt changes, neighborhood connectivity, obstacle distance, communication obstruction, and altitude stability. Nodes located in the direction from the starting point to the rescue target have path advancement value; nodes with high neighborhood connectivity have passability value; nodes whose distance from obstacle boundaries meets the safety interval have safe passage value; nodes with low communication obstruction have data transmission value; and nodes with continuous altitude changes have flight stability value. All of these values ​​contribute to the node ranking, rather than a single distance factor determining node priority.

[0054] For elevation gradient features, the system segments them according to their actual role in path planning. Nodes with low elevation gradient abrupt changes and high neighborhood connectivity are identified as stable passage nodes, used to form straight or gently curved flight segments; nodes with moderate elevation gradient abrupt changes and located outside obstacle boundaries are identified as risk boundary control nodes, used to guide flight paths around densely obstructed areas; nodes with high elevation gradient abrupt changes are identified as danger edge nodes, not used as flight overpasses, but their locations are retained in subsequent densely obstructed area identification. This processing allows the system to retain node information reflecting terrain boundaries while avoiding the misselection of collapse edges, steep slope edges, and rockfall tops as flight path nodes.

[0055] After ranking the candidate nodes, a set of critical flight nodes is generated. This set consists of mission-designated nodes, stable access nodes, and risk boundary control nodes. Mission-designated nodes include the rescue starting point, rescue target point, and communication relay point; stable access nodes are selected from grid nodes with high neighborhood connectivity, obstacle distances meeting safety requirements, and continuous altitude changes; risk boundary control nodes are selected from nodes outside dense obstacle areas, valley entrances, outside collapse boundaries, and near the outer edge of rockfall piles. For candidate nodes with overlapping spatial locations or spacing less than 4 meters and the same altitude layer, the system retains only the node with the highest ranking to avoid excessive concentration of critical flight nodes in the same small area. The selected set of critical flight nodes simultaneously meets the requirements of controlled quantity, continuous spatial distribution, and effective path control.

[0056] After obtaining the set of critical flight nodes, the system performs a difference operation between it and the terrain grid node feature matrix to identify obstacle-dense areas outside the critical flight nodes. The system first maps the set of critical flight nodes back to the original grid space and finds the spatially closest and topologically reachable critical flight node for each ordinary grid node. Then, the system compares the differences in elevation gradient, neighborhood connectivity, obstacle occupancy status, and buffer zone status between ordinary grid nodes and their corresponding critical flight nodes. The more significant the elevation abrupt change, the worse the connectivity, and the more severe the obstacle or buffer zone markings on an ordinary grid node, the greater the difference between it and the critical flight node; this node is then identified as deviating from safe passage characteristics.

[0057] The essence of difference calculation is to use key flight nodes to represent a safe or controllable passage state, and then find grid regions that differ significantly from this state. When the elevation gradient of a regular grid node is higher than that of the corresponding key flight node, and its neighborhood connectivity is lower than that of the corresponding key flight node, the system determines that the node is located at a location with abrupt terrain change and restricted passage. When the regular grid node also has an obstacle occupancy marker or a buffer zone marker, the system further determines that it belongs to an obstacle-affected area. When multiple nodes with the above differences are spatially continuous, the system merges these nodes into an obstacle cluster area. The system writes the difference judgment results into a spatial layer to form a spatial distribution map of densely obstructed areas.

[0058] The generation of the spatial distribution map of dense obstacle areas also includes region merging and boundary closure. The system checks whether grid nodes with high disparity results are adjacent; horizontal, vertical, or diagonal adjacency is considered to indicate spatial continuity. Continuously distributed high disparity nodes are merged into a single obstacle candidate region. When the merged region contains more than five consecutive obstacle nodes, or when the horizontal projected area reaches 20 square meters, the system identifies the region as a dense obstacle area. For regions formed by continuous distribution of rockfall piles, fallen tree clusters, and collapse boundaries, the system generates closed boundaries along their outermost grid nodes; for isolated, scattered obstacle nodes that do not form a continuous distribution, the system retains only single-point risk markers and does not merge them into dense obstacle areas. Thus, the spatial distribution map of dense obstacle areas records not only the location of obstacles but also the outer contour and continuous range of the obstacle region.

[0059] Subsequently, the system performs a weighted fusion of the spatial distribution map of densely obstructed areas with the feature matrix of terrain grid nodes to determine the terrain complexity distribution model. During fusion, the system writes back the spatial extent, boundary location, and outer buffer range of the densely obstructed areas to the corresponding grid nodes. Nodes falling inside the densely obstructed areas are marked with the highest regional risk; nodes within 2 meters of the outer boundary of the densely obstructed areas are marked with a boundary risk; nodes within 2 to 5 meters of the outer boundary are marked with a buffer risk; and nodes outside the 5-meter range and without other obstruction influence are not marked with a regional risk. These regional risk marks, along with the node's original elevation gradient, neighborhood connectivity, obstacle occupancy status, buffer state, and communication obstruction status, are used in the complexity determination.

[0060] The terrain complexity distribution model is determined node by node. Nodes exhibiting abrupt elevation changes, insufficient neighborhood connectivity, obstacle occupation, or the highest risk marker are classified as high-complexity nodes and treated as detour or no-fly zones in subsequent path planning. Nodes located outside the boundary of obstacle-dense areas, with partial connectivity and not occupied by physical obstacles are classified as medium-complexity nodes; these are used to form buffer detour routes or observation nodes. Nodes with gentle elevation changes, sufficient neighborhood connectivity, not falling within obstacle influence range, and low communication obstruction are classified as low-complexity nodes; these are used to form priority passage corridors. After determining the complexity of all nodes, the system obtains a terrain complexity distribution model covering the entire planning area.

[0061] The terrain complexity distribution model is also used to express the travel trends between regions. The system examines the continuous changes in complexity level along adjacent grid nodes. When low-complexity nodes are continuously connected to the starting and destination directions, the continuous area is identified as the main travel corridor; when medium-complexity nodes are continuously distributed outside densely obstacle areas, the continuous area is identified as a detour buffer zone; when high-complexity nodes form a closed or semi-closed distribution, the area is identified as an avoidance zone. Through this spatial continuity judgment, the model output is no longer the risk of isolated nodes, but rather the travel areas, detour areas, and avoidance areas that can participate in path search.

