A dynamic path planning method and system for multi-drone collaborative delivery

Abstract

This invention provides a dynamic path planning method and system for multi-UAV collaborative delivery, relating to the field of UAV technology. The dynamic path planning method specifically includes: deploying a reconnaissance UAV in front of a carrier UAV to form a collaborative formation; the reconnaissance UAV performs advance environmental perception of the airspace to be explored ahead, integrating a ground risk weight map, a space airflow risk and noise impact map to generate a comprehensive cost map, and generating a time-sensitive three-dimensional safety bubble cube; the reconnaissance UAV sends the flight path segment and the safety bubble cube to the carrier UAV for flight execution; simultaneously, the reconnaissance UAV moves forward to repeat the advance environmental perception and safety bubble cube generation steps for the next flight segment; the above steps are executed cyclically until the mission is completed; and a dynamic replanning mechanism is triggered when an unacceptable risk is detected ahead or the flight deviates from the safety envelope at any stage.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and more specifically, to UAV flight path planning technology, and particularly to a dynamic path planning method and system for collaborative logistics delivery of multiple UAVs in complex urban environments. Background Technology

[0002] Unmanned aerial vehicles (UAVs), as a rapidly developing technology, have demonstrated enormous application potential in numerous industries. Especially in the logistics and delivery sector, UAVs, with their advantages of high efficiency, high speed, and the ability to effectively avoid ground traffic congestion, are considered one of the key technological paths to solving the last-mile delivery problem. Automated goods delivery via UAVs not only promises to significantly reduce labor costs but also improve delivery timeliness and service quality. Therefore, building a stable, safe, and efficient UAV delivery network has become a current research hotspot.

[0003] One of the core technological challenges for drones performing delivery tasks in urban environments lies in path planning. Urban environments are highly complex and dynamic. Dense clusters of high-rise buildings create complex urban canyon effects, while numerous static and dynamic obstacles exist, such as high-voltage power lines, communication towers, construction cranes, moving vehicles, and pedestrians. Traditional path planning methods typically rely on pre-built two-dimensional or three-dimensional static map information, employing graph search algorithms or sampling algorithms to calculate a globally optimal path from the starting point to the destination. However, these methods heavily depend on the accuracy and timeliness of map data, making it difficult to handle frequent, unmapped temporary obstacles or dynamic changes in the real world, thus limiting their safety and reliability in practical applications.

[0004] To address the shortcomings of static path planning, dynamic path planning technology has emerged. Dynamic path planning allows drones to perceive their surroundings in real time during flight using onboard sensors such as vision, lasers, or radar. When a sudden obstacle is detected, the drone can immediately replan its local path to avoid it. This method enhances the drone's adaptability to unknown environments. However, the single-drone dynamic path planning approach also has inherent limitations. The drone's own sensors have limited detection range and capabilities, and its perception is usually confined to the short-range space around the aircraft, essentially a passive, reactive obstacle avoidance strategy. This strategy struggles to pre-assess and avoid complex environmental situations further along the flight path, such as local microclimate changes, airflow turbulence, or potentially high-risk areas. This is especially true for heavy-payload drones performing missions, whose maneuverability is relatively poor, making obstacle avoidance at the last minute extremely risky.

[0005] With the expanding scale of drone applications, multi-drone collaborative operations have become an important development direction for improving the overall efficiency and capability of systems. Currently, research on multi-drone systems mainly focuses on task allocation and scheduling, swarm formation control, and collision avoidance. For example, a central control system can rationally allocate multiple delivery tasks to different drones within a swarm and plan their respective macroscopic paths to avoid mutual interference. While these technologies improve the macroscopic efficiency of task execution, in terms of environmental perception and safety assurance, each drone typically still independently addresses environmental risks along its own flight path. Existing collaborative mechanisms generally lack a deep strategy that combines multiple drones in a functionally heterogeneous manner, utilizing the perception capabilities of some drones to provide other drones with forward-looking, beyond-line-of-sight environmental safety information assurance. Therefore, how to leverage the collaborative advantages of multi-drone systems to construct a forward-looking dynamic path planning method capable of proactively avoiding various known and unknown risks to ensure the flight safety of heavy-load drones in complex environments, especially in low-visibility conditions such as nighttime when sensor performance is limited, is a pressing technical problem to be solved in this field. Summary of the Invention

