Urban logistics unmanned aerial vehicle path planning method

Abstract

The application discloses a kind of urban logistics unmanned aircraft path planning methods, belong to unmanned aircraft path planning technical field, S1, construct city global air-ground integration risk base static three-dimensional grid model;S2, construct based on risk level matching space-time hierarchical dynamic twin model;S3, construct the coupling path planning cost function of double risk hard constraint preposition;S4, global benchmark path planning;S5, global-local linkage local dynamic obstacle avoidance and hierarchical re-planning;S6, risk check and closed-loop optimization of airworthiness traceability;The application is based on the space-time hierarchical dynamic twin modeling mechanism of risk level matching, according to the ground third-party risk level of grid, modeling accuracy and data update frequency are differentiated matching, high-precision modeling is adopted in high-risk area to ensure safety redundancy, meter-level lightweight modeling is adopted in low-risk area to reduce calculation amount, while effectively reducing the overall calculation amount of global modeling, the fine management and control of high-risk area are realized.

Description

Technical Field

[0001] This invention belongs to the field of drone path planning technology, specifically, it relates to a method for urban logistics drone path planning. Background Technology

[0002] With the rapid and standardized development of the low-altitude economy industry and the continued surge in demand for urban instant logistics and same-city last-mile delivery, urban logistics drones have become an important development direction for urban smart logistics systems due to their core advantages of high efficiency, flexibility, and immunity to ground traffic congestion. Path planning technology is the core foundation for the safe, compliant, and efficient operation of urban logistics drones. It directly determines the safety level, delivery efficiency, and operating costs of drone flights, and is also a core element for adapting to urban low-altitude flight control policies and meeting civil aviation airworthiness certification requirements.

[0003] Currently, the path planning technology for urban logistics drones still has certain shortcomings in practical applications. Existing path planning schemes mainly fall into two categories: one is based on static preset routes and fixed no-fly zones. This type of scheme can only adapt to static and fixed urban airspace environments. It lacks the ability to predict and respond to dynamic obstacles such as birds, other aircraft, and temporary construction buildings, as well as dynamic scenarios such as sudden weather, temporary airspace control, and GPS signal obstruction in urban canyons. It can only achieve passive avoidance through post-event replanning, which can easily lead to flight safety accidents. The other type is based on full-domain refined 3D modeling. Although this type of scheme can improve the accuracy of environmental perception, full-domain refined 3D modeling of the city will lead to an explosion of the state space and a surge in computational load, which cannot meet the timeliness requirements of real-time planning. On the other hand, lightweight modeling will lead to insufficient safety redundancy, failing to meet the collision probability ≤10 specified in the ISO18491 airworthiness standard. -7 The stringent requirement of flight hours has created a core problem that the industry has long struggled to solve: the balance between modeling accuracy and real-time planning.

[0004] Existing urban logistics drone path planning technologies generally prioritize shortest flight distance, lowest energy consumption, and fastest delivery time as core optimization objectives. They merely incorporate legally designated no-fly zones as simple constraints in the planning process, rarely considering safety risks to third parties on the ground as a prerequisite hard constraint and core optimization objective. In densely populated urban areas, existing solutions are prone to routes flying over sensitive areas such as schools, hospitals, and residential areas. This not only fails to meet the compliance requirements for low-altitude urban flights but also raises public concerns about drone flight safety, becoming a core obstacle to the large-scale deployment of urban logistics drones. Summary of the Invention

[0005] To address the aforementioned problems and technical deficiencies, this invention employs the following technical solution: a method for urban logistics drone path planning, comprising the following steps:

[0006] S1. Construct a static three-dimensional grid model of the city’s air-ground integrated risk base: Divide the target city’s low-altitude logistics airspace into multiple three-dimensional grids, collect static airspace constraint elements and ground third-party risk elements simultaneously, quantify the ground third-party risk value corresponding to each grid and perform hierarchical control, map the results of airspace static constraints and ground third-party risk classification to the three-dimensional grid, generate the flyability label, risk level and constraint conditions of each grid, and form a city-wide static risk base model.

[0007] S2. Construct a spatiotemporal hierarchical dynamic twin model based on risk level matching: Based on the risk level of each grid in the static risk base model, dynamically match the modeling accuracy and data update frequency corresponding to each grid; access multi-source dynamic environmental data, generate dynamic environmental prediction results for a preset duration through a spatiotemporal fusion prediction model, update the prediction results to the static risk base model in real time, and generate a spatiotemporally continuous hierarchical dynamic digital twin model.

[0008] S3. Construct a coupled path planning cost function with dual-risk hard constraints in advance: Set three types of pre-filtering conditions: compliance hard constraints, airworthiness hard constraints, and physical hard constraints. Paths that do not meet any of the pre-filtering conditions are directly eliminated. For paths that meet the pre-filtering conditions, construct a multi-objective optimization cost function that integrates path length, energy consumption, cumulative risk of dynamic collision in the air, and cumulative risk of third parties on the ground.

[0009] S4. Global baseline route planning: Based on the hierarchical dynamic digital twin model, hard no-fly zones are eliminated to compress the route planning space. A multi-objective optimization algorithm is adopted, with the coupled route planning cost function as the optimization objective, to generate a global baseline route that meets all pre-filtering conditions and complete dual risk pre-verification.