[0062] After determining the terrain complexity distribution model, the system performs feature dimensionality reduction on the terrain grid nodes. Dimensionality reduction is not simply deleting features, but rather merging multiple low-level features expressing the same accessibility meaning into comprehensive features required for path planning. Elevation gradient, slope change, and terrain roughness, all used to reflect abrupt terrain changes, are merged into comprehensive terrain change features; neighborhood connectivity, number of access connections, and upper / lower layer connection states, all used to reflect spatial accessibility, are merged into comprehensive access connectivity features; obstacle occupancy status, buffer zone status, and obstacle-dense area risk, all used to reflect collision and detour risks, are merged into comprehensive obstacle risk features; communication obstruction status and altitude layer stability remain as independent constraints. After merging, the number of node features is reduced, but information directly related to flight safety and path generation is retained.

[0063] The reduced-dimensionality nodes retain different content according to their complexity level. For low-complexity nodes, the system retains the center coordinates, comprehensive terrain change features, comprehensive connectivity features, and communication obstruction status for rapid generation of the main flight segment. For medium-complexity nodes, the system retains comprehensive obstacle risk features and detour direction information in addition to the above content for generating local obstacle avoidance segments. For high-complexity nodes, the system retains the center coordinates, comprehensive obstacle risk features, and no-fly or detour markers, but no longer retains detailed data such as surface material that does not participate in route decision-making. This processing ensures that high-complexity areas still participate in risk constraints without increasing the computational burden of subsequent path search. The final simplified terrain feature representation includes a set of key flight nodes, boundaries of obstacle-dense areas, low-complexity passage corridors, medium-complexity detour buffer zones, high-complexity avoidance areas, and the connectivity relationships between these areas.

[0064] S3 includes extracting the proportion of densely obstructed areas from simplified terrain features. If the proportion exceeds a threshold, new terrain change information is acquired through a sensor array. The adjustment range of the local topology is determined based on the terrain change information. The adjustment range is used to guide the correction operator to locally correct the simplified terrain features. The simplified terrain features are reconstructed using the correction operator to obtain a feature mapping model. The global consistency of the feature mapping model is then checked to determine the updated terrain feature distribution.

[0065] In this embodiment, S3 is used to address the issue that the original simplified terrain features cannot accurately reflect the current flight environment after new changes occur in the mountain disaster area environment. The simplified terrain features obtained in S2 already include key flight nodes, boundaries of obstacle-dense areas, low-complexity passage corridors, medium-complexity detour buffer zones, high-complexity avoidance areas, and passage connections between adjacent areas. After entering S3, the system first reads all grid nodes marked as obstacle-dense areas from the simplified terrain features and extracts the outer contour, number of covered grids, and horizontal projection range of each obstacle-dense area.

[0066] When extracting the area proportion of densely obstructed regions, the system uses the current path planning range as the statistical object. The current path planning range is jointly determined by the starting point, the rescue target point, and the preset safety outer boundary. The system first calculates the horizontal projected area of ​​all effective grids within this range, and then calculates the horizontal projected area of ​​the grids marked as densely obstructed regions. For overlapping portions of multiple densely obstructed regions, the system only calculates them once to avoid duplicate calculations. Subsequently, the system compares the total horizontal projected area of ​​the densely obstructed regions with the total horizontal projected area of ​​the current path planning range to obtain the degree of occupancy of the densely obstructed regions within the current planning range. This degree of occupancy is the area proportion mentioned in the claims.

[0067] In this implementation, the threshold is set at 35%. When the area with dense obstacles accounts for more than 35% of the total area, it indicates that the current level of obstacle concentration in the disaster area has substantially affected the original passageway, key flight nodes, and detour buffer zones. The system then triggers the sensor array to collect new terrain change information. When the area accounts for less than 35%, the system does not initiate local topology reconstruction, but only retains the original simplified terrain features and proceeds to the subsequent global path search. Therefore, the threshold determination is not simply a risk warning, but is used to decide whether to reacquire environmental change information and whether to initiate a local correction process.

[0068] When acquiring new terrain change information, the sensor array prioritizes the boundaries of densely obstructed areas, the areas adjacent to the original global flight path, and the areas surrounding key flight nodes. A 32-line lidar performs point cloud supplementation in these areas to obtain the spatial contours, height changes, and boundary positions of new obstacles. A multispectral camera acquires surface spectral changes to identify exposed rock masses, soil slippage, fallen vegetation, and muddy surfaces. A visible light camera acquires texture images to identify rescue vehicles, temporary tents, road breaks, and fallen trees. An infrared thermal imaging camera acquires heat source information to rule out the possibility of personnel heat sources being mistaken for physical obstacles. An inertial measurement unit, barometric altimeter, and real-time dynamic differential satellite positioning module provide the attitude, altitude, and spatial coordinates at the time of acquisition. After time alignment and coordinate transformation, the above data is written into a local spatial coordinate system consistent with the original simplified terrain features.

[0069] The process of identifying newly added terrain changes uses the original terrain features as a comparison benchmark. The system spatially overlays the newly added point cloud with the original terrain grid and compares the height, boundary, and occupancy status of the same location. When the newly added point cloud rises more than 0.5 meters relative to the original terrain surface and has a continuous coverage area of ​​more than 1 square meter, the system identifies the location as a newly added physical obstacle; when the newly added point cloud drops more than 1 meter relative to the original terrain surface and the descent boundary extends continuously, the system identifies the location as a terrain collapse; when the newly added point cloud is distributed in strips or patches and multispectral imagery shows vegetation features, the system identifies it as fallen trees; when the newly added point cloud has a regular outer contour and visible light imagery shows textures of artificial facilities, the system identifies it as an area occupied by temporary rescue facilities. For each newly added object, the system records the center location, outer boundary, height change, horizontal coverage, change category, and acquisition time.

[0070] When determining the adjustment range of the local topology based on terrain change information, the system does not remodel the entire planning area. Instead, it identifies the local areas that need correction around the newly added objects. For newly added physical obstacles, the grid they actually occupy is determined as the direct adjustment zone; for fallen rocks, the system extends 2 meters outward from the outer boundary of the fallen rocks to form a safety buffer zone; for terrain collapses, the system extends 5 meters outward from the collapse boundary to form a no-fly buffer zone; for fallen trees, the system extends 3 meters outward from the outer edge of the tree crown to form an obstacle avoidance buffer zone; for temporary rescue facilities, the system extends 2 meters outward from the outer contour of the facility to form an avoidance buffer zone. The direct adjustment zone and the buffer zones together constitute the main adjustment range of the local topology.

[0071] In addition to the direct adjustment zone and buffer zone, the system also includes two consecutive layers of adjacent grids outside the main adjustment range in the connection adjustment zone. The connection adjustment zone is used to check whether existing access connections still hold. For example, if a low-complexity access corridor in the original simplified terrain feature passes outside a newly added rockfall boundary, even though the corridor is not directly occupied by rocks, the safe distance between it and the rockfall boundary is insufficient; therefore, the access connection corresponding to this corridor needs to be deleted or downgraded. Similarly, if a critical flight node was originally located outside a collapse boundary but falls within a 5-meter no-fly zone, then this critical flight node will no longer participate in subsequent path searches. By setting up connection adjustment zones, the system can correct nodes and edges indirectly affected by new changes, rather than only modifying the grid containing the obstacle itself.