[0006] This invention discloses a dynamic path planning method for multi-UAV collaborative delivery, the method comprising: Deploy a reconnaissance drone in front of a carrier drone to form a coordinated formation and begin the mission along the global reference flight path; The reconnaissance drone conducts advance environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. The ground risk weight map, the space airflow risk and noise impact map are integrated to generate a comprehensive cost map. Combined with real-time perceived air obstacle information, a time-sensitive three-dimensional safety bubble cube is generated in the low-cost area of ​​the comprehensive cost map, and the next flight path segment is planned within the safety bubble cube. The reconnaissance UAV sends the flight path segment and the safety bubble cube to the carrier UAV for flight execution. At the same time, the reconnaissance UAV moves forward to repeat the leading environment perception and safety bubble cube generation steps for the next flight segment. The above steps are repeated until the mission is completed. If at any stage an unacceptable risk is detected ahead or the flight deviates from the safe envelope, a dynamic replanning mechanism is triggered, in which the carrier UAV hovers and the reconnaissance UAV searches for an alternative path.

[0007] The process of fusing multimodal sensor data to generate a ground risk weight map specifically includes: performing multi-stage parallel feature extraction and interactive fusion of lidar data, visible light images, and infrared thermal imaging data through a hierarchical interactive encoder containing geometric structure branches, visual semantic branches, and dynamic thermodynamic branches. In at least one fusion stage, the dynamic thermodynamic features extracted by the dynamic thermodynamic branch... Used to generate gating signals Geometric features to guide the branches of the geometric structure semantic features of the visual semantic branch The fusion is performed, and the calculation method is as follows: in, The fusion features output at this stage As a standard processing unit, For element-wise multiplication, The process involves element-wise addition; and the ground risk weight map is generated based on the final fusion features. .

[0008] The analysis of its own flight status data to calculate space airflow risk specifically includes: calculating the real-time control torque vector based on the flight control motor control commands of the reconnaissance UAV. and in the sliding time window The total variance is calculated internally to obtain the turbulence intensity index. : Furthermore, the angular velocity error is calculated based on the inertial measurement unit data of the reconnaissance drone. The wind field direction characteristics were obtained by low-pass filtering. : in, For the average control torque, It is a low-pass filter. The cutoff frequency is used; and the space airflow risk map is constructed based on the turbulence intensity index.

[0009] The assessment and generation of the noise impact map specifically includes: calculating the noise impact map of the reconnaissance drone at a specific receiving point. The sound pressure level generated at the location : And based on the perceived ambient background noise sound pressure level Calculate the acoustic influence index : in, For sound power level, As a directional factor, , , These are geometric diffusion, atmospheric absorption, and obstacle diffraction attenuation, respectively; and the noise impact map is generated based on the acoustic impact index calculated from all noise-sensitive points. .

[0010] The process of integrating the ground risk weight map, space airflow risk, and noise impact map to generate a comprehensive cost map specifically includes: For the space airflow risk map and the noise impact diagram After normalization, we get and ; By analyzing the ground risk weight map and after normalization and The weighted summation yields the comprehensive cost diagram. : in, This is a preset risk weighting coefficient.

[0011] The step of generating the three-dimensional safety bubble cube specifically includes: The comprehensive cost diagram Above, search for results that meet the preset cost threshold. Two-dimensional safe flight area The two-dimensional safe flight area Extending vertically, and utilizing the global point cloud acquired by the lidar. Constrain the expanded three-dimensional space, which exists in any ground grid. Maximum safe height satisfy: in, This refers to the set of lidar points located in the vertical space above the grid. Let be the perpendicular coordinates of the point. To provide a safety margin, the safety bubble cube is generated. .

[0012] The flight of the carrier drone also includes: the carrier drone continuously checking its current position while flying along the flight path segment. Is it located within the safety bubble cube? Within the boundaries, and examine the current moment. Is it within the effective time window of the safety bubble cube? Internally, this means continuously verifying the following conditions: in, The timestamp for the generation of the safety bubble cube is used; if any condition is not met, the dynamic replanning mechanism is triggered.