[0010] S5. Global-Local Linkage for Local Dynamic Obstacle Avoidance and Hierarchical Replanning: Establish a global-local dual-layer linkage constraint mechanism, with local path planning inheriting the three types of pre-filtering conditions of the global baseline route throughout the process; perform local path fine-tuning and dynamic obstacle avoidance based on real-time UAV perception data and dynamic environment prediction results; set a three-level replanning trigger mechanism based on risk thresholds, dynamically triggering path replanning of the corresponding level according to changes in environmental risk.

[0011] S6. Airworthiness traceability risk verification and closed-loop optimization: Full-process risk recording and traceability verification of all executed paths, generating flight risk reports; actual flight data is fed back to the spatiotemporal fusion prediction model and the coupled path planning cost function optimization module to achieve continuous self-optimization of the model.

[0012] Preferably, in step S1, the low-altitude logistics airspace of the target city is divided into a three-dimensional grid with a horizontal granularity of 100m×100m and a vertical granularity of 5m. Each grid corresponds to a unique spatiotemporal coordinate and attribute label. A quantitative model coupled with entropy weighting and analytic hierarchy process is used to standardize and quantify the ground third-party risk of each horizontal grid, generating a risk value R_ground of 0-1. Based on the risk value, three levels of rigid control rules are set.

[0013] Hard no-fly grid: R_ground≥0.9, set as a prerequisite hard constraint that no one can fly over;

[0014] Restricted Flying Grids: 0.3≤R_ground<0.9, flying is only allowed within a preset height and specified time slot, and the cumulative flying time of a single path in this type of grid must not exceed a preset threshold;

[0015] Flyable grid: R_ground < 0.3, which is the priority for airspace planning.

[0016] Furthermore, in step S1, the static constraints in the air include DEM (Digital Elevation Model), 3D building models, legally designated no-fly or restricted-fly zones, static obstacles, and high-risk areas of GPS signal obstruction in urban canyons; the third-party risk factors on the ground include land use type, static population density, legal control level of sensitive locations, distribution of emergency landing sites, and building density.

[0017] Preferably, in step S2, the matching rules between risk level and modeling accuracy and update frequency are as follows:

[0018] High-risk areas: the 500m radius around the no-fly zone and the core area of ​​the restricted-fly zone, using centimeter-level high-precision modeling with a data update frequency of ≤1Hz;

[0019] Medium-risk area: The buffer zone around the restricted grid is limited, and decimeter-level modeling is used with a data update frequency of ≤5Hz;

[0020] Low-risk area: Flyable grid across the entire area, using meter-level lightweight modeling, with a data update frequency of ≤10Hz.

[0021] Furthermore, in step S2, the accessed multi-source dynamic environmental data includes real-time meteorological monitoring and forecast data, temporary airspace control notices, real-time flight plans and location data of same-city logistics drone clusters, real-time ground pedestrian flow heat data, updated data of temporary urban buildings and construction areas, real-time monitoring and historical spatiotemporal pattern data of bird activities, and real-time drone positioning and sensor data; the spatiotemporal fusion prediction model is a spatiotemporal fusion Transformer model, which outputs dynamic environmental prediction results for the next 5-15 minutes. The dynamic environmental prediction results include prediction of the spatiotemporal trajectory of dynamic obstacles, the range and duration of the impact of sudden weather events, updates of temporary airspace control boundaries, risk value upgrades caused by sudden changes in ground pedestrian flow heat, and updates of the locations of temporary buildings and obstacles.

[0022] Preferably, in step S3, the three types of pre-filtering conditions are as follows:

[0023] Compliance hard constraints: The entire route must not enter the airspace corresponding to the hard no-fly grid, and the cumulative overflight time within the restricted grid must not exceed the preset threshold;

[0024] Airworthiness hard constraint: The cumulative probability of air collision throughout the entire path, P_collision, ≤ / flight hours;

[0025] Physical constraints: The entire path must meet the UAV dynamics constraints, maximum endurance constraints, and maximum climb or descent angle constraints, and priority should be given to avoiding high-risk areas where GPS signals are blocked.

[0026] Preferably, in step S3, the coupled path planning cost function The expression is:

[0027]

[0028] in, This is the total path length. The total energy consumption along the path. Accumulate risk value for dynamic collisions in the air, calculated as the integral of the instantaneous collision probability between each waypoint and a dynamic obstacle on the path; The cumulative risk value for ground-based third parties is calculated based on the horizontal grid corresponding to each waypoint along the path. The integral of the product of the value and the flight time over that waypoint; , , , The weights are dynamically adaptive and automatically adjusted based on the real-time risk level of the current environment.

[0029] Preferably, in step S5, the global-local dual-layer linkage constraint mechanism is as follows: the local path planning takes the global baseline route as a reference and sets the airspace boundary for local adjustment to ensure that the local path adjustment will not break the hard no-fly constraint, will not cause the ground cumulative risk increment to exceed the preset threshold, and will not cause the collision probability to exceed the airworthiness hard constraint requirements.