[0072] After the adjustment range is determined, the correction operator performs local corrections on the simplified terrain features in four aspects: node status, region boundaries, access connections, and critical flight nodes. The correction of node status involves rewriting the grids actually occupied by newly added obstacles as obstacle nodes, rewriting the grids within the collapse boundary and its 5-meter extension as no-fly nodes, and rewriting the grids within the extension range of fallen rocks, fallen trees, and temporary rescue facilities as obstacle avoidance buffer nodes. The correction of region boundaries involves regenerating the boundary of densely obstructed areas based on the outer contour of the newly added changed objects, and merging obstacle nodes that are in contact with each other or whose spacing is less than one grid side length into continuous obstacle areas.

[0073] The correction of accessible connections involves checking whether each existing connection within the adjustment area crosses obstacle nodes, no-fly zones, or buffer zones with insufficient safety distances. Connections crossing these nodes are deleted; connections where both the starting and ending points are accessible nodes and the connection path does not pass through obstacles or no-fly zones are retained. After a connection is deleted, the system searches for adjacent accessible nodes within the connection adjustment area and re-establishes an alternative connection. Alternative connections must simultaneously meet the following conditions: continuous flight altitude, lateral spacing satisfying grid adjacency, no crossing of obstacle boundaries, and no entry into no-fly buffer zones. The correction of critical flight nodes involves deleting critical flight nodes that fall into obstacle nodes or no-fly zones, downgrading critical flight nodes that fall into obstacle avoidance buffer zones to risk boundary candidate nodes, and selecting nodes from the edge of the adjustment area whose neighborhood connectivity meets the requirements to add to the candidate set.

[0074] After the correction operator completes the local correction, the system reconstructs and simplifies the terrain features using the corrected node states, boundary states, and connection states, forming a feature mapping model. This feature mapping model records the correspondence between terrain features before and after correction. For each corrected grid node, the system records its original node type, updated node type, original complexity level, updated complexity level, original access connectivity, updated access connectivity, and the source of change that triggered the correction. Sources of change include newly added falling rocks, terrain collapse, fallen trees, temporary rescue facility occupation, and supplementary sensor observations. Nodes not within the adjustment range maintain their original state, but their current verification time is recorded so that subsequent checks can determine whether the node participated in the current round of global consistency verification.

[0075] After the feature mapping model is established, the system performs a global consistency check. First, it checks the consistency of access connections. The system traverses all updated access connections, confirming that no accessible nodes directly connect to obstacle nodes or no-fly zones, that connection segments do not cross the boundaries of densely obstructed areas, and that altitude changes between adjacent connection nodes are within the drone's single climb or descent capability. If a connection is found to cross a no-fly zone, the system deletes the connection. If a low-complexity access corridor is found to be broken due to connection deletion, the system searches for alternative nodes outside the connection adjustment area to maintain reachability between the starting point, communication relay point, and rescue target point.

[0076] Secondly, the consistency of the area boundaries is verified. The system checks whether the boundary of the obstacle-dense area surrounds its internal obstacle nodes, whether the no-fly buffer zone completely covers the 5-meter extension beyond the collapsed boundary, and whether the obstacle avoidance buffer zone completely covers the extension of fallen rocks, fallen trees, and temporary rescue facilities. If there are gaps in the boundary, the system fills in the missing boundary according to the relationship between adjacent grids; if a grid is simultaneously written as a passable node and an obstacle avoidance buffer node, the system processes it as an obstacle avoidance buffer node; if a grid is simultaneously written as an obstacle avoidance buffer node and a no-fly node, the system processes it as a no-fly node. Through this process, the updated area markings do not have conflicting states.

[0077] The consistency of complexity distribution is verified again. The system checks whether low-complexity passageways cross areas with dense newly added obstacles, whether medium-complexity detour buffer zones are located outside areas with dense obstacles, and whether high-complexity avoidance zones cover newly added obstacles and their no-fly zones. If an original low-complexity passageway enters the area affected by a newly added obstacle, the system upgrades its corresponding node to a medium-complexity or high-complexity node and re-marks the detour direction. If a newly added obstacle renders an original critical node ineffective, the system selects new passable nodes from the edge of the adjustment range to supplement the set of critical flight nodes. After this verification, the complexity level is consistent with the actual obstacle distribution.

[0078] Finally, the system verifies the consistency of time and data source. It compares the acquisition time of newly added terrain change information with the generation time of the original simplified terrain features, only including data acquired later than the original feature generation time in the overlay update. For cases where multiple update records exist for the same grid node, the system retains the record with the latest acquisition time, valid spatial positioning, and satisfactory sensor confidence. When there are category differences between LiDAR point cloud and image recognition results, the system uses the LiDAR point cloud, which reflects 3D contours, to determine obstacle boundaries and heights, and uses image recognition results to determine obstacle categories. This avoids expired data overwriting new data and also prevents spatial boundary errors caused by misidentification of a single image. After global consistency verification passes, the system determines the feature mapping model as the updated terrain feature distribution. This distribution includes the updated set of key flight nodes, boundaries of obstacle-dense areas, low-complexity passage corridors, medium-complexity detour buffer zones, high-complexity avoidance areas, and the passage connections between these areas.

[0079] S4 includes obtaining a global flight node set based on terrain feature distribution, which contains the spatial coordinates and connectivity data of each node; traversing the global flight node set using a path cost calculation method to determine the path cost values ​​between each node; generating the distance weight distribution of each potential route based on the path cost values ​​and a preset distance estimation method; conducting a feasibility assessment of the potential routes by combining the distance weight distribution and obstacle weight allocation techniques; and selecting the optimal path based on the feasibility assessment results to obtain a preliminary global flight route.

[0080] In this embodiment, S4 is used to generate a preliminary global flight path based on the updated terrain feature distribution. The system first reads the terrain feature distribution obtained in S3 and extracts nodes that can participate in the flight search. The extracted nodes include starting point nodes, rescue target nodes, communication relay nodes, low-complexity passageway nodes, medium-complexity detour buffer nodes, and critical flight nodes that meet the safety distance requirements; obstacle nodes, no-fly zones, collapse buffer zones, and nodes less than 5 meters away from the rockfall boundary are not included in the global flight node set.

