A Low-Altitude Path Planning Method and System Based on Dynamic Adaptive Graph Neural Network

Abstract

This invention discloses a low-altitude path planning method and system based on a dynamic adaptive graph neural network, belonging to the field of artificial intelligence and robotic autonomous navigation technology. Addressing the problems of high computational cost of fixed grid maps, slow dynamic response of static octrees, and the failure of traditional heuristic functions on non-uniform graphs, this invention proposes the following solutions: constructing a resolution-adaptive non-uniform graph using an octree; employing Morton codes to achieve rapid localization of dynamic obstacles and millisecond-level map updates; introducing an attention mechanism into the graph attention network (GAT) to adaptively assign weights to different neighbor nodes, aggregating features and outputting a high-precision heuristic cost estimate; finally, integrating this estimate into the algorithm for efficient path search. This method significantly reduces memory usage and computational complexity, improving planning efficiency and path safety in dynamic environments.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and autonomous robot navigation, specifically relating to three-dimensional path planning technology for unmanned aerial vehicles (UAVs) in complex urban low-altitude environments. More specifically, this invention relates to a low-altitude path planning method and system based on a dynamic adaptive graph neural network. Background Technology

[0002] With the booming development of the low-altitude economy, the demand for drones in complex urban environments has increased dramatically. Unlike high-altitude flight, the urban low-altitude environment is characterized by dense obstacles (such as buildings, trees, and power lines), complex topology, and significant dynamic changes (such as moving vehicles and temporary construction cranes). Therefore, how to achieve efficient and real-time three-dimensional path planning while ensuring flight safety has become a key bottleneck restricting the industry's development.

[0003] Existing three-dimensional path planning technologies for low-altitude unmanned aerial vehicles (UAVs) can be mainly classified into the following categories, but all of them have obvious limitations.

[0004] The first category is path planning methods based on fixed 3D grids. This type of method was the mainstream approach in early path planning, and its core idea is to uniformly discretize the continuous 3D space into fixed-size cubic units (voxels). However, this type of method suffers from a serious "storage-computation contradiction." To accurately describe small obstacles, the grid size must be reduced, leading to a cubic explosion in the number of nodes (the curse of dimensionality), exceeding the load of onboard computing units; if the grid size is increased, small obstacles cannot be identified, posing a safety hazard.

[0005] The second category is path planning methods using static octrees. These methods introduce an octree structure to address storage issues, recursively partitioning the tree to use high-resolution nodes near obstacles and low-resolution nodes in open areas. However, traditional octree mapping is primarily suited for static environments. When dynamic obstacles appear, it often requires reconstructing the entire tree or a large area, resulting in enormous computational overhead and failing to meet the millisecond-level obstacle avoidance requirements of high-speed UAVs. Furthermore, traditional heuristic functions (such as Euclidean distance) fail on non-uniform meshes, leading to… The algorithm's search efficiency has dropped significantly.

[0006] The third category comprises preliminary attempts using graph neural networks (GNNs). These methods attempt to learn heuristic functions for path planning using GNNs. However, most existing GNN planning methods are trained on regular, fixed grid graphs, which cannot directly handle non-uniform graph structures like octrees, where node sizes vary and the number of neighbors changes dynamically. Furthermore, simple GNN models (such as GCNs) cannot distinguish the importance of different neighboring nodes, making it difficult to provide effective guidance information in complex topologies. Summary of the Invention