[0030] The three-level replanning triggering mechanism based on risk thresholds is as follows:

[0031] Level 1 Trigger: When the probability of a collision between a sudden obstacle sensed in real time and the baseline flight path exceeds the warning threshold, or when the local environmental risk slightly increases, a local replanning is triggered, and the path is finely adjusted only within the current airspace.

[0032] Level 2 trigger: When temporary airspace control is updated, the ground risk value in a local area suddenly rises to the level of a hard no-fly zone, or GPS signal obstruction exceeds expectations, segment replanning is triggered, and only the affected flight routes are replanned in segments.

[0033] Level 3 trigger: When a sudden major meteorological disaster occurs, the airspace control is adjusted across the entire region, or the drone malfunction causes a decrease in endurance, a global replanning is triggered to regenerate a complete global baseline route.

[0034] Furthermore, in step S5, when the UAV enters the urban canyon area and experiences GPS signal loss or positioning error exceeding a preset threshold, local replanning is automatically triggered, prioritizing planning to open airspace, while switching to a low-maneuverability path adapted to visual and IMU combined navigation.

[0035] Furthermore, in step S6, the full-process risk recording and traceability verification specifically involves: recording waypoint data, risk calculation data, constraint satisfaction status, replanning triggering reasons and decision-making basis for the global baseline path and all local replanning paths, and generating a traceable flight report.

[0036] The actual flight data fed back includes actual collision risk events, actual trajectories of dynamic obstacles, actual areas of GPS signal obstruction, actual changes in ground pedestrian flow, actual energy consumption and flight duration, which are used to continuously optimize the prediction accuracy of the spatiotemporal fusion prediction model, the accuracy of the ground risk quantification model, and the adaptation logic of the dynamic weights of the coupled path planning cost function.

[0037] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0038] (1) The present invention is based on a spatiotemporal hierarchical dynamic twin modeling mechanism with risk level matching. According to the ground third-party risk level of the grid, the modeling accuracy and data update frequency are matched differently. In high-risk areas, centimeter-level high-precision modeling is used to ensure safety redundancy, while meter-level lightweight modeling is used in low-risk areas to reduce the amount of computation. While effectively reducing the overall amount of computation of the whole domain modeling, it realizes the fine management of all elements in high-risk areas. At the same time, the spatiotemporal fusion Transformer model realizes the long-term prediction of the dynamic environment in the next 5-15 minutes, upgrading the path planning from the passive response of the existing technology to the active prediction, effectively improving the response capability to dynamic scenarios such as dynamic obstacles, sudden weather, and temporary airspace control, and thus effectively solving the problems of static modeling response lag and the surge in the amount of computation of the whole domain modeling.

[0039] (2) This invention quantifies and classifies the risks of third parties on the ground and embeds them into the basic attributes of the airspace grid from the bottom-level modeling stage of path planning. It upgrades the ground compliance requirements from the post-verification items of existing technologies to the pre-hard constraints of path planning. Paths that do not meet the compliance requirements are directly eliminated. At the same time, it takes the accumulated risks of third parties on the ground as the core optimization target of the cost function, realizing the synchronous optimization of airworthiness safety and ground compliance management throughout the entire path planning process. Through this solution, the flight time of UAVs over sensitive ground areas such as schools, hospitals, and residential areas can be effectively reduced, ensuring the personal and property safety of third parties on the ground and greatly improving the compliance and public acceptance of UAV flights in cities.

[0040] (3) This invention, through a global-local dual-layer linkage constraint mechanism, ensures that local path adjustments inherit the compliance and airworthiness hard constraints of the global flight path throughout the entire process, effectively avoiding the problem of existing technologies breaking the ground compliance bottom line in order to avoid air risks; at the same time, through the pre-positioned airworthiness hard constraints, the cumulative air collision probability throughout the flight is stably controlled at 8.2×10 -9 / Flight hours, far below the stipulated 10 -7 The flight hour threshold fully meets the safety requirements for civil aviation airworthiness; a special emergency adaptation mechanism has been designed for GPS signal obstruction scenarios in urban canyons, further improving flight robustness in complex urban environments. Attached Figure Description

[0041] In the attached diagram:

[0042] Figure 1 This is a flowchart of an embodiment of the present invention. Detailed Implementation

[0043] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed in various different configurations.

[0044] Example 1

[0045] like Figure 1 As shown, a method for path planning of urban logistics drones includes the following steps:

[0046] Step S1: Construct a static 3D grid model of the city's overall air-ground integrated risk baseline:

[0047] This step is executed in the spatial database of the ground control platform, laying the underlying compliance and security foundation for route planning. The specific implementation process is as follows:

[0048] Standardized 3D raster division: The low-altitude logistics airspace of the target city, covering 30km × 20km and with an altitude of 0-120m, is divided into raster layers with a horizontal granularity of 100m × 100m and a vertical granularity of 5m. This results in 300 × 200 = 60,000 horizontal raster layers and 24 vertical height layers, totaling 1.44 million 3D raster layers. Each raster corresponds to a unique spatiotemporal coordinate (X, Y, H), where X and Y are plane coordinates in the CGCS2000 national geodetic coordinate system, and H is the altitude of the 1985 national elevation datum. Each raster is also configured with editable attribute labels.