[0081] When the global flight node set is formed, the system writes spatial coordinates, altitude layer, node type, and connectivity data for each node. Spatial coordinates are determined by the node's center position within the terrain feature distribution; the altitude layer is determined by the node's corresponding flight altitude; and the node type distinguishes between passing nodes, bypass nodes, communication relay nodes, and target nodes. Connectivity data is obtained by checking adjacent nodes: the system checks each node to determine if there is continuous passable space between the current node and its adjacent nodes, whether the two nodes traverse areas with dense obstacles, no-fly zones, or communication obstructions, and whether the altitude change exceeds the UAV's single climb or descent capability. Node pairs that meet the passability conditions are retained as connectable nodes; node pairs that do not meet the passability conditions are not connected.

[0082] After determining the global set of flight nodes, the system traverses the set using a path cost calculation method. The traversal begins at the starting node and expands layer by layer along established connectable relationships to adjacent nodes. For each candidate flight segment pointing from the current node to an adjacent node, the system reads the segment's spatial length, altitude change, obstacle proximity, estimated passage time, and communication status, converting these factors into corresponding path cost values. Longer spatial lengths result in higher distance costs; greater altitude changes result in higher climb or descent costs; closer to rockfall areas, landslide areas, or dense forest areas, the higher the obstacle costs; and communication costs increase when passing through areas with communication obstruction. The system combines all costs and uses this as the path cost value for the candidate flight segment within the node connectivity relationships.

[0083] The calculation of path cost values ​​relies on specific flight segment data. The flight segment length is determined by the spatial coordinates of two adjacent nodes. The system calculates the differences between the two nodes in the lateral, longitudinal, and altitude directions, and then uses these three directional differences to determine the 3D spatial distance. Flight altitude changes are determined by the altitude layer differences between the two nodes; the estimated passage time is determined by the flight segment length, UAV cruise speed, climb time, descent time, and turn time; obstacle proximity is determined by the minimum distance from the flight segment centerline to the boundary of the nearest dense obstacle area. These results are written into the corresponding node connections, enabling subsequent searches to directly compare the passage costs of different flight segments.

[0084] The system generates a distance weight distribution for potential flight routes based on path cost values ​​and a preset distance estimation method. A potential flight route is formed by connecting multiple consecutive nodes. The system sequentially reads the segment distances between adjacent nodes according to the flight sequence and accumulates the distances of each segment throughout the entire route. Simultaneously, the system records the altitude and turning changes for each segment. For any segment exhibiting continuous climbs, continuous descents, or large-angle turns, additional throughput is added to the distance weight of that segment. Therefore, the distance weight distribution reflects not only the length of the flight route but also the actual flight load in a mountainous environment.

[0085] During obstacle weighting, the system performs risk correction on potential routes based on obstacle type and distance. Routes crossing rockfall areas or landslide areas are directly deemed infeasible; routes entering within 2 meters of rockfall boundaries or 5 meters of landslide boundaries are marked as high-risk; routes passing through fallen trees, dense forests, or the buffer zone around temporary rescue facilities are marked as medium-risk; and routes maintaining a safe distance from dense obstacle areas and located within low-complexity passageways are marked as low-risk. The system incorporates these risk markers into the evaluation data of potential routes to correct route ranking results obtained solely based on distance.

[0086] When assessing the feasibility of potential flight routes, the system evaluates the entire route, not just individual segments. The system checks whether the route connects continuously from its origin to the rescue target point, whether it passes through necessary communication relay nodes, whether it avoids obstacle nodes and no-fly zones, and whether it meets the requirements for the UAV's maximum climb capability, maximum turning angle, safe distance, and endurance. Any route with broken connections, crossing no-fly zones, insufficient safe distance, excessive altitude changes, or unsatisfactory communication continuity is excluded from the candidate routes. Through this process, the remaining candidate routes all meet the practical flight requirements.

[0087] The optimal route is determined based on the feasibility assessment results. The system first selects the route with the lowest safety risk from the candidate routes; when safety risk levels are consistent, it selects the route with the lowest total distance weight; when total distance weights are consistent, it selects the route with fewer altitude changes, fewer turns, and continuous communication; when the above conditions are still consistent, it selects the route with fewer critical flight nodes and a more continuous path structure. The final selected node sequence is connected according to flight order to form a preliminary global flight route from the origin to the rescue target point.

[0088] S5 includes extracting terrain feature data based on the geospatial extent of the global flight route, constructing terrain sub-maps of local risk areas, determining the risk level value of each sub-map through spatial overlap analysis of terrain complexity and dynamic obstacles within the terrain sub-maps, and if the risk level value exceeds a preset threshold, then weighting the branches within the terrain sub-maps using route connectivity and weather interference to obtain the route feasibility score for each branch, and predicting the time consumption of the branches based on the mapping relationship between the route feasibility score and flight energy consumption ratio, combined with flight timestamps, to determine the flight time series of each branch, and prioritizing the branches through correlation analysis between the flight time series and airspace congestion to obtain the priority sequence of each branch, and reconstructing local flight segments based on the priority sequence to obtain optimized local flight segments, thus achieving precise planning of local flight segments.

[0089] In this embodiment, S5 is used to refine the planning of flight segments that pass through local risk areas in the preliminary global flight path. After reading the global flight path obtained in S4, terrain feature data within a range of 20 meters on both sides of the centerline and 10 meters above and below the flight altitude are extracted, using the centerline of the route as a reference. The data within this range includes elevation changes, slope changes, terrain complexity level, obstacle boundaries, dynamic obstacle locations, passable nodes, buffer nodes, no-fly zones, and the passability connections between nodes. The above range is determined based on the minimum obstacle avoidance distance, turning radius, and effective sensor recognition width of the UAV in mountainous rescue flights, ensuring that the extracted area covers the actual detour space of the current flight segment.

[0090] Based on the risk locations traversed by the global flight route, a terrain sub-map set is constructed for local risk areas. Specifically, when the global flight route enters areas affected by rockfalls, landslides, dense forests, narrow valleys, communication blockages, or dynamic obstacles, the system extracts the corresponding local grid data centered on that risky flight segment and combines the passable nodes, obstacle nodes, buffer nodes, and passability connections into a single terrain sub-map. Multiple risky flight segments are processed into multiple terrain sub-maps according to the sequence of the global flight route, thus obtaining the terrain sub-map set. Each terrain sub-map retains entry and exit nodes for easy reconnection with the global flight route later.

[0091] The risk level is determined by the spatial overlap between terrain complexity and dynamic obstacles. The system first reads the complexity level of each node within the terrain submap, and then maps the impact range of newly added falling rocks, expanded collapse areas, fallen trees, rescue vehicles, and temporary facilities onto the same submap. The larger the overlap area between the impact range of a dynamic obstacle and high-complexity nodes, no-fly zones, or buffer zones, the higher the risk level of that submap; the closer the overlapping area is to the centerline of the global flight path, the higher the risk level; and the closer the height of the dynamic obstacle is to the current flight altitude of the UAV, the higher the risk level. The system limits the risk level to between 0 and 100.