[0013] In the deployment steps, the global reference route Through a preprocessed compliance subgraph The A* search algorithm is used to solve for minimizing the overall toll cost. The obtained, the The calculation method is as follows: in, For flight segment distance, For the maximum possible distance, Let be the static risk potential field function. The maximum risk value, and These are the weighting coefficients.

[0014] In the deployment step, the reconnaissance drone flies in front of the carrier drone to form a coordinated formation, which is achieved by controlling the reconnaissance drone to reach an instantaneous target position vector. This is achieved through a calculation method as follows: in, Let be the real-time position vector of the transport drone. To provide the global reference route The projection operator, The preset safe leading distance, The unit tangent vector of the global reference route at the projection point.

[0015] This invention also discloses a dynamic path planning system for multi-drone collaborative delivery, the system comprising: Formation Deployment Module: Deploy a reconnaissance UAV in front of a carrier UAV to form a coordinated formation and begin the mission along the global reference flight path; Weight generation module: The reconnaissance UAV performs advance environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. Path planning module: It integrates the ground risk weight map, the space airflow risk and noise impact map to generate a comprehensive cost map, and combines the real-time perceived air obstacle information to generate a time-sensitive three-dimensional safety bubble cube in the low-cost area of ​​the comprehensive cost map, and plans the next flight path segment within the safety bubble cube. Iterative optimization module: The reconnaissance UAV sends the flight path segment and the safety bubble cube to the carrier UAV for flight execution, while the reconnaissance UAV moves forward to repeat the leading environment perception and safety bubble cube generation steps for the next flight segment; Anomaly planning module: The above steps are executed repeatedly until the mission is completed. If an unacceptable risk is detected ahead or the flight deviates from the safe envelope at any stage, a dynamic replanning mechanism is triggered, in which the carrier UAV hovers and the reconnaissance UAV searches for an alternative path.

[0016] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned dynamic path planning method for multi-drone collaborative delivery.

[0017] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned dynamic path planning method for multi-drone collaborative delivery.

[0018] Compared with existing technologies, the dynamic path planning method and system for multi-drone collaborative delivery disclosed in this invention fundamentally solves the contradiction between the limited perception capabilities and high safety risks of a single drone in complex urban nighttime environments through a heterogeneous collaborative formation consisting of a lightweight reconnaissance drone and a heavy-payload transport drone. In existing technologies, the drone performing the delivery task typically bears all responsibility for environmental perception and obstacle avoidance. For heavy-payload drones with poor maneuverability, their reactive obstacle avoidance mode poses significant safety hazards. This invention decouples perception and transport tasks, allowing the more flexible reconnaissance drone with extremely low crash risk to first enter unknown airspace for reconnaissance. This ensures that the several-kilogram transport drone always flies within a safety space that has been verified in real time. This is a forward-looking system-level deployment method that maximizes safety margins, greatly improving the inherent safety of urban drone delivery.

[0019] This invention establishes a comprehensive risk assessment system that far surpasses traditional obstacle avoidance methods, encompassing multiple dimensions. Existing technologies largely focus on detecting physical obstacles, neglecting other equally dangerous risk sources. This invention, through a hierarchical interactive fusion network specifically designed for nighttime scenarios, not only achieves high-precision assessment of dynamic and static ground risks but also creatively introduces two key environmental risk dimensions: First, by indirectly extrapolating the flight attitude and control torque of the reconnaissance drone, it achieves quantitative perception of invisible airborne microclimates and turbulent fields, resolving the significant hidden danger of drone instability due to sudden airflow disturbances; second, by establishing an acoustic model and combining it with environmental background noise, it achieves proactive prediction and avoidance of flight path noise impacts, addressing the social acceptability challenge of nighttime delivery. This method of unifying and fusing ground physical risks, airflow risks, and social environmental risks constructs a deep understanding of the flight environment, elevating path decisions from simply whether a path can be cleared to a higher dimension of whether it is sufficiently safe, stable, and quiet.

