A logistics distribution-oriented unmanned aerial vehicle path planning method, device and medium

Abstract

The embodiment of the application discloses a kind of unmanned aerial vehicle path planning methods, equipment and medium for logistics distribution.Belong to logistics distribution technical field, solve when facing city high-density logistics distribution, the overall distribution efficiency and safety of unmanned aerial vehicle cluster in existing logistics distribution is lower.The problem includes, the geographical information and environmental information corresponding to logistics distribution area are multidimensional fusion, and dynamic environment model is constructed;Based on the global state information corresponding to logistics distribution task and dynamic environment model, the unmanned aerial vehicle distribution scheme corresponding to logistics distribution task and cluster coordination rule are generated by global coordination agent;The distribution scheme and cluster coordination rule received are analyzed by local execution agent, reference execution trajectory is generated, and unmanned aerial vehicle distribution strategy is generated based on reference execution trajectory and local real-time situation map;Risk is superimposed to each path in unmanned aerial vehicle distribution strategy, and the optimal distribution path is determined by superposition result.

Description

Technical Field

[0001] This application relates to the field of logistics and distribution technology, and in particular to a method, device and medium for unmanned aerial vehicle (UAV) path planning for logistics and distribution. Background Technology

[0002] With the rapid growth of e-commerce and on-demand delivery demands, drone logistics delivery has been widely adopted in the logistics field due to its advantages such as flexibility, efficiency, and independence from ground conditions. Especially in complex urban environments, drone path planning technology is crucial for ensuring its safe and efficient operation.

[0003] However, traditional methods for large-scale, high-density real-time logistics delivery scenarios often employ either centralized planning or a fully distributed paradigm. Centralized planning calculates collision-free paths for the entire swarm, essentially treating the planning of multiple drones as a complex combinatorial optimization problem. Its computational complexity increases exponentially with the number of drones, resulting in poor scalability and a massive computational burden. Fully distributed planning, on the other hand, allows each drone to make decisions based solely on local perception and simple rules. While this approach offers good scalability, it lacks global coordination, making it difficult to achieve overall efficiency optimization at the swarm level and prone to congestion. Therefore, when facing high-density logistics delivery in cities, the overall delivery efficiency and safety of existing drone swarms in logistics delivery are relatively low. Summary of the Invention

[0004] This application provides a method, device, and medium for drone path planning in logistics delivery, which addresses the following technical problem: when facing high-density logistics delivery in cities, the overall delivery efficiency and safety of drone swarms in existing logistics delivery are relatively low.

[0005] The embodiments of this application adopt the following technical solutions: This application provides a drone path planning method for logistics delivery. It includes: fusing geographic and environmental information corresponding to the logistics delivery area in multiple dimensions to construct a dynamic environment model; generating drone allocation schemes and cluster coordination rules for the logistics delivery task through a global coordination agent based on the global state information and the dynamic environment model; generating a local real-time situation map within the local delivery area through a local execution agent; analyzing the received allocation schemes and cluster coordination rules through the local execution agent to generate a reference execution trajectory, and generating a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; and superimposing risks onto each path in the drone delivery strategy based on the dynamic environment model, and determining the optimal delivery path based on the superimposed risk value.

[0006] In one implementation of this application, the geographic and environmental information corresponding to the logistics and distribution area is fused in multiple dimensions to construct a dynamic environmental model. Specifically, this includes: acquiring multi-source spatiotemporal data corresponding to the logistics and distribution area; wherein the multi-source spatiotemporal data includes at least one of a digital elevation model, meteorological data, dynamic airspace information, and historical traffic flow data; constructing a static geographic raster layer corresponding to the logistics and distribution area based on the digital elevation model; wherein the static geographic raster layer includes at least one of topographic elevation, obstacle distribution, and no-fly zones; spatiotemporally fusing meteorological data and dynamic airspace information to generate a dynamic environmental factor layer updated at preset time intervals; wherein the dynamic environmental factors include at least one of wind field intensity, wind field direction, precipitation area, and temporary no-fly zones; extracting features from historical traffic flow data, predicting the conflict probability of each airspace unit within a preset future time period based on the extracted features, and generating a dynamic traffic situation layer; and multi-dimensionally fusing the static geographic raster layer, the dynamic environmental factor layer, and the dynamic traffic situation layer to generate a dynamic environmental model.

[0007] In one implementation of this application, feature extraction is performed on historical traffic flow data, and the conflict probability corresponding to each airspace unit within a preset future time period is predicted based on the extracted features. Specifically, this includes: dividing the logistics and distribution area into multiple first airspace units based on an equal-area grid; selecting second airspace units from the multiple first airspace units based on key nodes in the logistics and distribution area, wherein the key nodes include at least one of logistics warehouses, distribution hotspots, and airway intersections; adaptively encrypting each second airspace unit and dividing it hierarchically in the vertical direction according to a preset height interval to construct a three-dimensional airspace; extracting historical traffic flow density features and historical conflict probability features from the historical traffic flow data, and determining multiple third airspace units in the three-dimensional airspace units that have a historical traffic flow density that is continuously less than a first threshold and are spatially adjacent based on the extracted features; if the variance of the historical conflict probability corresponding to the multiple third airspace units is less than a second threshold, merging the multiple third airspace units to obtain merged airspace units; and extracting the spatiotemporal features of each unit through a spatiotemporal convolutional network based on the divided and merged airspace units to predict the conflict probability corresponding to each airspace unit within a preset future time period.