[0092] In this embodiment, the preset threshold for the risk level is set to 70. This threshold is determined based on the minimum safe distance between UAVs, sensor positioning error, flight control response distance, and the safety margin of the rescue mission. Specifically, the system first determines the basic safety boundary based on the UAV's external dimensions and minimum obstacle avoidance distance, then superimposes the sensor positioning error and flight control response distance to form a risk buffer boundary. When the overlap between terrain complexity and dynamic obstacles causes the original global flight path to enter this risk buffer boundary, the corresponding risk level is classified as 70 or higher. Therefore, a risk level exceeding 70 indicates that the original global flight segment no longer meets the requirements for stable passage, necessitating a re-evaluation of the sub-map branches.

[0093] When the risk level value of a certain terrain submap exceeds 70, the system performs a route feasibility score on the branches within that submap. A branch refers to a candidate local route from the entrance node to the exit node of the terrain submap. The system first determines the connectivity of the branch, including whether the entrance and exit nodes are continuously reachable, whether the branch crosses obstacle nodes or no-fly zones, whether the altitude changes between adjacent nodes meet the UAV's climb or descent capabilities, and whether the turning positions meet the minimum turning radius. Branches with complete connectivity, continuous turns, and not entering no-fly zones receive a high feasibility score; branches with broken connections, insufficient safety distances, or abrupt altitude changes are excluded.

[0094] Meteorological interference is used to correct the feasibility score of branch routes. The system reads wind speed, wind direction, gust variation, rainfall intensity, visibility, and valley airflow disturbance data for the area on the topographic submap and matches them with the branch route direction. When the angle between the crosswind direction and the branch route increases, the gust variation intensifies, visibility decreases, or valley airflow disturbance is concentrated near the branch route, the meteorological interference score of that branch route increases. The system uses route connectivity as a positive evaluation factor and meteorological interference as a negative evaluation factor, and performs a weighted calculation on each branch route to obtain a route feasibility score. This score reflects whether the branch route can be stably traversed by the UAV under the current terrain and meteorological conditions.

[0095] Based on the correlation between route feasibility scores and flight energy consumption ratios, time consumption is predicted for each branch route. Branch routes with high feasibility scores are characterized by continuous passage, gentle altitude changes, and fewer obstacle avoidance maneuvers, resulting in lower flight energy consumption ratios. Branch routes with low feasibility scores have longer detour distances, more climbs, more frequent turns, or higher wind loads, resulting in higher flight energy consumption ratios. The system combines the UAV's flight timestamps upon entering the terrain submap to sequentially determine the estimated entry time, estimated passage time, and estimated departure time for each branch route, forming a corresponding flight time series.

[0096] Airspace congestion is used to determine whether there are airspace conflicts for each route during its expected passage period. The system determines the airspace occupancy of each route at different time periods based on the flight paths of other rescue drones, communication relay drones, rescue helicopters, temporary no-fly windows, and hovering operation points within the same area. If the flight time series of a route overlaps with the occupancy periods of other aircraft or with temporary no-fly windows, the priority of that route is reduced; if a route has low airspace occupancy, continuous communication, and no temporary no-fly restrictions during its expected passage period, the priority of that route is increased.

[0097] Based on the correlation analysis between flight time series and airspace congestion, branch routes are prioritized. During prioritization, impassable branch routes are first excluded, and then priority is determined in the order of high route feasibility score, short estimated transit time, low airspace congestion, and low energy consumption. For branch routes with similar scores, the system prioritizes those with smaller deviations from the original global flight route; if the deviations are the same, branch routes with fewer altitude changes and fewer turns are selected. After prioritization, the system obtains a priority sequence for each branch route.

[0098] Local flight segments are reconstructed based on a priority sequence. The branch ranked first is designated as the primary local flight segment, while lower-ranked branches serve as backup segments. The system connects the entry node of the primary local flight segment to the preceding segment of the global flight route and its exit node to the following segment, checking the continuity of altitude, heading angle, and speed at the connection point. If there are abrupt changes in altitude or heading at the connection point, the system smooths the connection segment, enabling the UAV to smoothly enter the local flight segment from the global route and smoothly return to the global route after passing through the risk area. This results in an optimized local flight segment, achieving precise planning for local risk areas.

[0099] S6 includes acquiring real-time point cloud data of a local flight segment and spatially registering it with a preset terrain benchmark model; determining the geometric boundaries of terrain collapse areas in the point cloud data; extracting elevation abrupt gradients and surface roughness changes from the terrain feature distribution based on the geometric boundaries; mapping the changes to a path cost matrix to determine the obstacle avoidance cost distribution of the local flight segment; calculating the overlap between the current trajectory and the obstacle avoidance cost distribution using a path deviation penalty function; if the overlap exceeds a preset risk threshold, determining a set of candidate paths; and iteratively calculating the cost function of the candidate path set through local optimization inspection to determine the corrected local flight path, thereby achieving dynamic adjustment of the local flight segment.

[0100] In this embodiment, S6 is used to identify the impact of newly added terrain collapses on the current flight path during the execution of a partial flight segment by the UAV, and to generate a corrected partial flight path when the original partial flight path no longer meets the safe passage conditions. The system first acquires real-time point cloud data ahead of the partial flight segment. This point cloud data is collected by the lidar onboard the UAV and includes the spatial locations of the mountain surface, slope edges, rockfall piles, collapse sections, and temporary obstacles within the current flight segment. The preset terrain reference model is derived from the terrain sub-map generated at the previous planning time, recording the reference elevation, slope, passable nodes, obstacle nodes, and original partial flight path locations within the area.

[0101] When spatially registering real-time point cloud data with a terrain reference model, the system first transforms the real-time point cloud into the local spatial coordinate system of the terrain reference model based on the UAV's current positioning coordinates, flight altitude, and attitude angle. Then, the system selects features that will not undergo significant displacement in a short period, such as mountain slopes, rock face edges, and road remnants, as registration references, aligning the corresponding reference points in the real-time point cloud with those in the terrain reference model. After registration, the same geographical location has consistent spatial coordinates in both the real-time point cloud and the terrain reference model, allowing the system to compare elevation changes between the two.