[0020] The leapfrog rolling path planning mechanism employed in this invention ensures the absolute timeliness of decision-making, fundamentally avoiding the information lag risk inherent in relying on global path planning. A complete path planned in one go using traditional methods may become invalid in the rapidly changing urban environment within the next second. This invention takes the opposite approach, decomposing long-distance tasks into a series of short-distance, time-sensitive safety bubble cube generation and traversal processes. Each path instruction received by the transport drone is accompanied by a valid timestamp of the multi-dimensional environmental information upon which it is generated. Once the information expires, the safety bubble automatically becomes invalid, forcing the system to re-perceive and re-decide. This rigorous, time-limited safety authentication and iterative execution closed-loop logic, coupled with a powerful dynamic replanning mechanism, ensures that the drone convoy operates on a solid and reliable safety foundation at every step throughout the entire delivery mission, thereby achieving highly robust and trustworthy unmanned operation in the complex and imperfect real world. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the dynamic path planning for multi-drone collaborative delivery according to the present invention; Figure 2 This is a verification diagram of the turbulence intensity index solution for the flight control torque of this invention; Figure 3This is a diagram showing the effect of attitude angular velocity error filtering and main wind direction feature extraction of the present invention; Figure 4 This is an image showing the results of the spatial infrared thermal imaging exploration of this invention. Detailed Implementation

[0023] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0024] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] It should be noted that various aspects of embodiments within the scope of the appended claims are described below. It will be apparent that the aspects described herein can be embodied in a wide variety of forms, and any particular structure and / or function described herein is merely illustrative. Based on this application, those skilled in the art will understand that one aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspects set forth herein can be used to implement the device and / or practice the method. Additionally, this device and / or method can be implemented using structures and / or functionalities other than one or more of the aspects set forth herein.

[0026] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. However, those skilled in the art will understand that practice can be carried out without these specific details.

[0027] This embodiment discloses a dynamic path planning method for multi-drone collaborative delivery, the method comprising: Deploy a reconnaissance drone in front of a carrier drone to form a coordinated formation and begin the mission along the global reference flight path; The reconnaissance drone conducts advance environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. The ground risk weight map, the space airflow risk and noise impact map are integrated to generate a comprehensive cost map. Combined with real-time perceived air obstacle information, a time-sensitive three-dimensional safety bubble cube is generated in the low-cost area of ​​the comprehensive cost map, and the next flight path segment is planned within the safety bubble cube. The reconnaissance UAV sends the flight path segment and the safety bubble cube to the carrier UAV for flight execution. At the same time, the reconnaissance UAV moves forward to repeat the leading environment perception and safety bubble cube generation steps for the next flight segment. The above steps are repeated until the mission is completed. If at any stage an unacceptable risk is detected ahead or the flight deviates from the safe envelope, a dynamic replanning mechanism is triggered, in which the carrier UAV hovers and the reconnaissance UAV searches for an alternative path.

[0028] The dynamic path planning method and system for multi-UAV collaborative delivery proposed in this embodiment are designed for low-altitude logistics delivery within urban built-up areas, specifically at night (preferably from 10:00 PM to 5:00 AM the following day). This specific scenario was chosen based on the following technical and practical considerations, which also constitute the design basis for the technical solution of this invention: At night, the number of moving targets such as vehicles and pedestrians on the urban ground decreases significantly, providing a more stable and predictable ground environment for drone delivery and reducing the risk of falling objects caused by moving targets below. However, the risk is not completely eliminated, and it is still necessary to accurately identify and avoid occasional nighttime vehicles and personnel.

[0029] Nighttime ambient lighting conditions are poor and light distribution is extremely uneven. Urban areas contain areas illuminated by streetlights, areas in building shadows, and completely dark areas, posing a significant challenge to traditional drones that rely on visible light sensors. Furthermore, nighttime weather conditions such as fog and low clouds are more likely to form, further reducing visibility.

[0030] At night, background noise levels drop significantly, making the aerodynamic noise generated by drones during flight particularly prominent and potentially disrupting the peace and quiet of residential areas. Therefore, noise impact must be considered a key constraint factor in path planning.

[0031] The low-altitude airspace of cities (referring to the airspace below 120 meters above the ground) is filled with static obstacles such as buildings, power lines, communication towers, and billboards. These obstacles become even more difficult to detect at night due to the lack of sufficient markings.

[0032] In the above application scenarios, this embodiment provides a complete system solution that can overcome the challenges of nighttime perception, proactively manage security risks, and take into account the impact on the social environment.

[0033] To achieve the above objectives, the dynamic path planning system for multi-UAV collaborative delivery disclosed in this embodiment consists of three main parts: the air terminal, the ground terminal, and the communication link.