[0049] Full collection of dual-dimensional constraint elements: Simultaneous collection of two core data types: static constraint elements in the air and third-party risk elements on the ground.

[0050] The static constraints in the air include: a 30m resolution ASTERGDEM digital elevation model, a 0.5m accuracy urban real-scene 3D building model, legally designated no-fly and restricted-fly zones published by the Civil Aviation Administration, static obstacles such as high-voltage transmission towers, communication towers, and bridges, and high-risk areas of GPS signal obstruction in urban canyons; the high-risk areas of GPS signal obstruction in urban canyons are pre-marked by areas where the ratio of building height to building spacing is ≥2, combined with areas with positioning errors ≥5m in historical flight data;

[0051] Ground-based third-party risk factors include: land use types determined by the national land space planning (land for primary and secondary schools, medical and health care, residential land, commercial and service land, park green space, and water area), static population density data at the community level from the population census, statutory control levels of sensitive locations, distribution of urban emergency landing sites (park open spaces, municipal squares, river beaches, etc.), and building density (the ratio of building footprint area to grid area).

[0052] Ground-based third-party risk quantification and hierarchical management: A quantification model coupled with the entropy weight method and the analytic hierarchy process is adopted to standardize and quantify the ground-based third-party risk of each horizontal grid, generating a risk value R_ground in the 0-1 interval. The specific quantification process is as follows:

[0053] Four core evaluation indicators were identified: land use type sensitivity level, static population density, building density, and distance to the nearest emergency landing site.

[0054] By combining the analytic hierarchy process (AHP) with scores from three experts in the field of low-altitude safety, the subjective weights of the indicators were determined as follows: land use type 0.4, static population density 0.3, building density 0.2, and emergency landing distance 0.1.

[0055] The objective weights of the indicators were calculated using the entropy weight method through 100 sets of sample raster data. The subjective weights and objective weights were then linearly coupled to obtain the comprehensive weights.

[0056] The positive indicators (higher values ​​indicate higher risk) and negative indicators (higher values ​​indicate lower risk) are respectively subjected to min-max standardization, and the R_ground value of each horizontal grid is finally calculated.

[0057] Based on the calculated R_ground value, set up three levels of rigid control rules:

[0058] Hard no-fly grid: R_ground≥0.9, including primary and secondary schools, hospital core areas, large commercial district core areas, and airport airspace core areas, set as an absolute no-flying prerequisite hard constraint, and no path may enter the airspace corresponding to this grid;

[0059] Restricted Flying Grid: 0.3≤R_ground<0.9, including ordinary residential areas and urban side roads, only allowed to fly over at a preset height of 60-80m, during designated time slots outside of morning and evening peak hours (7:00-9:00, 17:00-19:00), and the cumulative flying time of a single path in this type of grid shall not exceed 20% of the total flight time;

[0060] Flyable grid: R_ground < 0.3, including parks, rivers, green belts along main urban roads, and open spaces, set as priority planning airspace, with no additional flight duration or height restrictions.

[0061] Generation of a static risk baseline model across the entire domain: All static constraints in the air, high-risk areas marked by GPS signal obstruction, and third-party risk classification and control rules on the ground are mapped to the corresponding three-dimensional grids. Each grid is then used to generate a flyability label (flyable / restricted / no-fly), risk level, and constraints. This process ultimately forms a static risk baseline model across the entire domain, which is stored in the spatial database of the ground control platform as the underlying foundation for path planning.

[0062] Step S2: Construct a spatiotemporal hierarchical dynamic twin model based on risk level matching:

[0063] This step is executed within the digital twin engine of the ground control platform, and its core purpose is to resolve the fundamental contradiction between the lag in static modeling response and the balance between modeling accuracy and real-time planning. The specific implementation process is as follows:

[0064] The risk level matching hierarchical modeling mechanism: Based on the risk level of each grid in step S1, the 3D modeling accuracy and data update frequency of each grid are dynamically matched. Accuracy differentiation is achieved through the degree of 3D point cloud thinning. The specific matching rules are as follows:

[0065] High-risk areas: the area within 500m of the no-fly zone and the core area of ​​the restricted-fly zone, using 5cm-level high-precision modeling, retaining 100% of the 3D point cloud data, with a data update frequency of ≤1Hz;

[0066] Medium-risk area: a 200m buffer zone around the restricted fly grid, modeling at the 20cm decimeter level, thinning the point cloud to 20% of the original data, and a data update frequency of ≤5Hz;

[0067] Low-risk area: Flyable grid across the entire area, using 1m-level lightweight modeling, point cloud thinning to 5% of the original data, and data update frequency ≤10Hz.

[0068] Through the aforementioned hierarchical modeling mechanism, while ensuring safety redundancy in high-risk areas, the overall computational load of global modeling is reduced by 62%, fundamentally resolving the contradiction between accuracy and real-time performance.