[0008] In one implementation of this application, based on the global state information and dynamic environment model corresponding to the logistics delivery task, a drone allocation scheme and cluster collaboration rules corresponding to the logistics delivery task are generated through a global coordination agent. Specifically, this includes: extracting task feature vectors and drone feature vectors corresponding to the logistics delivery from the global state information; wherein the task feature vector includes at least one of task urgency and task complexity, and the drone feature vector includes at least one of drone endurance and drone load status; matching the logistics task with the drone based on the correlation between the task feature vector and the drone feature vector; and generating a dynamic environment model based on the dynamic environment model. Multiple reference paths are generated, and the predicted information output by the dynamic traffic situation layer is superimposed on each reference path to construct a path prediction spatiotemporal conflict map. A weighted bipartite graph matching model is constructed with the goal of minimizing the global comprehensive cost through a global coordination agent, and the initial UAV allocation scheme is obtained by solving the model. The edge weights of the weighted bipartite graph matching model are related to the costs corresponding to each reference path. The spatiotemporal hotspots where paths intersect are identified in the path prediction spatiotemporal conflict map by the global coordination agent, and cluster coordination rules are generated based on these hotspots. The cluster coordination rules are at least related to the UAV flight speed and the UAV flight altitude.

[0009] In one implementation of this application, a local real-time situation map corresponding to the local delivery range is generated by a local execution agent. Specifically, this includes: constructing a dynamic spatiotemporal grid map with drones within the local delivery range as key points; associating detected drone dynamic target points with received neighboring drone status broadcast messages, and injecting the status information corresponding to the successfully associated target drones into the dynamic spatiotemporal grid map to generate a dynamic layer; identifying drone dynamic target points that cannot be associated as non-cooperative dynamic obstacles, and generating a non-cooperative dynamic obstacle layer based on their flight trajectories; and fusing the dynamic spatiotemporal grid map of the previous moment as a priori, and the dynamic layer and non-cooperative dynamic obstacle layer of the current moment as observation evidence to obtain a local real-time situation map.

[0010] In one implementation of this application, a local execution agent analyzes the received allocation scheme and cluster coordination rules to generate reference execution trajectories. Specifically, this includes: constructing a local trajectory optimization model with the objective functions of minimizing trajectory tracking error and minimizing energy consumption, and using the cluster coordination rules as constraints in the optimization process; within the planning period, inputting the local real-time situation map and the cluster coordination rules into the local trajectory optimization model, and solving multiple reference execution trajectories for future prediction periods in a rolling manner.

[0011] In one implementation of this application, within the planning period, the local real-time situation map and the cluster collaboration rules are jointly input into the local trajectory optimization model to solve multiple reference execution trajectories for the future prediction period. Specifically, this includes: within the planning period, generating multiple reference execution trajectories corresponding to the future prediction period based on the local real-time situation map and the cluster collaboration rules, forming a candidate trajectory set; quantifying each reference execution trajectory in the candidate trajectory set to determine the individual scores corresponding to each reference execution trajectory in different feature domains; wherein, the feature domain includes at least one of flight efficiency, energy consumption, safety risk, and cluster collaboration degree; dynamically adjusting the quantization weights according to the current mission stage and real-time status of the UAV, and determining the comprehensive score corresponding to each reference execution trajectory based on the adjusted weights; selecting several reference trajectories as reference execution trajectories for the future prediction period in descending order of comprehensive scores.

[0012] In one implementation of this application, risk prediction is performed on multiple paths in a drone delivery strategy, and the optimal delivery path is determined based on the risk results. Specifically, this includes treating each delivery path in the drone delivery strategy as a spatiotemporal path point sequence, and associating each path point with a corresponding spatiotemporal influence domain; wherein the spatiotemporal influence domain consists of the location of the path point and its predicted location within a preset future time period; based on a dynamic environment model, multiple dynamic risk fields are superimposed on the spatiotemporal path point sequence to obtain the inherent risk value corresponding to each path point; conflict hotspots that may intersect with other drone paths in each path are identified, and a risk propagation model is constructed based on the conflict hotspots to determine the interaction risk value of the path; the inherent risk value and interaction risk value corresponding to each path are weighted and fused to obtain a comprehensive risk score, thereby determining the optimal delivery path.

[0013] This application provides a drone path planning device for logistics delivery, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to: perform multi-dimensional fusion of geographic and environmental information corresponding to the logistics delivery area to construct a dynamic environment model; generate a drone allocation scheme and cluster coordination rules corresponding to the logistics delivery task through a global coordination agent based on the global state information and dynamic environment model corresponding to the logistics delivery task; generate a local real-time situation map corresponding to the local delivery area through a local execution agent; and analyze the received allocation scheme and cluster coordination rules through the local execution agent to generate a reference execution trajectory, and generate a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; and, based on the dynamic environment model, perform risk superposition on each path in the drone delivery strategy, and determine the optimal delivery path through the risk value obtained by superposition.

[0014] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: fuse geographical and environmental information corresponding to a logistics delivery area in multiple dimensions to construct a dynamic environment model; based on the global state information and dynamic environment model corresponding to the logistics delivery task, generate a drone allocation scheme and cluster collaboration rules corresponding to the logistics delivery task through a global coordination agent; generate a local real-time situation map corresponding to the local delivery area through a local execution agent; analyze the received allocation scheme and cluster collaboration rules through a local execution agent to generate a reference execution trajectory, and generate a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; and, based on the dynamic environment model, superimpose risks onto each path in the drone delivery strategy, and determine the optimal delivery path based on the superimposed risk values.