[0102] The geometric boundaries of the terrain collapse area are determined by the difference between the real-time elevation and the reference elevation. The system reads the real-time elevation region by region in the registered point cloud and compares it with the reference elevation at the same location in the terrain reference model. When the real-time elevation of a certain region drops by more than 1 meter relative to the reference elevation, and the continuous horizontal coverage area of ​​the dropped region reaches 4 square meters, the system identifies the region as a candidate collapse area. The aforementioned 1-meter elevation drop threshold is determined based on the minimum safe altitude margin for UAV low-altitude rescue flight and the natural undulation error of mountainous terrain, and is used to exclude ordinary uneven ground; the 4-square-meter area threshold is determined based on the UAV fuselage outline, the influence range of rotor downwash, and the point cloud positioning error, and is used to exclude misjudgments caused by single-point noise or small-area missing stones.

[0103] After identifying candidate subsidence areas, the system checks the continuity of elevation changes along the outer edge of these areas. If a continuous drop boundary exists at the outer edge of the candidate area, and a stable height difference is formed between the inner and outer sides of the boundary, the system determines this outer edge as the geometric boundary of the terrain subsidence area. For cases where the boundary has discontinuities, the system fills in the discontinuities based on the elevation change directions of adjacent grids, ensuring the subsidence boundary forms a complete spatial extent. This geometric boundary is used to define the locations of subsequent high-risk areas, no-fly buffer zones, and candidate detour areas.

[0104] The system extracts elevation abrupt changes and surface roughness variations from the terrain feature distribution based on geometric boundaries. The elevation abrupt change gradient is determined by the difference in elevation between the inner and outer sides of the collapse boundary; a larger difference indicates a steeper collapse section and a higher risk of passage. The surface roughness variation is determined by the difference between the point cloud height dispersion within the current collapse area and the original surface roughness in the baseline model. When the point cloud height distribution is discrete and accompanied by fractured boundaries, the system determines that the area has loose soil or a risk of secondary collapse. These variations are written to the corresponding grid nodes, updating the original terrain feature distribution to reflect the current collapse state as local features.

[0105] The path cost matrix is ​​generated jointly by the elevation abrupt change gradient, surface roughness variation, collapse geometric boundary location, and relative track position. The system defines the grid inside the collapse geometric boundary as an impassable area, the grid extending 5 meters outward from the collapse boundary as a no-fly buffer zone, and the grid extending 5 to 10 meters outward as a high obstacle avoidance cost zone. The 5-meter no-fly buffer distance is determined based on the UAV's minimum obstacle avoidance distance, positioning error, flight control response distance, and the safety margin for further expansion of the collapse edge; the 10-meter high obstacle avoidance cost distance is used to cover the operational space for the UAV to decelerate, turn, and climb during low-altitude flight. Through this mapping, the system obtains the obstacle avoidance cost distribution within a local flight segment.

[0106] The overlap between the current trajectory and the obstacle avoidance cost distribution is used to determine whether the original local flight path is still executable. The system selects trajectory sampling points at equal intervals along the current trajectory and reads the obstacle avoidance cost of the grid where each sampling point is located. When a sampling point falls into a collapsed area, a no-fly buffer zone, or a high obstacle avoidance cost area, it is recorded as a risk overlap point. The system combines the number of risk overlap points, the proportion of risk overlap points in the entire trajectory, the distance between risk overlap points and the geometric boundary of the collapsed area, and the degree of deviation of the current trajectory from the original planned flight segment to obtain the overlap between the current trajectory and the obstacle avoidance cost distribution.

[0107] In this implementation, the risk threshold is set at 60%. This threshold is determined based on the UAV's obstacle avoidance safety distance, point cloud ranging error, positioning error, flight control response distance, and unstable outward expansion distance of the collapse boundary. Specifically, when the number of overlapping risk points is sufficient to cause the UAV's continuous flight segment to enter the no-fly buffer zone, or when the distance between the current flight path and the collapse boundary is less than the safety margin, the system will classify the corresponding overlap as 60% or higher. Therefore, an overlap exceeding 60% indicates that the current flight path substantially overlaps with the collapse risk area, and the original local flight segment will no longer be used as the execution route.

[0108] When the overlap exceeds the risk threshold, the system determines a set of candidate paths. Candidate paths are generated from the drone's current position within a local spatial range: 30 meters in front, 15 meters to the left and right, and 8 meters vertically above and below the current position. The 30 meters in front is determined based on the drone's current cruising speed and flight control response distance, ensuring the drone can change course before reaching the collapse area. The 15 meters to the left and right are determined based on the narrow mountain passages and the effective recognition range of local sensors. The 8 meters vertically above and below the altitude are determined based on the drone's safe climb and descent capabilities within this flight segment, as well as the extent of mountain obstruction. All candidate paths avoid the collapse area and its no-fly zone, and reconnect with unaffected passable nodes in the original local flight segment.

[0109] Local optimization inspection is used to determine the corrected local flight path from the candidate path set. The system sequentially checks the obstacle avoidance cost, path deviation, segment length, altitude change, number of turns, communication status, and connection status with subsequent nodes for each candidate path. Candidate paths that cross no-fly zones, have insufficient safety distances, have turning radii that do not meet the UAV's maneuverability requirements, or have discontinuous altitude changes are deleted. For the retained candidate paths, the system iteratively compares them in the order of lowest obstacle avoidance cost, controlled deviation from the original flight path, shortest detour distance, gentlest altitude change, and continuous communication status. In each round of comparison, the system adjusts the intermediate nodes of the candidate paths to keep them away from collapse boundaries and reduce unnecessary turns.

[0110] After completing the local optimization inspection, the system selects the candidate path with the lowest total cost and that satisfies flight constraints as the corrected local flight path. The entrance of this local flight path connects to the current position of the UAV, and the exit connects to the subsequent nodes in the original local flight segment that are not affected by the collapse. The system further checks the continuity of altitude, heading angle, and speed at the connection point; if there are abrupt changes in altitude or heading, the connection segment is smoothed.

[0111] S7 includes acquiring elevation data and local flight path trajectories of mountainous areas, constructing a 3D terrain raster model, projecting the local flight path onto the surface of the 3D terrain raster model to obtain a spatial matching degree matrix; based on the spatial matching degree matrix, using a cubic spline interpolation algorithm to smooth the local flight path and generate corrected flight path trajectory data; inputting the corrected flight path trajectory data into a multi-objective optimization function, and using a genetic algorithm to determine the optimal cooperative travel path coordinate sequence; associating and matching the optimal cooperative travel path coordinate sequence with historical data, and obtaining a UAV cooperative travel plan for mountainous area rescue by correcting the offset.