[0069] Real-time access to multi-source dynamic environmental data: Through standardized API interfaces and private network communication, multi-source dynamic environmental data is accessed synchronously, including: minute-level real-time meteorological monitoring and 15-minute forecast data from the local meteorological bureau (wind speed, wind direction, rainfall, visibility, and severe convective weather warnings), temporary airspace control notices from the Civil Aviation Administration's low-altitude flight service platform, real-time flight plans and location data of urban logistics drone clusters from the urban low-altitude collaborative management and control platform, minute-level ground pedestrian flow heat map data from the internet map platform, temporary building / construction site update data from the urban housing and construction department, real-time bird activity monitoring data and historical migration pattern data from the urban wildlife monitoring station, and real-time positioning and sensor data transmitted back by drone onboard equipment.

[0070] Long-term dynamic environment proactive prediction: A pre-trained spatiotemporal fusion Transformer model is used as input, taking multi-source spatiotemporal sequence data from the past 30 minutes as input, and outputting dynamic environment prediction results for the next 5-15 minutes. The spatiotemporal fusion Transformer model adopts an Encoder-Decoder architecture. The Encoder consists of 6 Transformer encoder layers, used to extract spatiotemporal features from multi-source data, and the Decoder consists of 4 Transformer decoder layers, used to generate prediction results for the future time series. The model is pre-trained using one year of historical flight and environmental data of the target city, with a single-frame inference time ≤50ms, meeting real-time requirements.

[0071] The output dynamic environment prediction results specifically include: spatiotemporal trajectory prediction of dynamic obstacles such as birds and other aircraft, impact range and duration of sudden weather events, boundary updates of temporary airspace control, risk value upgrades caused by sudden changes in ground pedestrian flow heat, and location updates of temporary buildings / obstacles.

[0072] The dynamic twin model is updated in real time: the dynamic environment prediction results are updated to the static risk base model in real time. For the predicted risk escalation area, the modeling accuracy and update frequency of the corresponding grid are automatically improved. Finally, a spatiotemporally continuous hierarchical dynamic digital twin model is generated, which provides an environmental foundation for subsequent path planning with proactive prediction, accuracy adaptation and dual risk coupling.

[0073] The dynamic twin model is updated in real time: the dynamic environment prediction results are updated to the static risk base model in real time. For the predicted risk escalation area, the modeling accuracy and update frequency of the corresponding grid are automatically improved. Finally, a spatiotemporally continuous hierarchical dynamic digital twin model is generated, which provides an environmental foundation for subsequent path planning with proactive prediction, accuracy adaptation and dual risk coupling.

[0074] Step S3: Construct the coupled path planning cost function with dual-risk hard constraints in advance:

[0075] The core of this step is to upgrade airworthiness safety and ground third-party compliance requirements from post-verification items of existing technologies to pre-defined hard constraints for path planning. The specific implementation process is as follows:

[0076] Three types of rigid pre-filtering conditions are set: compliance hard constraints, airworthiness hard constraints, and physical hard constraints. During the path planning process, candidate paths that do not meet any of these conditions are directly eliminated and will not proceed to subsequent optimization stages. The specific thresholds are as follows:

[0077] Strict compliance constraints: The entire flight path must not enter the airspace corresponding to the restricted flight grid, and the cumulative overflight time within the restricted flight grid must not exceed 20% of the total flight time;

[0078] Airworthiness hard constraint: The cumulative probability of air collision throughout the entire path, P_collision, ≤ / flight hours;

[0079] Physical constraints: The entire path must meet the drone's dynamic constraints (maximum climb / descent angle ≤ 15°, maximum flight speed ≤ 15m / s) and maximum endurance constraints (remaining battery power ≥ 20% safety threshold throughout the flight), and prioritize avoiding high-risk areas with GPS signal obstruction.

[0080] Construction of a multi-objective optimization cost function coupled with dual risks: For paths that satisfy the pre-filtering conditions, the following coupled path planning cost function is constructed. As the optimization objective of path planning:

[0081]

[0082] The specific calculation method for each item in the formula is as follows:

[0083] Total path length, which is the sum of the Euclidean distances between all adjacent waypoints in the path, in meters;

[0084] Total energy consumption along the path, calculated based on a pre-calibrated UAV energy consumption model, taking into account factors such as flight speed, climb / descent attitude, environmental wind resistance, and payload, and is expressed in Wh.

[0085] The cumulative risk value of dynamic collisions in the air is calculated as the integral of the instantaneous collision probability between each waypoint and a dynamic obstacle on the path and the flight time of that waypoint; the instantaneous collision probability is calculated based on the probability of overlap between the error ellipse of the dynamic obstacle trajectory prediction and the position of the UAV waypoint.

[0086] Ground-based third-party cumulative risk value, calculated as the horizontal grid corresponding to each waypoint on the path. The sum of the products of the value and the flight time to that waypoint;

[0087] , , , For dynamic adaptive weights, the initial default weights are... =0.2、 =0.2、 =0.3、 =0.3, automatically adjusted based on the real-time risk level of the current environment: when predicting dense dynamic obstacles or sudden weather events, Automatically boosted to 0.6, with other weights compressed proportionally; the path is located in a densely populated area ( When ≥0.6), Automatically increased to 0.6, with other weights compressed proportionally; in time-sensitive scenarios such as emergency medical supplies delivery, Automatically boosted to 0.5, with the remaining weights compressed proportionally.