[0015] The above-mentioned technical solutions adopted in this application embodiment can achieve the following beneficial effects: By multi-dimensionally fusing geographic information and environmental information, this application embodiment constructs a dynamic environment model, providing a unified and accurate environmental perception basis for all subsequent decisions, and can proactively respond to dynamic factors such as weather changes and airspace control. The global coordination agent, based on the global state and dynamic model, can not only generate efficient drone allocation schemes but also formulate cluster collaboration rules, reducing potential conflicts between drones from the source and avoiding local optima and cluster inefficiencies that may result from distributed decision-making. Under the constraints of collaboration rules, the local execution agent analyzes and makes decisions in conjunction with real-time situation maps, generating adaptive strategies containing specific behavioral logic, thereby ensuring global collaboration while endowing drones with high robustness in dealing with emergencies. By quantifying and superimposing dynamic environmental risks onto each candidate path, the trade-off between safety and efficiency of different strategies can be accurately evaluated, thereby selecting the path with the lowest overall risk and improving the safety level of the entire delivery process. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings: Figure 1 A flowchart of a drone path planning method for logistics delivery is provided as an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a drone path planning device for logistics delivery, provided as an embodiment of this application.

[0017] Figure label: 200: Unmanned aerial vehicle (UAV) path planning equipment for logistics delivery; 201: Processor; 202: Memory. Detailed Implementation

[0018] This application provides a method, device, and medium for unmanned aerial vehicle (UAV) path planning for logistics delivery.

[0019] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0020] Figure 1 A flowchart illustrating a drone path planning method for logistics delivery, provided as an embodiment of this application. Figure 1 As shown, the drone path planning method for logistics delivery includes the following steps: S101. Multi-dimensionally integrate the geographic and environmental information corresponding to the logistics and distribution area to construct a dynamic environment model.

[0021] In one implementation of this application, multi-source spatiotemporal data corresponding to the logistics and distribution area is acquired; wherein, the multi-source spatiotemporal data includes at least one of digital elevation model, meteorological data, dynamic airspace information, and historical traffic flow data. Based on the digital elevation model, a static geographic raster layer corresponding to the logistics and distribution area is constructed; wherein, the static geographic raster layer includes at least one of topographic elevation, obstacle distribution, and no-fly zone. Meteorological data and dynamic airspace information are spatiotemporally fused to generate a dynamic environmental factor layer updated at preset time intervals; wherein, the dynamic environmental factors include at least one of wind field intensity, wind field direction, precipitation area, and temporary no-fly zone. Features are extracted from historical traffic flow data, and based on the extracted features, the conflict probability corresponding to each airspace unit in the future preset time period is predicted, and a dynamic traffic situation layer is generated. The static geographic raster layer, dynamic environmental factor layer, and dynamic traffic situation layer are multidimensionally fused to generate a dynamic environmental model.

[0022] Specifically, data is collected asynchronously or synchronously from multiple heterogeneous data sources through a data interface layer. Specifically, the digital elevation model (DEM) is obtained from geographic information systems (GIS) or satellite remote sensing data products, providing high-precision terrain elevation information. Meteorological data, such as gridded wind fields, precipitation, and visibility data covering the logistics and distribution area, is obtained in real time by subscribing to a meteorological service API. Dynamic airspace information, such as temporary no-fly zones and flight route adjustments, is obtained in digital format by accessing the airspace management department's notification system. Historical traffic flow data is extracted from the drone operation management platform's data warehouse, recording the drone traffic density, speed, and interval history of each airspace unit within a specific time period. All data is accompanied by a valid timestamp and spatial reference information upon acquisition. Using unified airspace grid division rules, the entire logistics and distribution area is discretized into regular three-dimensional grid units. Based on the acquired DEM, an elevation value is assigned to each ground projection grid unit, constructing a high-level terrain structure. Building upon this foundation, image recognition and 3D modeling technologies are used to automatically identify and vectorize prominent obstacles such as buildings and high-voltage power line towers from high-resolution orthophotos or point cloud data. Their 3D contours are then mapped onto a raster grid, identified as obstacle grids. Simultaneously, the geometric boundaries of no-fly zones are also rasterized and marked as no-fly grids. Finally, terrain, obstacle, and permanent no-fly zone information are integrated into a unified static geographic raster data layer.

[0023] Furthermore, meteorological data and dynamic spatial information are uniformly converted to the same spatial coordinate system as the static geographic raster layer. Next, a spatiotemporal interpolation algorithm is used to interpolate discrete meteorological observation point data into a grid field covering the entire region with spatiotemporal continuity. For dynamic spatial information, its geometric extent is associated with the effective time period to form spatiotemporal volume data. Finally, the processed meteorological grid field is overlaid with the dynamic spatial spatiotemporal volume to generate a dynamic environmental factor layer that is refreshed at preset time intervals. Each spatiotemporal unit in this layer contains environmental constraint attributes for the current and short-term future.

[0024] Furthermore, feature extraction is performed on historical traffic flow data, and the conflict probability corresponding to each airspace unit within a preset future time period is predicted based on the extracted features. Specifically, the logistics and distribution area is divided into multiple first airspace units based on an equal-area grid. Based on key nodes in the logistics and distribution area, second airspace units are selected from the multiple first airspace units, where key nodes include at least one of logistics warehouses, distribution hotspots, and airway intersections. Each second airspace unit is adaptively densified and hierarchically divided in the vertical direction at preset height intervals to construct a three-dimensional airspace. Historical traffic flow density features and historical conflict probability features are extracted from the historical traffic flow data, and based on the extracted features, multiple third airspace units with historical traffic flow densities consistently below a first threshold and spatially adjacent are identified in the three-dimensional airspace units. If the variance of the historical conflict probabilities corresponding to the multiple third airspace units is less than a second threshold, the multiple third airspace units are merged to obtain merged airspace units. Based on the divided and merged airspace units, spatiotemporal features of each unit are extracted using a spatiotemporal convolutional network to predict the conflict probability corresponding to each airspace unit within a preset future time period.