[0112] In this implementation, step S7 is used to reconnect the corrected local flight path obtained in step S6 with the global flight path, forming a final route suitable for the coordinated passage of multiple UAVs. The system first reads the elevation data of the mountainous area and the local flight path trajectory. The elevation data of the mountainous area comes from a structured terrain grid dataset and real-time point cloud update results, including the ground elevation, slope, obstacle boundaries, no-fly zones, buffer zones, and passable altitude layers at each grid location. The local flight path trajectory consists of several consecutive track points, each track point recording spatial coordinates, flight altitude, heading angle, speed, and time sequence.

[0113] The system constructs a 3D terrain grid model based on elevation data of mountainous areas. During construction, the system uses the horizontal position of the terrain grid as a basis, superimposing passable spaces at different height levels into the same spatial frame; each grid cell is recorded with ground elevation, safe ground clearance, obstacle status, risk status, and passability status. Grids corresponding to rockfall areas, collapse areas, fallen tree areas, and communication obstruction areas are marked with risk tags; spaces that meet the requirements of safe height, unobstructed occupancy, and connectivity with adjacent grids are marked with passability tags. Thus, the mountain surface and UAV-flyable airspace are uniformly represented as a computable 3D terrain grid model.

[0114] When a local flight path is projected onto the surface of a 3D terrain grid model, the system locates the corresponding grid cell based on the spatial coordinates of each flight point and reads the ground elevation, safe altitude range, and risk markers of that grid cell. If the flight altitude of a flight point is higher than the corresponding ground elevation and meets the safe takeoff altitude requirement, and the flight point does not fall into an obstacle grid, no-fly zone grid, or high-risk buffer zone, then the flight point is determined to be a matching point. If the flight point's altitude is insufficient, it crosses an obstacle area, or enters a no-fly zone, then the flight point is determined to be a mismatched point. The system records the above judgment results in the order of the flight points, forming a spatial matching degree matrix.

[0115] The spatial matching degree matrix represents the correspondence between local flight paths and the 3D terrain grid model. The matrix records the altitude matching status, risk matching status, connectivity matching status, and deviation status relative to the global flight path for each waypoint. The altitude matching status determines whether the waypoint meets safe takeoff requirements; the risk matching status determines whether the waypoint avoids rockfalls, landslides, and no-fly zones; the connectivity matching status determines whether there is continuous passable space between adjacent waypoints; and the deviation status determines the connection deviation between the local flight path and the global flight path. Using this matrix, the system can identify locations within the local flight path that require smoothing, correction, or reconnection.

[0116] The system smooths local flight paths based on the spatial matching matrix. For locations in the matrix that show abrupt changes in altitude, heading angle, insufficient connectivity between adjacent trajectory points, or discontinuities with the global flight path, the system selects multiple normally matching track points before and after that location as interpolation control points. A cubic spline interpolation algorithm generates continuous trajectories based on the spatial coordinates and temporal order of these control points, ensuring continuity in altitude, heading angle, and curvature for the local flight path. For track points already identified as mismatched, the system first moves them to the nearest passable grid cell before participating in the interpolation process, preventing the smoothed flight path from re-entering obstacle or no-fly zones.

[0117] After smoothing, the system generates corrected flight path data. This data includes the smoothed waypoint coordinates, flight altitude, heading angle, segment length, estimated passage time, and risk markers. The corrected flight path data retains the local obstacle avoidance effect formed in step 6, while eliminating sharp turns, sudden climbs, sudden descents, and trajectory breaks, enabling the UAV to smoothly enter the local detour segment from the global route and smoothly return to the global route after passing through the risk area.

[0118] Subsequently, the corrected flight path data is input into a multi-objective optimization function. This function simultaneously evaluates safety risk, flight distance, energy consumption, flight time, communication status, and airspace conflicts among multiple UAVs. Safety risk is determined based on the distance between the flight path and rockfall areas, landslide areas, and no-fly zones; flight distance is determined by the sum of spatial distances between consecutive waypoints; energy consumption is determined by the number of climbs, descents, turns, and segment length; flight time is determined by segment length, cruise speed, and turn time; communication status is determined based on whether the flight path passes through communication obstruction areas; and airspace conflicts are determined based on whether different UAVs occupy the same altitude layer and the same local area within the same time period.

[0119] A genetic algorithm is used to determine the optimal path coordinate sequence from multiple candidate cooperative travel paths. A candidate path is represented as multiple path coordinate points arranged in flight order, and the paths of multiple UAVs are combined into a cooperative travel scheme. Each scheme undergoes a safety constraint check; schemes that cross no-fly zones, have insufficient safe distances from obstacles, experience airspace conflicts with other UAVs, or suffer communication interruptions are eliminated. The remaining schemes are evaluated for adaptability based on flight distance, energy consumption, travel time, and cooperative conflict levels. The system continuously updates candidate schemes through selection, crossover, and mutation processes. After multiple iterations, the path coordinate sequence with the best overall evaluation is retained as the optimal cooperative travel path coordinate sequence.

[0120] Once the optimal coordinated travel path coordinate sequence is determined, the system correlates and matches it with historical data. Historical data includes past mountain rescue routes, verified safe passages, locations prone to rockfalls, collapse records, communication blind spot records, and multi-UAV coordinated flight records. The system compares the current path coordinate sequence with the spatial location in the historical data to determine whether the current path is close to historical risk points, whether it passes through verified safe passages, whether it enters historical communication blind spots, and whether it conflicts with existing coordinated flight altitude levels. This correlation and matching is used to further refine the real-time planning results using existing rescue experience.

[0121] The offset correction is determined based on the spatial relationship between the current path and historical risk points and historical safe passages. If the current path is close to historical rockfall points, landslide points, or communication blind spots, the system will offset the corresponding flight segment away from the risk point and check whether the offset flight segment is still within the passable grid. If the current path deviates from a historically verified safe passage, and the safe passage is not affected by new obstacles, the system will correct the corresponding flight segment towards the safe passage. If the offset causes the flight path to enter a no-fly zone, increases airspace conflicts, or disrupts communication continuity, the system will cancel the offset and retain the original optimized path.

[0122] After completing the correlation matching and offset correction, the system obtains a collaborative passage plan for UAVs in mountainous areas. This plan includes the final path coordinate sequence, flight altitude for each segment, estimated transit time, communication relay connection points, risk avoidance zones, and airspace allocation results for multiple UAVs. When multiple UAVs are performing missions, the system allocates passage paths according to different altitude layers, entry times, and mission areas, ensuring that reconnaissance UAVs, supply delivery UAVs, and communication relay UAVs maintain a safe distance within the same mountainous area and complete collaborative passage according to the final plan.