[0088] Step S4, Global Baseline Path Planning:

[0089] This step, based on a hierarchical dynamic digital twin model and a coupled cost function, generates a globally optimal baseline route that satisfies both constraints throughout the entire journey. The specific implementation process is as follows:

[0090] Planning space dimensionality reduction: Based on the hierarchical dynamic digital twin model, the airspace corresponding to all hard no-fly grids is eliminated, and the state space of path planning is compressed from 1.44 million three-dimensional grids in the whole domain to about 400,000 flyable / restricted grids, which greatly reduces the planning calculation dimensionality and improves planning efficiency.

[0091] Multi-objective global path optimization: An improved NSGA-III multi-objective optimization algorithm is adopted, with the drone's starting point (logistics warehouse), ending point (delivery point), and passing delivery nodes as inputs, and the coupled cost function constructed in step S3 as the optimization objective. The algorithm parameters are set as follows: population size 100, number of iterations 200, crossover probability 0.9, and mutation probability 0.1. The algorithm outputs a Pareto optimal path solution set that satisfies all pre-filtering conditions.

[0092] Optimal route selection and pre-verification: Based on the dynamic adaptive weights corresponding to the current environment, the TOPSIS method is used to select the globally optimal baseline route from the Pareto optimal path solution set; the global baseline route is a continuous sequence of waypoints, each waypoint containing spatiotemporal coordinates (X,Y,H,t), flight speed, heading angle, and the waypoint interval is 50m.

[0093] The generated global baseline route undergoes full-process dual-risk pre-verification, outputting a complete risk assessment report, including the instantaneous collision probability, ground risk value, cumulative collision probability, cumulative ground risk, and constraint condition satisfaction for each waypoint. This ensures that the route meets 100% of the pre-filtering requirements. Finally, the verified global baseline route is sent to the UAV flight control system and ground management platform.

[0094] Step S5: Local dynamic obstacle avoidance and hierarchical replanning with global-local linkage:

[0095] The core of this step is to address the disconnect between global planning and local obstacle avoidance, ensuring that local path adjustments inherit global compliance and airworthiness constraints throughout the process, while also achieving real-time response in dynamic scenarios. The specific implementation process is as follows:

[0096] A global-local dual-layer linkage constraint mechanism is established: the local path planning inherits the three types of pre-filter conditions of the global baseline route throughout the entire process. At the same time, with the global baseline route as the center, a horizontal airspace boundary of ±20m and an altitude boundary of ±5m are set. Local path adjustments must not exceed the airspace boundary and must not cause the cumulative ground risk increase to exceed 10% of the global baseline value or the air collision probability to exceed the airworthiness threshold, thus completely avoiding local avoidance from breaking the compliance and airworthiness bottom line.

[0097] Millisecond-level local dynamic obstacle avoidance: Centered on the current position of the UAV, within a high-precision modeling area of ​​200m, an improved model predictive control algorithm is used to achieve local dynamic obstacle avoidance. The algorithm parameters are set as follows: 10 steps in the control time domain, 20 steps in the prediction time domain, control period of 50ms, and local planning response time ≤100ms. The algorithm input is the real-time perception data of the UAV's onboard LiDAR and binocular vision camera, as well as the dynamic environment prediction results generated in step S2. The output is a locally fine-tuned waypoint sequence, which realizes real-time avoidance of sudden obstacles, while meeting the linkage constraint requirements throughout the process.

[0098] A risk threshold-based three-level replanning trigger mechanism: Three-level replanning trigger conditions are set, dynamically triggering path replanning at the corresponding level based on changes in environmental risk. This avoids invalid calculations and improves response efficiency. The specific triggering rules are as follows:

[0099] Level 1 Trigger (Local Replanning): The probability of a collision between a sudden obstacle sensed in real time and the baseline flight path is ≥ Triggered when the ground risk value in a local area rises to ≤0.2, the path is finely adjusted only within the current 200m range, and automatically returns to the global baseline route after the adjustment is completed;

[0100] Level 2 Trigger (Segmented Replanning): Triggered when receiving a temporary airspace control notice, when the ground risk value in a local area suddenly rises to ≥0.9 (hard no-fly zone), or when the GPS positioning error is ≥5m and lasts for more than 3 seconds. Only the affected segments of the route within 1km before and after the affected segment are replanned, while the unaffected segments retain the baseline route.

[0101] Level 3 Trigger (Global Replanning): Triggered when receiving a major weather warning such as heavy rain / winds of level 6 or above, adjustments to airspace control across the entire region, or when a drone's battery failure causes a decrease in endurance of ≥30%. The original baseline route is cleared, and the global planning process in step S4 is re-executed to generate a brand new global baseline route.

[0102] Emergency adaptation for GPS signal obstruction scenarios: When the drone enters an urban canyon area and loses its GPS signal, or when the RTK positioning error exceeds 10m and lasts for 2 seconds, it will automatically trigger a first-level local replanning, prioritizing the planning to open airspace such as surrounding parks and rivers. At the same time, the flight speed will be reduced to 5m / s, and the drone will switch to the visual + IMU combined navigation mode to plan a low-maneuver straight flight path, avoiding sharp turns and ensuring flight safety.