[0007] 1. The technical problem that the invention aims to solve This invention aims to address the problems in existing technologies, such as the huge computational resource consumption caused by using fixed grids, the slow dynamic response caused by static octrees, and the low search efficiency caused by the failure of traditional heuristic functions on non-uniform graphs. It provides a low-altitude path planning method and system based on a dynamic adaptive graph neural network. This invention utilizes a dynamically updated octree data structure to adaptively represent the three-dimensional space and combines it with a Graph Attention Network (GAT) for heuristic path search. This technical solution is particularly suitable for applications requiring high-precision environmental modeling and millisecond-level dynamic response, such as low-altitude logistics delivery, urban emergency rescue, power facility inspection, and security monitoring. 2. Technical Solution To achieve the above objectives, the technical solution provided by this invention is as follows: The present invention provides a low-altitude path planning method based on a dynamic adaptive graph neural network, comprising the following steps: Step 1: Use an octree data structure to perform multi-resolution recursive partitioning of the three-dimensional flight airspace and construct a non-uniform dual graph in which the node resolution adapts to the environmental complexity. Step 2: Introduce Morton codes to encode spatial nodes. Utilize the spatial locality principle of Morton codes to achieve rapid localization of leaf nodes containing dynamic obstacles. Perform local re-subdivision and merging only on the affected areas to achieve millisecond-level map updates and obtain a dynamically non-uniform dual graph updated in real time. Step 3: Deploy a graph attention network (GAT) on the dynamically non-uniform dual graph after real-time updates. The graph attention network (GAT) uses an attention mechanism to adaptively aggregate feature representations of neighbor nodes of different sizes and numbers, and heuristically infers a high-precision heuristic cost estimate containing environmental semantics. Step 4: Use the heuristic cost estimate output by the graph attention network GAT as... The algorithm's heuristic function performs path search on the dynamic non-uniform dual graph, quickly generating collision-free optimal waypoint paths.

[0008] Furthermore, step 1 includes: Step 1.1: Initialize the root node: Define the entire 3D flight airspace as a cubic space, which will serve as the root node of the octree. Let the side length of the root node be... The minimum resolution is ; Step 1.2, Obstacle Detection: For the current node, the intersection test between the obstacle geometry model and the current node's bounding box is performed to detect whether the current node's spatial range contains the obstacle boundary. Step 1.3, Recursive Partitioning Determination: If a node is completely occupied by obstacles, mark the node as "occupied" and terminate the partitioning; if a node is completely free space, mark the node as "free" and terminate the partitioning; if a node contains both obstacles and free space, and the node size is greater than... Then, the node is divided into 8 child nodes along the three coordinate axes of the three-dimensional space, and obstacle detection and recursive subdivision judgment are recursively performed on each child node. Step 1.4: Constructing a Non-Uniform Dual Graph: After partitioning, all leaf nodes marked as "empty" are used as nodes in the graph. If two leaf nodes are spatially adjacent, an undirected edge is created between them to form a non-uniform dual graph. ,in For a set of nodes, Let it be the set of edges.

[0009] Furthermore, in step 2, the Morton code uses a bit interleaving algorithm to convert the three-dimensional coordinates of the node into a unique index key value, thereby encoding the spatial node.

[0010] Furthermore, the process of achieving rapid dynamic obstacle localization in step 2 is as follows: Obtain the three-dimensional coordinates of the obstacle from the sensor; The world coordinates of the obstacle are converted into the discretized index coordinates of the octree, and then converted into Morton codes using a bit interleaving algorithm. The Morton codes are used as keys to search in the hash table storing the leaf nodes of the octree, instantly locating the leaf node containing the obstacle coordinates.

[0011] Furthermore, the process of achieving local re-subdivision in step 2 is as follows: If the located leaf node was originally marked as "idle" and its size is larger than the minimum resolution. If so, the leaf node will be split into 8 child nodes; Calculate a new Morton code for each newly generated child node and check whether the obstacle point cloud falls within its spatial range. If the obstacle point cloud falls within the child node, mark it as "occupied" or continue recursively splitting; if no obstacle point cloud falls within the child node, mark it as "free". Recursively perform splitting and marking on child nodes marked as partially occupied until all child nodes are completely free or occupied, or the minimum resolution is reached. .

[0012] Furthermore, the process of updating the graph in step 2 is as follows: Remove the split original leaf nodes and all their associated edges from the non-uniform dual graph; Insert all newly generated child nodes marked as "free" into the non-uniform dual graph as new graph nodes; For each newly inserted node, find its spatial neighbor nodes in the octree. If the neighbor nodes are also in an "empty" state, then create an undirected edge.

[0013] Furthermore, in step 3, the input node features of the graph attention network (GAT) include at least one of the following: the three-dimensional coordinates of the node center, the size of the node, the Euclidean distance from the node center to the nearest obstacle, the Euclidean distance from the node center to the target endpoint, and the octree depth of the node.

[0014] Furthermore, in step 3, the graph attention network GAT processes node features through multiple graph attention layers. Each layer contains K independent attention heads. Each attention head calculates the attention coefficient between a node and its neighboring nodes. After normalization by the Softmax function, the features of the neighboring nodes are weighted and aggregated to update the node feature representation. The graph attention network GAT calculates attention coefficients and adaptively assigns different weights to different neighboring nodes of the current node, thereby aggregating the features of neighboring nodes. The calculation of the attention coefficient includes: calculating the correlation score between the current node and its neighboring nodes, and processing it through a normalization function to obtain the normalized attention coefficient.