[0025] Specifically, the overall spatial scope of the logistics and distribution area is first determined, and an equal-area grid division method suitable for the geographical projection of the area is selected, such as a regular grid under the universal transverse Mercator projection. The size of the grid unit is set according to the required airspace management accuracy, such as 100 meters × 100 meters. Subsequently, the horizontal projection plane of the entire distribution area is divided into several grids of this size, each grid being a first airspace unit, and its corresponding two-dimensional spatial identifier is labeled, thereby establishing a baseline airspace index system covering the entire area. The geographical coordinates of key nodes are identified through a geographic information system or clustered from operational data. Logistics warehouses are known fixed points; distribution hotspots are determined through spatial density analysis of historical order data; and airway intersections are identified through network analysis of historical UAV flight paths to identify intersections with concentrated traffic. Then, with the coordinates of these key nodes as the center, an influence radius is set, and all first airspace units whose center points fall within this radius are selected and marked as second airspace units that require key monitoring. For each selected second airspace unit, adaptive densification is performed on its horizontal plane. Specifically, the original equal-area grid is further subdivided into higher-resolution subgrids, for example, a 100m × 100m grid is refined into a 25m × 25m subgrid. Simultaneously, different flight altitude layers are defined vertically according to preset height intervals. Ultimately, each cubic space uniquely defined by the horizontal subgrid and vertical altitude layers constitutes a refined three-dimensional airspace unit, thus forming a high-resolution three-dimensional airspace representation in key areas. Through spatiotemporal statistical analysis of historical traffic flow data, the average historical traffic density of each three-dimensional airspace unit during the same time period, as well as the frequency of historically close-interval events within that unit, are extracted. Then, a relatively low first threshold is set to filter out units with historical traffic density consistently below this threshold.

[0026] Furthermore, using spatial clustering algorithms, sets of units that are spatially adjacent in three dimensions are identified from these low-flow units, and these sets are designated as candidate sets of third-space units. The variance of the historical conflict probability of each spatially adjacent third-space unit in the candidate set is calculated. A small second threshold is set as the upper limit of the variance to determine whether the risk levels of these units are uniform. If the variance is less than the second threshold, it indicates that these adjacent units not only have sparse flow but also similar conflict risk levels, meeting the conditions for merging. Subsequently, these third-space units that meet the conditions are geometrically merged into a larger, irregular three-dimensional spatial unit, i.e., merged spatial unit, and regarded as a new spatial management unit for subsequent prediction. The final spatial unit system obtained after partitioning and merging is used as the basic spatiotemporal graph node. A spatiotemporal graph is constructed using the historical flow data, conflict data, and spatial adjacency relationships of each unit over multiple consecutive time slices. This graph is input into a spatiotemporal convolutional network, which captures the spatial dependencies between spatial units through spatial dimension graph convolution and captures their temporal evolution patterns through temporal dimension one-dimensional convolution. The network ultimately learns a mapping function from historical spatiotemporal features to the probability of conflict in a future preset time period, thereby achieving accurate prediction of the future conflict probability for each airspace unit. Finally, these predicted conflict probability values ​​are filled into the corresponding airspace units, forming a gridded dynamic traffic situation layer covering future time periods.

[0027] Each airspace unit is assigned a unique identifier, with time as the fourth dimension index. Based on this spatiotemporal index, the fusion engine correlates and overlays basic geographic information from the static geographic raster layer, meteorological and airspace constraints at the corresponding time in the dynamic environmental factor layer, and predicted conflict probabilities for the corresponding time period in the dynamic traffic situation layer. The fusion result forms a four-dimensional dynamic environmental model. This model, at any given spatiotemporal coordinates, can query whether the terrain is passable, whether there are obstacles or no-fly zones, the current wind field and precipitation conditions, and the expected traffic conflict risk at that spatiotemporal point, thus providing a comprehensive, quantitative, and forward-looking environmental cognitive foundation for route planning.

[0028] S102. Based on the global state information and dynamic environment model corresponding to the logistics and delivery task, a drone allocation scheme and cluster collaboration rules corresponding to the logistics and delivery task are generated through a global coordination agent.

[0029] In one implementation of this application, task feature vectors and drone feature vectors corresponding to logistics delivery are extracted from the global state information. The task feature vectors include at least one of task urgency and task complexity, and the drone features include at least one of drone endurance and drone load status. Based on the correlation between the task feature vectors and drone feature vectors, logistics tasks are matched with drones. Multiple reference paths are generated based on a dynamic environment model, and the prediction information output by the dynamic traffic situation layer is superimposed on each reference path to construct a path prediction spatiotemporal conflict map. A weighted bipartite graph matching model is constructed using a global coordination agent with the objective of minimizing the global comprehensive cost, and an initial drone allocation scheme is obtained by solving the model. The edge weights of the weighted bipartite graph matching model are related to the costs corresponding to each reference path. The global coordination agent identifies spatiotemporal hotspots where paths intersect in the path prediction spatiotemporal conflict map and generates cluster coordination rules based on these hotspots. The cluster coordination rules are at least related to the drone's flight speed and flight altitude.

[0030] Specifically, the raw state information is quantitatively extracted. The process is as follows: for logistics delivery tasks, task urgency is calculated by taking the time difference between the current time and the promised latest delivery time and normalizing it in conjunction with business priority; task complexity is calculated by comprehensively weighting the spatial accessibility of the delivery destination, the estimated flight distance, and the density of known static obstacles in the route area. For drones, their endurance feature vector is characterized by the ratio of the current remaining battery power to the average power consumption per unit distance; load status is described by the ratio of the current payload to the maximum payload capacity, and the accumulated value of the payload assigned to the task but not yet executed. Finally, each task and drone is transformed into a numerical feature vector. Next, a task-drone fit matrix is ​​constructed, where each element represents the degree of matching between a task and a drone. This fit is calculated using a predefined fit function, which comprehensively considers the matching relationship between task urgency and drone readiness status, task complexity and drone endurance and remaining payload capacity. For example, high-urgency tasks are preferentially matched with drones with high readiness status; high-weight tasks must be matched with drones with sufficient payload capacity. This step outputs a set of candidate task-drone pairings.