[0034] The airborne segment, the core component that performs actual flight and perception missions, consists of at least one reconnaissance drone and one carrier drone in a heterogeneous configuration.

[0035] The ground-based system, serving as the command and computing center, consists of a collaborative task management platform deployed on servers or high-performance computers.

[0036] The communication link is responsible for establishing a stable, low-latency data and command exchange channel between the air and ground ends, as well as between different UAVs within the air end.

[0037] The air terminal is the key execution unit of this invention, and its heterogeneous hardware configuration is the physical basis for realizing the method of this invention.

[0038] Reconnaissance drones play the role of pathfinders and environmental modelers in the system, and their core design principles are lightweight, high perception, and high mobility.

[0039] In a preferred embodiment, the reconnaissance drone is primarily a quadcopter. The arms and fuselage structure should be constructed of carbon fiber composite materials, and the unloaded takeoff weight (including mission payload) should be controlled below 1.5 kg to ensure that even in the event of a malfunction under extreme conditions, its fall kinetic energy remains low. The reconnaissance drone should be equipped with four T-Motor F40 ProIII brushless motors, matched with APC 10x4.5MR low-noise propellers. This power combination ensures maneuverability while reducing sound pressure levels during flight through optimized aerodynamic design. The reconnaissance drone should integrate an NVIDIA Jetson AGXOrin developer kit as an onboard AI computer. This unit is responsible for receiving and processing multi-sensor data in real time, performing high-intensity sensor data fusion, environmental recognition, risk assessment, and turbulence field calculation algorithms. Flight control is handled by a CUAV Nora+ flight control hardware, running PX4 open-source firmware, and communicating with the Jetson AGX Orin via high-speed serial port.

[0040] The reconnaissance drone's payload is integrated into a three-axis stabilized gimbal, which includes the following three sensors: The visible light camera, preferably an industrial camera module using a Sony IMX586 sensor and a fixed-focus lens with an F / 1.6 aperture, is used to acquire color images of street-lit areas in high-sensitivity mode, providing the system with semantic information such as road, lane markings, and vehicle colors.

[0041] The infrared thermal imaging camera, preferably employing a FLIR Boson infrared sensor, has a resolution of 640x512 pixels and a thermal sensitivity of less than 40mK. This camera is unaffected by any lighting conditions and is specifically designed for accurately detecting and identifying targets with thermal signatures, such as people, animals, and moving vehicles, in completely dark areas.

[0042] The preferred lidar is the Livox Avia hybrid solid-state lidar. It features a 360° horizontal field of view and a 70° vertical field of view, a ranging capability of up to 190 meters, and can generate 240,000 points per second. Its core function is to perform 3D environmental structure modeling unaffected by lighting conditions, accurately detecting building outlines and tree branches, and exhibiting a high detection rate, especially for small obstacles such as overhead power lines and cables that are extremely difficult for traditional cameras to identify.

[0043] The reconnaissance drone is equipped with a dual-antenna RTK-GNSS module, enabling centimeter-level differential positioning in open urban areas. It also integrates a communication module supporting 5G NR networks for high-bandwidth, low-latency data exchange with ground control and the drone itself.

[0044] The core design principles of unmanned aerial vehicles (UAVs) in the system are high load capacity, high stability, and high reliability.

[0045] In a preferred embodiment, a hexacopter UAV platform is selected. The hexacopter structure provides higher power redundancy and wind resistance stability. It is designed with a maximum payload of 5 kg and is equipped with a standardized cargo mounting and automatic release device under the fuselage. It employs six high-power disk brushless motors, matched with 18-inch carbon fiber folding propellers, to ensure ample power even under full load. Its onboard computing requirements are lower than those of reconnaissance UAVs, and it is equipped with a standard flight controller responsible for receiving commands, executing flight missions, and maintaining its own attitude stability. It is equipped with basic millimeter-wave radar for autonomous obstacle avoidance in emergencies, and a standard RTK-GNSS module for precise positioning. Its perception capabilities serve only as safety redundancy and do not participate in active path planning decisions. A 5G NR communication module is also integrated to ensure stable reception of flight commands issued by reconnaissance UAVs or ground control.