[0015] Furthermore, in step 4, The algorithm uses a cost function To evaluate each candidate node Priority:

[0016] in: Indicates starting from the origin To the current node The actual cumulative cost; This represents the output of the graph attention network GAT from the current node. To the finish line The estimated remaining cost.

[0017] This invention provides a low-altitude path planning system based on a dynamic adaptive graph neural network, used to implement the aforementioned low-altitude path planning method. The system includes: UAV onboard terminal and ground control station; The UAV's onboard unit is equipped with: a sensor module, a navigation module, and an onboard computing unit; The sensor module is used to detect environmental information in real time, including lidar and / or binocular camera; The navigation module is used to provide the UAV's positioning and attitude information, including an IMU / GPS navigation module; The airborne computing unit internally deploys: The dynamic map manager is configured to perform the following operations: use an octree data structure to perform multi-resolution recursive partitioning of the three-dimensional flight airspace, construct and maintain a dynamic non-uniform dual graph; and introduce Morton codes to achieve millisecond-level local updates of the map. The GAT inference engine is configured to perform the following operations: load and run a pre-trained graph attention network GAT model, perform inference on the dynamic non-uniform dual graph, and output a heuristic cost estimate. The path planner is configured to perform the following operation: use the heuristic cost estimate output by the GAT inference engine as... The algorithm's heuristic function performs path search on the dynamic non-uniform dual graph to generate collision-free optimal waypoint paths; The ground control station is connected to the UAV's onboard terminal and is used to issue flight missions and preload high-precision static maps to the UAV's onboard terminal.

[0018] 3. Beneficial effects Compared with the prior art, the technical solution provided by this invention has the following advantages: (1) The present invention provides a low-altitude path planning method based on dynamic adaptive graph neural network. By using octree adaptive graph construction, the number of nodes in the graph is reduced by 1-2 orders of magnitude compared with a fixed grid graph of the same precision in a typical urban low-altitude environment, which greatly reduces memory usage and computational complexity.

[0019] (2) The low-altitude path planning method based on dynamic adaptive graph neural network of the present invention, combined with the local dynamic update mechanism based on Morton code, enables the system to complete map update and graph topology repair within milliseconds after detecting new obstacles, improving the map update speed by 1-2 orders of magnitude (<5ms) and meeting the requirements of high-speed obstacle avoidance.

[0020] (3) The low-altitude path planning method based on dynamic adaptive graph neural network of the present invention is more accurate than the traditional Euclidean distance heuristic model. The number of path search nodes expanded in complex environments is reduced, which can effectively reduce invalid search paths and improve planning efficiency. At the same time, the environmental semantic information learned by GAT (such as avoiding dead ends and staying away from areas with dense obstacles) makes the planned path smoother and safer, which greatly improves flight safety. Attached Figure Description

[0021] Figure 1 This is a diagram showing the overall hardware and software architecture of the system according to an embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the graph attention network model architecture.

[0023] Figure 3 This is a flowchart of the overall path planning method. Detailed Implementation

[0024] To further understand the content of this invention, a detailed description of the invention will be provided in conjunction with the accompanying drawings and embodiments.

[0025] Example 1 See Figure 1 The path planning system in this embodiment mainly includes an UAV onboard terminal and a ground control station.

[0026] The UAV's onboard unit is equipped with a LiDAR, binocular camera, IMU / GPS navigation module, and a high-performance embedded computing unit. The computing unit contains a dynamic map manager, a GAT inference engine, and a path planner.

[0027] The ground control station is responsible for task distribution and preloading of high-precision static maps (BIM / DEM).

[0028] See Figure 3 Based on the aforementioned path planning system, this embodiment provides a low-altitude path planning method based on a dynamic adaptive graph neural network, which specifically includes the following steps: Step 1: Utilize an octree data structure to perform multi-resolution recursive partitioning of the 3D flight airspace, constructing a non-uniform graph where node resolution adaptively varies with environmental complexity. Specifically: During system initialization, or when a new mission area is assigned by the ground control station, the dynamic map manager in the airborne computing unit performs multi-resolution spatial recursive partitioning of the three-dimensional flight airspace using an octree data structure. The system first constructs an initial octree based on a high-precision prior static map (such as BIM / DEM data).