[0031] Furthermore, for each candidate task—UAV pairing—obtained above, several feasible reference paths under static environmental constraints are generated using the A* or fast random tree algorithm based on the dynamic environment model. Subsequently, the predicted airspace conflict probabilities for different future time periods provided by the dynamic traffic situation layer are mapped to each path point of each reference path, thus assigning each path a conflict risk curve that varies over time. Based on this, all reference paths for all candidate pairings are overlaid in a unified spatiotemporal coordinate system to analyze the intersection of paths at specific spatiotemporal nodes, represented by a graph structure: nodes represent path segments, edges represent potential spatiotemporal intersections between path segments, and the edge weights characterize the probability or severity of conflict, thereby constructing a path prediction spatiotemporal conflict map.

[0032] Furthermore, a weighted bipartite graph is constructed, where one set of vertices represents all tasks to be assigned, and the other set represents all available UAVs. The resulting candidate pairings are used as edges connecting task vertices and UAV vertices. The weight of each edge is a comprehensive quantified value, composed of multiple cost factors of the reference path selected for the pairing, including but not limited to: the estimated flight time cost of the path, the total energy cost, and the overall risk cost of the path obtained based on the conflict graph. The problem is formalized into a weighted bipartite graph matching problem with the objective of minimizing the global total cost by a global coordination agent, and solved to ultimately output an initial assignment scheme with one-to-one correspondence between tasks and UAVs. The global coordination agent performs cluster analysis on the constructed path prediction spatiotemporal conflict graph, identifying regions with high edge weights and concentrated connectivity; these regions are the spatiotemporal hotspots where multi-UAV interactions may occur in the future. For each identified hotspot region, the agent generates corresponding preventative coordination rules. For example, for an airspace node where multiple paths intersect within a specific time period, the rules may stipulate that drones flying in different directions must adhere to different designated flight altitude layers above this node to achieve spatial isolation.

[0033] S103. Generate a local real-time situation map corresponding to the local delivery range through local execution intelligent agents.

[0034] In one implementation of this application, a dynamic spatiotemporal grid map is constructed using drones within a local delivery range as key points. Detected dynamic target points of drones are associated with received status broadcast messages from neighboring drones, and the status information corresponding to successfully associated target drones is injected into the dynamic spatiotemporal grid map to generate a dynamic layer. Dynamic target points of drones that cannot be associated are identified as non-cooperative dynamic obstacles, and a non-cooperative dynamic obstacle layer is generated based on their flight trajectories. Using the dynamic spatiotemporal grid map from the previous moment as a priori, and the dynamic layer and non-cooperative dynamic obstacle layer from the current moment as observational evidence, they are fused to obtain a local real-time situation map.

[0035] Specifically, using the precise navigation and positioning information of the current UAV as a spatiotemporal reference, a local three-dimensional airspace is established, centered on the UAV's current location and covering a specific radius horizontal range and a specific altitude. This airspace is discretized into a dynamic spatiotemporal grid map composed of regular cubic units. Each grid unit not only records its spatial coordinates but also contains multiple state attributes, such as airspace occupancy probability, average elevation, and semantic labels (e.g., idle, static obstacle, cooperative target, non-cooperative target, etc.). This map is initialized with the estimated state of the previous moment at the initial moment and dynamically updated in each processing cycle as the UAV moves, with its coordinate system origin always remaining relatively fixed relative to the UAV. Next, dynamic target points detected by airborne sensors are clustered and tracked to form dynamic targets with continuous trajectories. Simultaneously, status messages periodically broadcast by neighboring UAVs are received via inter-UAV communication links. These messages contain their high-precision position, speed, identification, and flight intention. Subsequently, a nearest neighbor association algorithm or a joint probabilistic data association algorithm is used to calculate the spatiotemporal and motion similarity between each sensor target point and each communication message. Points with similarity exceeding a preset threshold are associated and matched with messages. For successfully associated cooperative targets, their identity ID, precise status, and flight intention, among other high-level semantic information, are injected into the corresponding grid cell occupied by the target in the dynamic spatiotemporal raster map, thereby generating a dynamic layer of cooperative targets rich in semantic information. For all UAV dynamic target points that fail to be successfully associated with any cooperative target status message, they are marked as non-cooperative dynamic obstacles, such as UAVs without cooperative communication equipment or birds. These points are then further tracked and filtered to estimate their trajectory, velocity, and heading. The tracked target position, velocity vector, and predicted short-term trajectory are mapped to the corresponding grid cell in the dynamic spatiotemporal raster map, and their attributes are marked as non-cooperative dynamic obstacles. Simultaneously, their motion uncertainty is recorded, thereby generating a non-cooperative dynamic obstacle layer.

[0036] Furthermore, a probabilistic grid update method based on Bayesian estimation is employed, using the state estimate of each grid cell in the previous dynamic spatiotemporal grid map as the prior probability. The currently generated dynamic layer and non-cooperative dynamic obstacle layer are considered as new observational evidence, and these observations are transformed into conditional probabilities. Subsequently, using the Bayesian update formula, the prior probabilities are fused with the likelihood probabilities of the current observations to calculate the posterior probability distribution of each grid cell at the current time, thereby updating the spatial occupancy status, target classification, and motion attributes of the entire map. This process effectively reduces the uncertainty of a single sensor and integrates temporal information, ultimately outputting a stable, reliable, and time-consistent local real-time situational map.