[0046] The collaborative task management platform is deployed on a cloud server or local workstation. Its software system includes: a task order management module, a high-precision 3D city map engine, a global initial path planning module, and a dual-drone collaborative flight monitoring and data visualization interface. This platform is responsible for planning a macroscopic reference path based on a high-precision map before the task begins, and for monitoring and managing the status of the two drones during task execution, but it does not directly participate in the segmented dynamic path generation.

[0047] This embodiment preferably employs a private network communication scheme based on 5G slicing. An independent network slice with Quality of Service (QoS) is allocated to this system through the operator's network to ensure ultra-low latency (end-to-end less than 20ms) and ultra-high bandwidth (uplink greater than 100Mbps) of the communication link. This is crucial for the real-time transmission of multiple high-definition sensor data from the reconnaissance drone and for the system to issue emergency control commands.

[0048] Next, this embodiment details the initial stage of the dynamic path planning method for multi-UAV collaborative delivery, namely task initialization and formation deployment. The task initialization and formation deployment are initiated by the collaborative task management platform and completed through the coordinated actions of the reconnaissance UAV and the carrier UAV. Specifically, this includes: S2.1 Task Reception and Parsing: The collaborative task management platform receives delivery task instructions from external logistics systems. This instruction is a structured data packet containing at least the following information: a unique task identifier, and cargo information (preferably, including weight). (and dimensions), three-dimensional coordinates of the takeoff point Target landing point three-dimensional coordinates The collaborative task management platform analyzes task information, especially based on cargo weight. Verify whether it is within the defined safe payload range of the transport drone.

[0049] S2.2 Global Reference Flight Route Planning: Before the UAV takes off, the collaborative mission management platform plans a flight route from... arrive Global reference route This route is not the final flight path, but a macroscopic, static, optimal guiding path.

[0050] The urban transportation network is modeled as a directed graph. , where the set of nodes Representing key waypoints, edge sets This represents a feasible flight segment connecting waypoints. Before performing a path search, the system first analyzes the graph. Pre-processing is performed to eliminate all flight segments that violate regulations. This includes any flight segments that cross known no-fly zones or fly below the legally mandated minimum safe altitude. All will be from the edge set Remove from the map and generate a submap containing only compliant flight segments. .

[0051] In the compliance subgraph The optimal path is found using the A* search algorithm. The toll cost function is a normalized weighted sum. For any compliant flight segment... Its dimensionless overall travel cost Defined as: in, and The preset weighting coefficients are used, and they satisfy the following conditions: These two coefficients represent the decision-maker's preference for flight efficiency and flight safety risks. In the nighttime scenario of this embodiment, safety and low social impact are primary considerations, therefore, the coefficients can be set... , .

[0052] The two core cost terms in the function and Defined respectively as normalized distance cost and normalized static risk cost. Normalized distance cost : Represents the relative level of segment length among all possible segments, calculated using the following formula: in, Indicates flight segment Euclidean distance, in meters. It is a picture The maximum possible distance across all flight segments, i.e. .

[0053] Normalized static risk cost This represents the average relative static risk level when flying over a certain flight segment, calculated using the following formula: in, It is a static risk potential field function defined on a map, representing the point of flight at night. A quantitative value (0-100 score) of the potential risks or impacts when flying over sensitive buildings such as hospitals and schools. This function is pre-calculated based on the semantic information of the map, and is particularly effective when flying over such buildings. Take the highest value. When flying over parks, rivers, or overhead, Take the lower value. The integral of this item is the flight segment. The total accumulated static risk.

[0054] It is the maximum possible value of the risk potential field function, corresponding to the most sensitive areas such as the airspace above the intensive care unit of a hospital. This also normalizes the risk value at any point to Interval. The factor makes the final result the average normalized risk per unit distance of the flight segment, thereby eliminating the influence of the flight segment length itself on the risk cost calculation, making it a dimensionless indicator that purely reflects the path risk density.

[0055] The collaborative task management platform ultimately adopts the A* search algorithm, based on its dimensionless comprehensive cost of travel. For measurement, in the graph Search from arrive The minimum cost path is used to obtain the global reference route. .

[0056] S2.3 Coordinated Formation Takeoff and Deployment Obtain global reference route Subsequently, the system commands the airborne drones to take off and establish precise forward and backward coordinated formations.