[0029] The specific process for step 1 is as follows: Step 1.1: Initialize the root node: Define the entire 3D flight airspace as a cubic space, which will serve as the root node of the octree. Let the side length of the root node be... The minimum resolution is (e.g., 0.1 meters).

[0030] Step 1.2, Obstacle Detection: For the current node, detect whether its spatial range contains obstacle boundaries. Specifically, perform an intersection test between the obstacle's geometric model (point cloud or triangular mesh) and the current node's bounding box.

[0031] Step 1.3, Recursive Partition Determination: If a node is completely occupied by obstacles, then mark the node as "Occupied" and terminate the partitioning. If a node contains completely free space, then mark the node as "Free" and terminate the partitioning. If a node contains both obstacles and free space (i.e., includes the obstacle boundary), and the node size is greater than... Then, the node is divided into 8 child nodes along the three coordinate axes of the three-dimensional space, and steps 1.2 and 1.3 are recursively executed for each child node until the recursive partitioning of the three-dimensional flight airspace is completed.

[0032] Step 1.4: Constructing a Non-Uniform Dual Graph: After partitioning, all leaf nodes marked as "empty" are treated as nodes in the graph. If two leaf nodes are spatially adjacent (sharing a face, edge, or vertex), an undirected edge is created between them. Thus, the octree structure is transformed into a searchable non-uniform dual graph structure. ,in For a set of nodes, Let it be the set of edges.

[0033] Step 2: Introduce Morton code (Z-order curve) to encode spatial nodes. Utilize its spatial locality principle to achieve the encoding of leaf nodes containing dynamic obstacles. It achieves rapid location with low time complexity and performs local repartitioning and merging only on the affected areas, realizing millisecond-level map updates. To efficiently store and retrieve nodes of varying sizes in computer memory, this embodiment uses Morton codes as the unique index key for each node. Morton codes are generated by mapping the three-dimensional coordinates of the nodes. It is generated by interleaving the binary bits.

[0034] Assume the coordinates of a node are .

[0035] Interleave the binary bits of the three coordinates in sequence: Take 1st place 1st place The first one, then The second one... and so on.

[0036] The generated Morton code is:

[0037] Just like shuffling playing cards, the three decks (x, y, z coordinates) are perfectly shuffled into one deck. The resulting integer (Morton code) contains complete three-dimensional position information of the nodes.

[0038] The formula for converting node spatial coordinates to Morton code: Let the three-dimensional coordinates of the node be... Each coordinate is used Morton code is represented by 1 bit in binary. The calculation formula is:

[0039] in, The node coordinates are respectively The Bit binary value.

[0040] In three-dimensional space, physically adjacent nodes typically have very similar Morton code values. Leveraging this spatial locality allows computers to fully utilize the CPU cache when reading map data, significantly improving read speed. Traditional octrees require storing a large number of pointers (parent nodes pointing to child nodes), consuming significant memory. Using Morton codes, all leaf nodes can be stored in a linear hash table (Hash Map) or sorted array, completely eliminating pointer overhead.

[0041] Furthermore, given the spatial coordinates of any node, its Morton code can be directly calculated through bitwise operations without traversing the root of the tree. This allows for the identification of leaf nodes containing dynamic obstacles. The system offers fast localization with minimal time complexity. After rapid localization, only the affected area (where obstacles exist) needs to be locally re-partitioned and merged, achieving millisecond-level map updates. Ultimately, a complete octree spatial representation structure indexed by Morton codes and the corresponding dynamic non-uniform dual graph are obtained. .

[0042] The following sections provide a detailed introduction to the processes of rapid dynamic obstacle localization, local re-partitioning, and graph topology update: During the drone's flight, the onboard lidar or binocular camera detects new obstacles (such as birds, moving vehicles, temporary tower cranes, etc.) or the disappearance of existing obstacles in real time. The dynamic map manager in the onboard computing unit performs the following operations based on the constructed initial octree structure and Morton code indexing mechanism to ensure the real-time performance and accuracy of the map.

[0043] Rapid obstacle localization process: When the lidar detects point cloud data of a new obstacle, the system maps the obstacle's location to the leaf nodes of an octree using the following steps: Step 2.1: Obtain the three-dimensional coordinates of the obstacle from the sensor. .