[0037] S104. The received allocation scheme and cluster coordination rules are analyzed by the local execution agent to generate a reference execution trajectory. Based on the reference execution trajectory and the local real-time situation map, a drone delivery strategy is generated.

[0038] In one implementation of this application, a local trajectory optimization model is constructed with the objective functions of minimizing trajectory tracking error and minimizing energy consumption, and cluster collaboration rules are used as constraints in the optimization process. Within the planning period, the local real-time situation map and the cluster collaboration rules are jointly input into the local trajectory optimization model to continuously solve for multiple reference execution trajectories within the future prediction period.

[0039] Specifically, the local trajectory optimization model in this embodiment is constructed using a model predictive control framework. The model's state variables include the UAV's position, velocity, attitude, and remaining battery power, while the control variables are the motor thrust or attitude angle commands. The objective function is formalized as a cost function that needs to be minimized within a finite time domain. This function mainly consists of a weighted sum of two parts: the first part is the trajectory tracking error term, used to penalize the deviation between the UAV's actual state and the desired reference state; the second part is the energy consumption term, calculated by integrating the motor thrust or power to estimate energy consumption. Simultaneously, swarm coordination rules, such as time windows for specific airspace nodes, specified flight altitude layers, or speed ranges, are transformed into hard or soft constraints in the optimization problem. Specifically, the time window can be transformed into time equality or inequality constraints for reaching specific path points, while altitude layers and speed ranges are transformed into boundary constraints on state variables, thereby embedding the coordination requirements into the optimization problem.

[0040] Furthermore, within the planning period, based on the local real-time situation map and cluster collaboration rules, multiple reference execution trajectories are generated for the future prediction period, forming a candidate trajectory set. Each reference execution trajectory in the candidate trajectory set is quantified to determine its individual score across different feature domains; these feature domains include at least one of flight efficiency, energy consumption, safety risk, and cluster collaboration. The quantization weights are dynamically adjusted based on the UAV's current mission stage and real-time status, and the comprehensive score for each reference execution trajectory is determined based on the adjusted weights. Several reference trajectories are selected as reference execution trajectories for the future prediction period, ranked from highest to lowest comprehensive score.

[0041] Specifically, at the start of each planning cycle, the aforementioned local trajectory optimization model is used to rapidly generate multiple reference execution trajectories for the future prediction period through parallel computation or by using different initial solution seeds. These trajectories, while satisfying dynamic constraints and basic obstacle avoidance, achieve diversity by differentiating the weight settings of their optimization objective functions or introducing different behavioral assumptions. For example, one trajectory might prioritize flight efficiency, while another focuses on smoothness to reduce energy consumption. All these trajectories must adhere to swarm coordination rules as a hard constraint and use dynamic obstacle prediction information provided by the local real-time situation map as an environmental constraint, thereby ensuring that the candidate trajectory set simultaneously possesses feasibility, diversity, and rule compliance. For flight efficiency, the total flight time of the trajectory or the deviation from the ideal shortest time is calculated and normalized to an efficiency score. For energy consumption, the estimated total energy consumption of the entire trajectory is calculated based on the UAV's power consumption model and converted into an energy consumption score. For safety risk, the minimum distance between each point on the trajectory and static and dynamic obstacles is calculated, and sensor uncertainties are considered, and the risk score is obtained through integration. For swarm coordination, the strictness of the trajectory's adherence to swarm coordination rules is evaluated, generating a coordination score. Each individual score is normalized to a preset range to facilitate subsequent comprehensive comparison.

[0042] Furthermore, a basic weight vector is established, corresponding to the initial importance of features such as flight efficiency, energy consumption, safety, and coordination. This weight vector is not fixed but dynamically adjusted by a weight adjustment strategy module based on the real-time context of the UAV. For example, when the UAV is in the cruise phase and has sufficient battery power, the flight efficiency weight is increased; when the battery power is below a threshold, the energy consumption weight is significantly increased; and when passing through a known high-conflict-risk intersection zone, the weights of safety and coordination are increased. The adjusted weight vector is weighted and summed with the individual score vectors of each trajectory to calculate the final comprehensive score of each reference execution trajectory in the current context. After calculating the comprehensive scores of all candidate trajectories, they are sorted in descending order. The trajectory ranked first is selected as the main execution trajectory for the current planning cycle and immediately sent to the flight control system for tracking. At the same time, the next one or two high-scoring trajectories are cached as immediate backup trajectories. This set of selected trajectories serves as the final reference execution trajectory scheme determined for this cycle, used for actual execution and emergency backup.

[0043] S105. Based on the dynamic environment model, risk is superimposed on each path in the drone delivery strategy, and the optimal delivery path is determined by the risk value obtained from the superposition.

[0044] In one implementation of this application, each delivery path in the drone delivery strategy is treated as a spatiotemporal path point sequence, and a corresponding spatiotemporal influence domain is associated with each path point; wherein, the spatiotemporal influence domain consists of the location of the path point and its predicted location within a preset future time period. Based on a dynamic environment model, multiple dynamic risk fields are superimposed on the spatiotemporal path point sequence to obtain the inherent risk value corresponding to each path point. Conflict hotspots that potentially intersect with other drone paths are identified, and a risk propagation model is constructed based on these conflict hotspots to determine the interaction risk value of the path. The inherent risk value and interaction risk value corresponding to each path are weighted and fused to obtain a comprehensive risk score, thereby determining the optimal delivery path.