[0123] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-source data fusion UAV path planning method for emergency rescue, characterized in that, include: S1. Real-time terrain undulation and temporary obstacle distribution information are obtained from the mountain disaster area environment through a multi-source sensor array carried by the UAV. The collected terrain data is spatially divided into blocks using grid partitioning technology and integrated into a structured terrain grid dataset to determine the preliminary spatial distribution of the UAV flight environment. S2. Based on the structured terrain grid dataset, key flight nodes and areas with dense obstacles are selected using a node priority sorting method. Combined with terrain complexity analysis, features are extracted from the grid data to obtain a simplified terrain feature representation. S3. If the proportion of densely obstructed areas in the simplified terrain feature representation exceeds a preset threshold, the newly added terrain change information is obtained from the sensor array on the UAV through a real-time data update mechanism. Combined with dynamic obstacle avoidance technology, the feature representation is locally corrected to determine the updated terrain feature distribution. S4. Based on the updated terrain feature distribution, the path cost calculation method is used to search for flight nodes in the global range. The feasibility of potential routes is evaluated by combining distance estimation method and obstacle weight allocation technology to obtain a preliminary global flight route. S5. By using the preliminary global flight route, obtain detailed terrain sub-maps of local risk areas, and use route feasibility scoring and time consumption prediction techniques to prioritize the branches within the sub-maps and determine the optimized local flight segments.

2. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that: S1 includes: Acquire raw point cloud data and multispectral images collected by a multi-source sensor array carried by a drone; Feature fusion is performed on raw point cloud data and multispectral imagery to extract real-time terrain undulation features and temporary obstacle distribution features; The real-time terrain undulation features and temporary obstacle distribution features are spatially divided into blocks using grid partitioning to obtain discretized terrain grid units. Discretized terrain grid units are logically associated according to preset topological relationships to form a structured terrain grid dataset; Based on the structured terrain mesh dataset, a 3D reconstruction was performed to determine the preliminary spatial distribution of the UAV flight environment.

3. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that: S2 includes: Obtain the node elevation gradient and neighborhood connectivity of the structured terrain grid dataset to determine the feature matrix of the terrain grid nodes; Based on the feature matrix of the terrain grid nodes, a node priority ranking algorithm is used to evaluate the weight of the grid nodes and determine the set of key flight nodes. By performing a difference operation between the key flight node set and the terrain grid node feature matrix, a spatial distribution map of densely obstructed areas is obtained. The terrain complexity distribution model is determined by weighted fusion of the spatial distribution map of densely obstacle areas and the feature matrix of terrain grid nodes; A terrain complexity distribution model is used to perform feature dimensionality reduction on terrain grid nodes, resulting in a simplified representation of terrain features.

4. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that: S3 includes: The proportion of areas with dense obstacles is extracted from simplified terrain features. If the proportion of areas exceeds the judgment threshold, new terrain change information is obtained through a sensor array. The adjustment range of the local topology is determined based on the terrain change information. The adjustment range is used to guide the correction operator to make local corrections to the simplified terrain features. The simplified terrain features are reconstructed using the correction operator to obtain the feature mapping model, and the global consistency of the feature mapping model is checked to determine the updated terrain feature distribution.

5. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that: S4 includes: The global flight node set is obtained based on the distribution of terrain features. The global flight node set contains the spatial coordinates and connectivity data of each node. The path cost calculation method is used to traverse the global flight node set to determine the path cost values ​​between each node. Based on the path cost values ​​and the preset distance estimation method, the distance weight distribution of each potential route is generated; By combining distance weight distribution and obstacle weight allocation techniques, a feasibility assessment of potential routes is conducted. The optimal path was selected based on the feasibility assessment results, resulting in a preliminary global flight route.

6. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that: S5 includes: Based on the geospatial extent of the global flight path, terrain feature data is extracted to construct terrain sub-maps of local risk areas. By analyzing the spatial overlap of terrain complexity and dynamic obstacles within the terrain sub-maps, the risk level value within each sub-map is determined. If the risk level value exceeds the preset threshold, the branch roads in the topographic sub-atlas are weighted using route connectivity and weather interference to obtain the route feasibility score for each branch road. Based on the mapping relationship between route feasibility score and flight energy consumption ratio, and combined with flight timestamps, time consumption prediction is performed on branch routes to determine the flight time series of each branch route. By analyzing the correlation between flight time series and airspace congestion, the routes are prioritized to obtain the priority sequence of each route. By reconstructing local flight segments based on priority sequences, optimized local flight segments are obtained, thus achieving precise planning of local flight segments.

7. The multi-source data fusion UAV path planning method for emergency rescue according to claim 1, characterized in that, This also includes S6. If rockfall or terrain collapse is detected in the real-time monitoring of the optimized local flight segment, relevant change features are extracted from the terrain feature distribution. Combined with path deviation penalty and local optimization inspection technology, the local flight segment is dynamically adjusted to determine the corrected local route. Specifically, this includes: Acquire real-time point cloud data of local flight segments and spatially register it with a preset terrain reference model; Determine the geometric boundaries of terrain collapse areas in the point cloud data; Extract the elevation gradient and surface roughness variation from the terrain feature distribution based on the geometric boundary; The mapping change is represented by the path cost matrix to determine the obstacle avoidance cost distribution for local flight segments.

8. The multi-source data fusion UAV path planning method for emergency rescue according to claim 7, characterized in that: S6 further includes: The overlap between the current trajectory and the obstacle avoidance cost distribution is calculated using a path deviation penalty function. If the overlap exceeds a preset risk threshold, a set of candidate paths is determined. By performing cost function iterative calculations on the candidate path set through local optimization inspection, the corrected local flight routes are determined, thereby realizing the dynamic adjustment of local flight segments.

9. The multi-source data fusion UAV path planning method for emergency rescue according to claim 7, characterized in that, This also includes S7. Based on the spatial matching degree between the corrected local flight path and the global flight path, path smoothing and multi-objective optimization techniques are used to integrate and adjust the two. Historical data is then used to refine the final route, resulting in a drone collaborative passage plan for mountain rescue operations. Specifically, this includes: Acquire elevation data and local flight path trajectories of mountainous areas, construct a three-dimensional terrain raster model, project the local flight path onto the surface of the three-dimensional terrain raster model, and obtain the spatial matching degree matrix. Based on the spatial matching degree matrix, a cubic spline interpolation algorithm is used to smooth the local flight path and generate corrected flight path trajectory data.

10. The multi-source data fusion UAV path planning method for emergency rescue according to claim 9, characterized in that: The S7 also includes: The corrected flight path data is input into a multi-objective optimization function, and the optimal cooperative travel path coordinate sequence is determined by a genetic algorithm. By associating and matching the optimal collaborative travel path coordinate sequence with historical data, and correcting the offset, a collaborative travel plan for drones in mountainous areas is obtained.

Images