[0044] Step 2.2: First, convert the world coordinates into discretized index coordinates of the octree:

[0045]

[0046]

[0047] in, This represents the minimum coordinate boundary of the spatial domain. To achieve the minimum resolution, a bit-interleaving algorithm is then used. Convert to Morton code .

[0048] Step 2.3, with The key is used to look up the information in the hash table storing the leaf nodes of the octagonal tree. Since the time complexity of a hash table lookup is O(n log n), the search operation is performed in the hash table. The system can instantly locate the leaf node containing the coordinates of the obstacle. .

[0049] Local repartitioning process: If leaf node The area was originally marked as "idle" because a new obstacle has intruded into it, requiring re-subdivision. Step 2.4: Check the leaf nodes Is the size larger than the minimum resolution? If yes, proceed to the splitting process in step 2.5; otherwise, directly mark the node as "occupied" and complete the update.

[0050] Step 2.5: Move the leaf nodes Divide the node into 8 child nodes along the three coordinate axes. Calculate a new Morton code for each child node by appending 3 bits (0-7) to the parent node's Morton code, representing the child node's relative position within the parent node.

[0051] Step 2.6: For each newly generated child node, check whether the obstacle point cloud falls within its spatial range; if the obstacle point cloud falls within the child node, mark it as "occupied" or continue recursively splitting; if no obstacle point cloud falls within the child node, mark it as "free".

[0052] Step 2.7: For child nodes marked as partially occupied, repeat steps 2.5 and 2.6 until all child nodes are completely free or occupied, or the minimum resolution is reached. .

[0053] Graph topology update process: Step 2.8, from the non-uniform dual graph Delete the split original leaf node and all its associated edges.

[0054] Step 2.9: Insert all newly generated child nodes marked as "free" into the non-uniform dual graph as new graph nodes. G .

[0055] Step 2.10: For each newly inserted node Find its spatial neighbors in the octree (adjacent leaf nodes sharing a face, edge, or vertex). If the neighboring node is also "free", then in the newly inserted node... A new undirected edge is established between the neighbor node and the neighbor node. Neighbor node search utilizes the bitwise operations of Morton codes. Complete within the specified time.

[0056] Updated Morton code-indexed octree structure and dynamic non-uniform dual graph It accurately reflects the latest distribution of obstacles in the current environment. The entire update process involves only local data changes and typically takes less than 5 milliseconds.

[0057] Step 3: The dynamically updated non-uniform dual graph obtained in step 2. The graph attention network (GAT) is deployed on top. This graph attention network model uses an attention mechanism to adaptively aggregate feature representations of neighbor nodes of different sizes and numbers, and heuristically infers high-precision heuristic cost estimates that include environmental semantics (such as avoiding dead ends and moving away from dense areas).

[0058] The GAT inference engine in the onboard computing unit will initiate path planning or replanning if any of the following occurs: -The UAV receives a new flight mission from the ground control station (including the coordinates of the start and end points); - Step 2 detected a change in obstacles, which may cause the currently planned path to collide with the obstacles; - The drone deviated from its original flight path by more than a preset threshold.

[0059] The Graph Attention Network (GAT) model used in this invention is adaptively designed for the non-uniform dual graph structure generated by octrees, and its architecture is as follows: Figure 2 As shown, it mainly includes the following components: Input layer: Receives node feature vectors ,in, The feature dimension is denoted by . The feature vector of each node contains: 3D coordinates of the node center ; The size of a node (the cube of the side length). ; Euclidean distance from the node center to the nearest obstacle ; Euclidean distance from the node center to the target endpoint ; octree depth of a node (Reflects the resolution level).

[0060] Attention layer (total) Layers): Each layer contains Each attention head independently learns the relationships between nodes from different subspaces. Each attention head independently calculates attention coefficients and aggregates neighbor features, ultimately combining them... The output of each attention head is spliced ​​or averaged.

[0061] Output layer: A fully connected layer that maps the output of the last graph attention layer to a scalar value. , which represents the estimated cost from the current node to the destination.

[0062] Traditional Graph Convolutional Networks (GCNs) aggregate the feature vectors of neighboring nodes by averaging them, essentially assuming that all neighboring nodes are equally important in terms of orientation. In contrast, the GAT implementation in this example introduces an "attention coefficient," which adaptively assigns different weights to different neighboring nodes.