[0045] Specifically, a complete delivery route is discretized into a spatiotemporal sequence of points arranged chronologically, where each point contains three-dimensional spatial coordinates and a precise timestamp. For each path point in the sequence, its spatiotemporal influence domain is defined as a four-dimensional ellipsoid centered at that point. This ellipsoid spatially considers the physical size and positioning error of the UAV, and temporally extends forward and backward by a small time interval to cover the actual location range that the UAV may occupy due to control errors or wind speed disturbances. Therefore, the spatiotemporal influence domain of each path point essentially represents the spatiotemporal volume that the UAV may occupy at that moment. For the spatiotemporal influence domain of each path point, multiple dynamic risk fields corresponding to the dynamic environment model are queried. First, the maximum wind shear intensity and precipitation level risk values ​​that may occur within the influence domain are extracted from the micro-meteorological risk field. Second, the probability of non-cooperative targets intruding into the influence domain is obtained from the dynamic obstacle risk field. Finally, the probability of navigation signal loss in the area is assessed from the signal interference risk field. By using weighted summation to fuse the risk values ​​from different sources, the inherent environmentally relevant comprehensive risk value within the spatiotemporal influence domain of the path point is calculated and used as the inherent risk value of that point.

[0046] By comparing and analyzing the planned paths of all drones, spatiotemporal nodes with overlapping spatiotemporal influence domains between paths are identified and marked as conflict hotspots. Based on these hotspots, a directed graph-based risk propagation model is constructed, where nodes represent conflict hotspots and their associated path segments, and edges represent risk propagation relationships. Simulations using this model simulate how initial local conflicts trigger chain reactions under random perturbations, thereby quantifying the expected costs of potential delays and conflicts caused by swarm interactions. This expected cost is quantified as the interaction risk value for the entire path. For each candidate path, the inherent risk value of all path points is integrated along the path to obtain the total inherent risk of the path. This value is then weighted and fused with the interaction risk value obtained through the risk propagation model. Finally, a comprehensive risk score is calculated for each path, and the path with the highest comprehensive score is determined as the optimal delivery path.

[0047] Figure 2 This is a structural schematic diagram of a drone path planning device for logistics delivery, provided as an embodiment of this application. Figure 2As shown, a drone path planning device 200 for logistics delivery includes: at least one processor 201; and a memory 202 communicatively connected to the at least one processor 201. The memory 202 stores instructions executable by the at least one processor 201, which, when executed, enable the at least one processor 201 to: fuse geographical and environmental information corresponding to the logistics delivery area in multiple dimensions to construct a dynamic environment model; generate a drone allocation scheme and cluster coordination rules corresponding to the logistics delivery task through a global coordination agent based on the global state information and dynamic environment model; generate a local real-time situation map within the local delivery area through a local execution agent; analyze the received allocation scheme and cluster coordination rules through the local execution agent to generate a reference execution trajectory, and generate a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; and, based on the dynamic environment model, superimpose risks onto each path in the drone delivery strategy, and determine the optimal delivery path based on the superimposed risk value.

[0048] This application provides a non-volatile computer storage medium storing computer-executable instructions. These instructions are configured to: fuse geographical and environmental information corresponding to a logistics delivery area in multiple dimensions to construct a dynamic environment model; based on the global state information and dynamic environment model corresponding to the logistics delivery task, generate a drone allocation scheme and cluster collaboration rules corresponding to the logistics delivery task through a global coordination agent; generate a local real-time situation map corresponding to the local delivery area through a local execution agent; analyze the received allocation scheme and cluster collaboration rules through a local execution agent to generate a reference execution trajectory, and generate a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; and, based on the dynamic environment model, superimpose risks onto each path in the drone delivery strategy, and determine the optimal delivery path based on the superimposed risk values.

[0049] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and non-volatile computer storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0050] The above descriptions are merely embodiments of this application and are not intended to limit the scope of this application. For those skilled in the art, various modifications and variations can be made to the embodiments of this application. These modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the spirit and scope of the technical solutions in the embodiments of this application.

Claims

1. A method for unmanned aerial vehicle (UAV) path planning for logistics delivery, characterized in that, The method includes: The process of constructing a dynamic environment model by multidimensionally fusing geographic and environmental information corresponding to a logistics distribution area includes: acquiring multi-source spatiotemporal data corresponding to the logistics distribution area; wherein the multi-source spatiotemporal data includes at least one of digital elevation model, meteorological data, dynamic airspace information, and historical traffic flow data; constructing a static geographic raster layer corresponding to the logistics distribution area based on the digital elevation model; wherein the static geographic raster layer includes at least one of terrain elevation, obstacle distribution, and no-fly zones; spatiotemporally fusing the meteorological data and the dynamic airspace information to generate a dynamic environmental factor layer updated at preset time intervals; wherein the dynamic environmental factors include at least one of wind field intensity, wind field direction, precipitation area, and temporary no-fly zones; extracting features from the historical traffic flow data, predicting the conflict probability corresponding to each airspace unit within a preset time period based on the extracted features, and generating a dynamic traffic situation layer; and multidimensionally fusing the static geographic raster layer, the dynamic environmental factor layer, and the dynamic traffic situation layer to generate the dynamic environment model. Based on the global state information corresponding to the logistics delivery task and the dynamic environment model, a global coordination agent generates a drone allocation scheme and cluster collaboration rules for the logistics delivery task. Specifically, this includes: extracting task feature vectors and drone feature vectors corresponding to the logistics delivery from the global state information; wherein the task feature vector includes at least one of task urgency and task complexity, and the drone feature vector includes at least one of drone endurance and drone load status; matching logistics tasks with drones based on the correlation between the task feature vectors and the drone feature vectors; and generating multiple reference paths based on the dynamic environment model. The system constructs a path prediction spatiotemporal conflict map by overlaying the predicted information output by the dynamic traffic situation layer onto each of the reference paths. A weighted bipartite graph matching model is constructed using a global coordination agent to minimize the overall global cost, and an initial UAV allocation scheme is obtained by solving the model. The edge weights of the weighted bipartite graph matching model are related to the costs corresponding to each of the reference paths. The global coordination agent identifies spatiotemporal hotspots where paths intersect in the path prediction spatiotemporal conflict map and generates cluster coordination rules based on these hotspots. These cluster coordination rules are at least related to the UAV's flight speed and altitude. A local real-time situation map is generated within the local delivery area by locally executing intelligent agents. Furthermore, the local execution agent analyzes the received allocation scheme and cluster coordination rules to generate a reference execution trajectory, and generates a drone delivery strategy based on the reference execution trajectory and the local real-time situation map; Based on the dynamic environment model, risks are superimposed on each path in the drone delivery strategy, and the optimal delivery path is determined by the risk value obtained from the superposition.