[0103] Step S6: Risk verification and closed-loop optimization for airworthiness traceability:

[0104] The core of this step is to address the problem of intelligent planning algorithms being black-box and unable to meet airworthiness traceability requirements, while simultaneously achieving continuous self-optimization of model performance. The specific implementation process is as follows:

[0105] Full-process risk traceability and verification: During the flight of the UAV, the ground control platform and the airborne flight control system synchronously record all flight data, including: all waypoint data of the global baseline route, waypoint data of each local replanning, risk calculation data of each waypoint, constraint condition satisfaction, replanning triggering reasons and decision basis, raw data of airborne sensors, and flight control operation command data.

[0106] After the flight mission is completed, a traceable flight report is automatically generated. The report includes the complete flight trajectory, risk change curve, constraint verification results, and replanning event details. All data is encrypted and stored for a period of ≥1 year.

[0107] Flight data closed-loop feedback and model self-optimization: Actual flight data, including actual collision risk events, actual trajectories of dynamic obstacles, actual areas of GPS signal obstruction, actual changes in ground pedestrian traffic, actual energy consumption, and flight duration, is fed back to the model optimization module of the ground control platform to achieve continuous self-optimization. Specific optimization content includes:

[0108] Incremental fine-tuning of the spatiotemporal fusion Transformer model is performed, and the model parameters are updated once every 100 flights to continuously improve the accuracy of dynamic environment prediction and control the trajectory prediction error within 5m.

[0109] Based on actual human flow thermal change data, the indicator weights of the ground risk quantification model are optimized to improve the accuracy of R_ground value calculation;

[0110] Based on actual flight energy consumption, duration, and risk data, the adjustment rules for the dynamic adaptive weights of the cost function are optimized to make the optimization objective more suitable for actual urban flight scenarios.

[0111] The above embodiments only illustrate preferred embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, improvements, and substitutions without departing from the concept of the present invention, and these all fall within the protection scope of the present invention.

Claims

1. A method for path planning of urban logistics drones, characterized in that, Includes the following steps: S1. Construct a static three-dimensional grid model of the city’s air-ground integrated risk base: Divide the target city’s low-altitude logistics airspace into multiple three-dimensional grids, collect static airspace constraint elements and ground third-party risk elements simultaneously, quantify the ground third-party risk value corresponding to each grid and perform hierarchical control, map the results of airspace static constraints and ground third-party risk classification to the three-dimensional grid, generate the flyability label, risk level and constraint conditions of each grid, and form a city-wide static risk base model. S2. Construct a spatiotemporal hierarchical dynamic twin model based on risk level matching: Based on the risk level of each grid in the static risk base model, dynamically match the modeling accuracy and data update frequency corresponding to each grid; access multi-source dynamic environmental data, generate dynamic environmental prediction results for a preset duration through a spatiotemporal fusion prediction model, update the prediction results to the static risk base model in real time, and generate a spatiotemporally continuous hierarchical dynamic digital twin model. S3. Construct a coupled path planning cost function with dual-risk hard constraints in advance: Set three types of pre-filtering conditions: compliance hard constraints, airworthiness hard constraints, and physical hard constraints. Paths that do not meet any of the pre-filtering conditions are directly eliminated. For paths that meet the pre-filtering conditions, construct a multi-objective optimization cost function that integrates path length, energy consumption, cumulative risk of dynamic collision in the air, and cumulative risk of third parties on the ground. S4. Global baseline route planning: Based on the hierarchical dynamic digital twin model, hard no-fly zones are eliminated to compress the route planning space. A multi-objective optimization algorithm is adopted, with the coupled route planning cost function as the optimization objective, to generate a global baseline route that meets all pre-filtering conditions and complete dual risk pre-verification. S5. Global-local linkage local dynamic obstacle avoidance and hierarchical replanning: Establish a global-local two-layer linkage constraint mechanism, and the local path planning inherits the three types of pre-filter conditions of the global baseline route throughout the process; perform local path fine-tuning and dynamic obstacle avoidance based on UAV real-time perception data and dynamic environment prediction results. A three-level replanning trigger mechanism based on risk thresholds is set up to dynamically trigger path replanning of the corresponding level according to changes in environmental risk. S6. Airworthiness traceability risk verification and closed-loop optimization: Record and verify the risks of all executed paths throughout the entire process and generate flight risk reports; Actual flight data is fed back to the spatiotemporal fusion prediction model and the coupled path planning cost function optimization module to achieve continuous self-optimization of the model.

2. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S1, the low-altitude logistics airspace of the target city is divided into a three-dimensional grid with a horizontal granularity of 100m×100m and a vertical granularity of 5m. Each grid corresponds to a unique spatiotemporal coordinate and attribute label. A quantitative model coupled with entropy weighting and analytic hierarchy process is used to standardize and quantify the ground third-party risk of each horizontal grid, generating a risk value R_ground of 0-1. Based on the risk value, three levels of rigid control rules are set. Hard no-fly grid: R_ground≥0.9, set as a prerequisite hard constraint that no one can fly over; Restricted Flying Grids: 0.3≤R_ground<0.9, flying is only allowed within a preset height and specified time slot, and the cumulative flying time of a single path in this type of grid must not exceed a preset threshold; Flyable grid: R_ground < 0.3, which is the priority for airspace planning.

3. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S1, the static constraints in the air include DEM (Digital Elevation Model), 3D building models, legally designated no-fly or restricted-fly zones, static obstacles, and high-risk areas in urban canyons where GPS signals are blocked; the third-party risk factors on the ground include land use type, static population density, legal control level of sensitive locations, distribution of emergency landing sites, and building density.

4. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S2, the matching rules between risk level and modeling accuracy and update frequency are as follows: High-risk areas: the 500m radius around the no-fly zone and the core area of ​​the restricted-fly zone, using centimeter-level high-precision modeling with a data update frequency of ≤1Hz; Medium-risk area: The buffer zone around the restricted grid is limited, and decimeter-level modeling is used with a data update frequency of ≤5Hz; Low-risk area: Flyable grid across the entire area, using meter-level lightweight modeling, with a data update frequency of ≤10Hz.

5. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S2, the accessed multi-source dynamic environmental data includes real-time meteorological monitoring and forecast data, temporary airspace control notices, real-time flight plans and location data of same-city logistics drone clusters, real-time ground pedestrian flow heat data, updated data of temporary urban buildings and construction areas, real-time monitoring and historical spatiotemporal pattern data of bird activities, and real-time drone positioning and sensor data. The spatiotemporal fusion prediction model is a spatiotemporal fusion Transformer model, which outputs dynamic environmental prediction results for the next 5-15 minutes. The dynamic environmental prediction results include prediction of the spatiotemporal trajectory of dynamic obstacles, the range and duration of the impact of sudden weather events, updates of temporary airspace control boundaries, risk value upgrades caused by sudden changes in ground pedestrian flow heat, and updates of the locations of temporary buildings and obstacles.

6. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S3, the three types of pre-filtering conditions are as follows: Compliance hard constraints: The entire route must not enter the airspace corresponding to the hard no-fly grid, and the cumulative overflight time within the restricted grid must not exceed the preset threshold; Airworthiness hard constraint: The cumulative probability of air collision throughout the entire path, P_collision, ≤ / flight hours; Physical constraints: The entire path must meet the UAV dynamics constraints, maximum endurance constraints, and maximum climb or descent angle constraints, and priority should be given to avoiding high-risk areas where GPS signals are blocked.

7. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S3, the coupled path planning cost function The expression is: in, This is the total path length. The total energy consumption along the path. Accumulate risk value for dynamic collisions in the air, calculated as the integral of the instantaneous collision probability between each waypoint and a dynamic obstacle on the path; The cumulative risk value for ground-based third parties is calculated based on the horizontal grid corresponding to each waypoint along the path. The integral of the product of the value and the flight time over that waypoint; , , , The weights are dynamically adaptive and automatically adjusted based on the real-time risk level of the current environment.

8. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S5, the global-local dual-layer linkage constraint mechanism is as follows: the local path planning takes the global baseline route as a reference and sets the airspace boundary for local adjustment to ensure that the local path adjustment will not break the hard no-fly constraint, will not cause the ground cumulative risk increment to exceed the preset threshold, and will not cause the collision probability to exceed the airworthiness hard constraint requirements. The three-level replanning triggering mechanism based on risk thresholds is as follows: Level 1 Trigger: When the probability of a collision between a sudden obstacle sensed in real time and the baseline flight path exceeds the warning threshold, or when the local environmental risk slightly increases, a local replanning is triggered, and the path is finely adjusted only within the current airspace. Level 2 trigger: When temporary airspace control is updated, the ground risk value in a local area suddenly rises to the level of a hard no-fly zone, or GPS signal obstruction exceeds expectations, segment replanning is triggered, and only the affected flight routes are replanned in segments. Level 3 trigger: When a sudden major meteorological disaster occurs, the airspace control is adjusted across the entire region, or the drone malfunction causes a decrease in endurance, a global replanning is triggered to regenerate a complete global baseline route.

9. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S5, when the UAV enters the urban canyon area and experiences GPS signal loss or positioning error exceeding a preset threshold, local replanning is automatically triggered, prioritizing planning to open airspace, while switching to a low-maneuverability path adapted to visual and IMU combined navigation.

10. A method for urban logistics drone path planning according to claim 1, characterized in that, In step S6, the full-process risk recording and traceability verification specifically involves: recording waypoint data, risk calculation data, constraint satisfaction status, replanning triggering reasons and decision-making basis for the global baseline path and all local replanning paths, and generating a traceable flight report. The actual flight data fed back includes actual collision risk events, actual trajectories of dynamic obstacles, actual areas of GPS signal obstruction, actual changes in ground pedestrian flow, actual energy consumption and flight duration, which are used to continuously optimize the prediction accuracy of the spatiotemporal fusion prediction model, the accuracy of the ground risk quantification model, and the adaptation logic of the dynamic weights of the coupled path planning cost function.

Images