[0063] The model "focuses" like the human eye. It calculates the correlation score between the current node and each of its neighboring nodes.

[0064] For example, for a neighboring node located at the entrance of a dead end, the model will assign it a very low weight (meaning "don't go that way").

[0065] The model assigns a high weight to a neighboring node that points to the target point and is surrounded by open space.

[0066] For nodes and its neighboring nodes (in Represents a node (neighbor set), attention coefficient The calculation process is as follows: First, through a shared linear transformation matrix Mapping node feature vectors to a higher-dimensional space:

[0067] Then, calculate the raw attention score. :

[0068] in: This is a learnable attention weight vector; This represents the vector concatenation operation, which is... and Connecting the head and tail to form a dimensional vector; Let be a linear rectified activation function with leakage, defined as .

[0069] Finally, the attention scores of all neighboring nodes are normalized using the Softmax function to obtain the attention coefficients. :

[0070] The node is updated by weighting and aggregating the feature vectors of neighboring nodes using attention coefficients. Feature representation :

[0071] in, It is a non-linear activation function (such as ELU or ReLU).

[0072] To enhance the expressive power of the model, use Each independent attention head is computed in parallel, and the final output feature representation is... for:

[0073] in, This indicates a splicing operation. and The first Attention coefficients and transformation matrices for each attention head.

[0074] The heuristic reasoning process in this embodiment is as follows: Step 3.1: The airborne computing unit preloads an offline-trained GAT model. In a large-scale simulation environment, the model is trained using labeled optimal path data for supervised learning. The learning objective is to minimize the mean square error between the predicted cost and the actual cost.

[0075] Step 3.2, when the path search algorithm ( It is necessary to evaluate a candidate node. When prioritizing a node, input the node and its surrounding local subgraph into the GAT model. Specifically, extract the node... Centered on All nodes within the jump range and their connections constitute a local subgraph.

[0076] Step 3.3: The GAT model aggregates nodes through multi-layer graph attention, comprehensively considering the nodes. The function takes its own feature representation, the feature representation of its neighboring nodes, and the local topology as input, and outputs a scalar value. This value represents the model's prediction of the number of slave nodes. The remaining flight cost to reach the destination (which can be distance, time, or energy consumption).

[0077] For dynamic non-uniform dual graphs any node in The GAT model can output a high-quality heuristic cost estimate. This value is more accurate than that obtained by the traditional Euclidean distance heuristic algorithm, and can more accurately reflect the actual difficulty of passage, effectively guiding subsequent path search.

[0078] Step 4: Calculate the heuristic cost estimate output by the GAT model. As The algorithm's heuristic function performs path search on a dynamically adaptive graph, quickly generating collision-free optimal waypoints.

[0079] The path planner in the airborne computing unit uses the heuristic cost estimate output by the GAT model. In dynamic adaptive dual graph Up, start The search algorithm performs a path search.

[0080] The algorithm uses the following cost function To evaluate each candidate node Priority:

[0081] in: Indicates starting from the origin To the current node The actual cumulative cost. This value is dynamically calculated during the search process, after each edge is traversed. Accumulate the weight of that edge. Edge weight can be defined as the Euclidean distance between the centers of two nodes, i.e. .

[0082] : Indicates starting from the current node To the finish line The estimated residual cost is output by the GAT model. This estimated cost contains predictive information about complex structures such as dead ends, narrow passages, and areas with dense obstacles.

[0083] use The process of the search algorithm performing path search is as follows: Step 4.1: Create two collections: an open list and a closed list; set the starting point... Add to open list, settings , The list is closed and initialized to empty.

[0084] Step 4.2: Select from the open list The node with the smallest value As the current extended node. The smaller the value, the more likely the node is to be on the optimal path.

[0085] Step 4.3, if node End point If the search is successful, proceed to step 4.6 to backtrack the path.

[0086] Step 4.4, set the nodes Moving from the open list to the closed list, for nodes Each neighbor node : If node Already on the close list, skip; Calculation via node Reaching the node The cost: ; If node Not on the open list, or If a better path is found, then update the node. Estimated cost value: , And record its parent node as ; will node Add to the open list (if you haven't already).

[0087] Step 4.5: If the open list is empty but the destination is still not found, then there is no feasible path and the planning fails; otherwise, return to step 4.2 to continue the search.