2. The method for unmanned aerial vehicle (UAV) path planning for logistics delivery according to claim 1, characterized in that, The step of extracting features from the historical traffic flow data and predicting the conflict probability of each airspace unit within a preset future time period based on the extracted features specifically includes: The logistics and distribution area is divided based on an equal-area grid to obtain multiple first spatial units; Based on the key nodes in the logistics and distribution area, a second airspace unit is selected from multiple first airspace units, wherein the key nodes include at least one of logistics warehouses, distribution hotspots, and air route intersections; Each of the second spatial units is adaptively encrypted and divided into layers in the vertical direction according to a preset height interval to construct a three-dimensional spatial domain. Historical traffic flow density features and historical conflict probability features are extracted from the historical traffic flow data. Based on the extracted features, multiple third spatial domain units in the three-dimensional spatial domain that have historical traffic flow density that is continuously less than a first threshold and are spatially adjacent are identified. If the variance of the historical conflict probabilities corresponding to the multiple third spatial domain units is less than the second threshold, then the multiple third spatial domain units are merged to obtain a merged spatial domain unit. Based on the divided and merged spatial domain units, the spatiotemporal features of each unit are extracted through a spatiotemporal convolutional network to predict the conflict probability of each spatial domain unit within a preset future time period.

3. The method for unmanned aerial vehicle (UAV) path planning for logistics delivery according to claim 1, characterized in that, The process of generating a local real-time situation map within a local delivery area through a local execution agent specifically includes: A dynamic spatiotemporal grid map is constructed using drones within a local delivery range as key points. The detected dynamic target points of drones are associated with the received status broadcast messages of neighboring drones, and the status information corresponding to the successfully associated target drones is injected into the dynamic spatiotemporal grid map to generate a dynamic layer. Based on the uncorrelated UAV dynamic target points, they are identified as non-cooperative dynamic obstacles, and a non-cooperative dynamic obstacle layer is generated based on their flight trajectories. Using the dynamic spatiotemporal raster map of the previous moment as a priori, and the dynamic layer and the non-cooperative dynamic obstacle layer of the current moment as observation evidence, the local real-time situation map is obtained by fusion.

4. The drone path planning method for logistics delivery according to claim 1, characterized in that, The step of analyzing the received allocation scheme and cluster coordination rules through the local execution agent to generate a reference execution trajectory specifically includes: A local trajectory optimization model is constructed with the objective function of minimizing trajectory tracking error or minimizing energy consumption, and the cluster collaboration rule is used as a constraint condition for the optimization process. During the planning period, the local real-time situation map and the cluster collaboration rules are input into the local trajectory optimization model to solve multiple reference execution trajectories for future prediction periods.

5. The drone path planning method for logistics delivery according to claim 4, characterized in that, Within the planning period, the local real-time situation map and the cluster collaboration rules are jointly input into the local trajectory optimization model to continuously solve for multiple reference execution trajectories within the future prediction period, specifically including: Within the planning period, based on the local real-time situation map and the cluster collaboration rules, multiple reference execution trajectories corresponding to the future prediction period are generated to form a candidate trajectory set; The reference execution trajectories in the candidate trajectory set are quantified to determine the individual scores of each reference execution trajectory in different feature domains; wherein, the feature domains include at least one of flight efficiency, energy consumption, safety risk and swarm coordination. Based on the current mission stage and real-time status of the UAV, the quantization weights are dynamically adjusted, and the comprehensive scores corresponding to each of the reference execution trajectories are determined based on the adjusted weights. Several reference trajectories are selected in descending order of comprehensive scores as reference execution trajectories for the future prediction period.

6. The method for unmanned aerial vehicle (UAV) path planning for logistics delivery according to claim 1, characterized in that, The process of applying risk aggregation to each path in the drone delivery strategy based on the dynamic environment model, and determining the optimal delivery path using the aggregated risk values, specifically includes: Each delivery path in the drone delivery strategy is taken as a spatiotemporal path point sequence, and a corresponding spatiotemporal influence domain is associated with each path point; wherein, the spatiotemporal influence domain is composed of the location of the path point and the predicted location within a preset future time period; Based on the dynamic environment model, multiple dynamic risk fields are superimposed on the spatiotemporal path point sequence to obtain the inherent risk value corresponding to each path point. The potential conflict hotspots that intersect with other UAV paths in each path are identified, and a risk propagation model is constructed based on the conflict hotspots to determine the interaction risk value of the path based on the risk propagation model. The inherent risk value and the interactive risk value corresponding to each path are weighted and fused to obtain a comprehensive risk score, thereby determining the optimal delivery path.

7. A drone path planning device for logistics delivery, characterized in that, The device includes a memory for storing computer program instructions and a processor for executing the program instructions, wherein when the computer program instructions are executed by the processor, the device is triggered to perform the method described in any one of claims 1-6.

8. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are capable of performing the method described in any one of claims 1-6.

Images