[0088] Step 4.6, from the endpoint Start by traversing back to the starting point along the parent node pointers. This yields a complete node sequence. Connecting the center coordinates of each node in the node sequence forms a three-dimensional waypoint path.

[0089] The algorithm guarantees finding the optimal path, provided the heuristic function satisfies acceptability. Acceptability requires that for any node... Heuristic cost Not greater than The true minimum cost to reach the destination ,Right now:

[0090] The GAT model in this invention is trained through supervised learning, with the learning objective being to fit the true cost. In practice, it can provide heuristic estimates that are close to the true value but slightly smaller than the true value, thus satisfying the admissibility condition.

[0091] Finally, a connection starting point is obtained. and the end point A safe and efficient three-dimensional flight path, consisting of a series of waypoint coordinates. The path is designed to avoid collisions with any obstacles, and guided by the GAT model heuristic, its search efficiency is significantly higher than traditional methods. The algorithm has been significantly improved.

[0092] This embodiment employs octree adaptive mapping, which reduces the number of nodes in a typical urban low-altitude environment (1km × 1km × 0.3km) by 1-2 orders of magnitude compared to a fixed raster map with the same precision (0.1m resolution). The number of nodes has been reduced to ~ The number of nodes greatly reduces memory usage (from TB to GB) and computational complexity.

[0093] The local map dynamic update mechanism based on Morton codes enables the system to complete map updates and graph topology repair within 5 milliseconds after detecting new obstacles, meeting the real-time obstacle avoidance requirements of UAVs flying at a speed of 10 m / s.

[0094] Compared to traditional Euclidean distance heuristic models, the GAT heuristic model reduces the number of path search nodes by 30% to 50% in complex environments and shortens planning time by more than 40%. Furthermore, the environmental semantic information learned by the GAT model (such as avoiding dead ends and staying away from areas with dense obstacles) makes the planned path smoother and safer, greatly improving flight safety.

Claims

1. A low-altitude path planning method based on a dynamic adaptive graph neural network, characterized in that, Includes the following steps: Step 1: Use an octree data structure to perform multi-resolution recursive partitioning of the three-dimensional flight airspace and construct a non-uniform dual graph in which the node resolution adapts to the environmental complexity. Step 2: Introduce Morton codes to encode spatial nodes. Utilize the spatial locality principle of Morton codes to achieve rapid localization of leaf nodes containing dynamic obstacles. Perform local re-subdivision and merging only on the affected areas to achieve millisecond-level map updates and obtain a dynamically non-uniform dual graph updated in real time. The process of performing local repartitioning is as follows: If the located leaf node was originally marked as "idle" and its size is larger than the minimum resolution. If so, the leaf node will be split into 8 child nodes; For each newly generated child node, a new Morton code is calculated by appending 3 bits to the parent node's Morton code to represent the child node's relative position within the parent node. It also checks whether the obstacle point cloud falls within its spatial range. If the obstacle point cloud falls within the child node, it is marked as "occupied" or the recursive splitting continues. If no obstacle point cloud falls within the child node, it is marked as "free". Recursively perform splitting and marking on child nodes marked as partially occupied until all child nodes are completely free or occupied, or the minimum resolution is reached. ; Step 3: Deploy a Graph Attention Network (GAT) on the dynamically non-uniform dual graph after real-time updates. The GAT utilizes an attention mechanism to adaptively aggregate feature representations of neighbor nodes of different sizes and numbers, and heuristically infers a high-precision heuristic cost estimate that includes environmental semantics. ; Among them, the graph attention network GAT processes node features through multiple graph attention layers. Each layer contains K independent attention heads. Each attention head calculates the attention coefficient between a node and its neighboring nodes. After normalization by the Softmax function, the features of the neighboring nodes are weighted and aggregated to update the node feature representation. The graph attention network GAT calculates attention coefficients and adaptively assigns different weights to different neighboring nodes of the current node, thereby aggregating the features of neighboring nodes. The calculation of the attention coefficient includes: calculating the correlation score between the current node and its neighboring nodes, and processing it through a normalization function to obtain the normalized attention coefficient; Step 4: Use the heuristic cost estimate output by the graph attention network GAT as... The algorithm's heuristic function performs path search on the dynamic non-uniform dual graph to quickly generate collision-free optimal waypoint paths. The algorithm uses a cost function To evaluate each candidate node Priority: in: Indicates starting from the origin To the current node The actual cumulative cost; This represents the output of the graph attention network GAT from the current node. To the finish line The estimated remaining cost.

2. The low-altitude path planning method based on a dynamic adaptive graph neural network according to claim 1, characterized in that, Step 1 includes: Step 1.1: Initialize the root node: Define the entire 3D flight airspace as a cubic space, which will serve as the root node of the octree. Let the side length of the root node be... The minimum resolution is ; Step 1.2, Obstacle Detection: For the current node, the intersection test between the obstacle geometry model and the current node's bounding box is performed to detect whether the current node's spatial range contains the obstacle boundary. Step 1.3, Recursive Partitioning Determination: If a node is completely occupied by obstacles, mark the node as "occupied" and terminate the partitioning; if a node is completely free space, mark the node as "free" and terminate the partitioning; if a node contains both obstacles and free space, and the node size is greater than... Then, the node is divided into 8 child nodes along the three coordinate axes of the three-dimensional space, and obstacle detection and recursive subdivision judgment are recursively performed on each child node. Step 1.4: Constructing a Non-Uniform Dual Graph: After partitioning, all leaf nodes marked as "empty" are used as nodes in the graph. If two leaf nodes are spatially adjacent, an undirected edge is created between them to form a non-uniform dual graph. ,in For a set of nodes, Let it be the set of edges.

3. The low-altitude path planning method based on a dynamic adaptive graph neural network according to claim 1, characterized in that, In step 2, Morton code uses a bit interleaving algorithm to convert the three-dimensional coordinates of the node into a unique index key value, thereby encoding the spatial node.

4. A low-altitude path planning method based on a dynamic adaptive graph neural network according to claim 1 or 3, characterized in that, The process of achieving rapid dynamic obstacle localization in step 2 is as follows: Obtain the three-dimensional coordinates of the obstacle from the sensor; The world coordinates of the obstacle are converted into the discretized index coordinates of the octree, and then converted into Morton codes using a bit interleaving algorithm. The Morton codes are used as keys to search in the hash table storing the leaf nodes of the octree, instantly locating the leaf node containing the obstacle coordinates.

5. The low-altitude path planning method based on a dynamic adaptive graph neural network according to claim 4, characterized in that, The process of updating the graph in step 2 is as follows: Remove the split original leaf nodes and all their associated edges from the non-uniform dual graph; Insert all newly generated child nodes marked as "free" into the non-uniform dual graph as new graph nodes; For each newly inserted node, find its spatial neighbor nodes in the octree. If the neighbor nodes are also in an "empty" state, then create an undirected edge.

6. The low-altitude path planning method based on a dynamic adaptive graph neural network according to claim 1, characterized in that, In step 3, the input node features of the graph attention network (GAT) include at least one of the following: the three-dimensional coordinates of the node center, the size of the node, the Euclidean distance from the node center to the nearest obstacle, the Euclidean distance from the node center to the target endpoint, and the octree depth of the node.

7. A low-altitude path planning system based on a dynamic adaptive graph neural network, used to implement the method as described in any one of claims 1-6, characterized in that, The system includes: UAV onboard terminal and ground control station; The UAV's onboard unit is equipped with: a sensor module, a navigation module, and an onboard computing unit; The sensor module is used to detect environmental information in real time, including lidar and / or binocular camera; The navigation module is used to provide the UAV's positioning and attitude information, including an IMU / GPS navigation module; The airborne computing unit internally deploys: The dynamic map manager is configured to perform the following operations: use an octree data structure to perform multi-resolution recursive partitioning of the three-dimensional flight airspace, construct and maintain a dynamic non-uniform dual graph; and introduce Morton codes to achieve millisecond-level local updates of the map. The GAT inference engine is configured to perform the following operations: load and run a pre-trained graph attention network GAT model, perform inference on the dynamic non-uniform dual graph, and output a heuristic cost estimate. The path planner is configured to perform the following operation: use the heuristic cost estimate output by the GAT inference engine as... The algorithm's heuristic function performs path search on the dynamic non-uniform dual graph to generate collision-free optimal waypoint paths; The ground control station is connected to the UAV's onboard terminal and is used to issue flight missions and preload high-precision static maps to the UAV's onboard terminal.

Images