[0057] S2.3.1 The carrier drone takes off, vertically taking off to the preset safe altitude. (Preferred, 30 meters above the ground), and at the takeoff point It hovers stably directly above it. Its current 3D position vector is denoted as... .

[0058] S2.3.2 The reconnaissance drone then took off, its mission being to fly to the front of the carrier drone, located at... The predetermined leading position on the flight path. This target position is dynamically calculated to ensure precise formation relationships. The instantaneous target position vector of the reconnaissance UAV. The definition is as follows: in, Is it a carrier drone at any time The real-time three-dimensional position vector. It is a projection operator representing the current position of the drone to be carried. Vertical projection to global reference path The three-dimensional position vectors of the points on the map ensure that the formation always unfolds along the macroscopic path.

[0059] This represents a preset safe leading distance. Setting this distance requires balancing two factors: a sufficiently large distance to provide ample reaction time for the following carrier drone; and a sufficiently small distance to ensure that the environmental information detected by the reconnaissance drone remains valid when the carrier drone arrives. In this embodiment, the distance is set... rice.

[0060] It is a global reference route. At the projection point The unit tangent vector at that point, i.e., the direction of the flight path. The reconnaissance UAV calculates this based on its flight control system. It will autonomously fly to the target location.

[0061] S2.3.3 When the actual location of the reconnaissance drone With the target location The condition is met once the error remains within the allowable range for a period of time: (in For position tolerance, preferably 0.5 meters) for a duration exceeding (Preferred time: 3 seconds) The collaborative task management platform confirms that the initial collaborative formation deployment is complete.

[0062] The following embodiment details the preliminary environmental perception phase in the dynamic path planning method for multi-UAV collaborative delivery. After the collaborative formation is deployed, the reconnaissance UAV at the front of the formation performs this phase. Before entering the next area to be explored, it uses its onboard multimodal sensing payload to comprehensively collect raw environmental data.

[0063] 3.1 Multimodal sensor data acquisition After the coordinated formation deployment is completed, the reconnaissance drone will be at its current position (i.e., the distance in front of the carrier drone). The system scans the airspace ahead for exploration and simultaneously activates its three onboard sensors to collect data. All sensors are mounted on a three-axis stabilization gimbal to ensure attitude stability during flight, thereby obtaining high-quality raw data. All collected data includes precise timestamps. and corresponding UAV global positioning information This is to achieve subsequent spatiotemporal alignment.

[0064] 3.1.1 Visible Light Image Acquisition The reconnaissance drone is equipped with a visible light camera (preferably using a Sony IMX586 sensor) to collect visible spectrum information of the environment. Considering the uneven lighting conditions at night, the camera's exposure parameters employ an adaptive adjustment strategy. Upon entering a new acquisition location, the system first quickly adjusts the exposure time based on the average brightness of the central image area. and gain This is to avoid overexposure or underexposure of the image. Each frame of image data acquired can be represented as a three-dimensional tensor: in, Indicates at time The captured single-frame RGB image, and These represent the height and width of the image, respectively, in pixels.

[0065] 3.1.2 Infrared Thermal Imaging Data Acquisition The long-wave infrared thermal imaging camera (preferably using a FLIR Boson infrared sensor) carried by the reconnaissance drone is responsible for collecting environmental thermal radiation information. This process is unaffected by any lighting conditions and is a key means of detecting vital signs and active heat sources at night. Each frame of thermal imaging data acquired can be represented as a two-dimensional matrix: in: Indicates at time A single-frame long-wave infrared image was acquired. and These represent the height and width of the image, respectively, in pixels. Each pixel value in the matrix represents the intensity of infrared energy radiated by the object in the corresponding field of view, which can be approximately converted into a temperature value after calibration.

[0066] The data Specifically designed for the precise detection and location of targets with distinct thermal signatures, such as occasional pedestrians or animals on the ground, or vehicles that have just been turned off and are still warm, under any lighting conditions, especially in completely dark areas.

[0067] 3.1.3 3D LiDAR Point Cloud Acquisition The reconnaissance drone is equipped with a lidar system (preferably Livox Avia) responsible for acquiring three-dimensional spatial structure information of the environment. The lidar operates at a fixed scanning frequency. (Preferred, 10Hz) A laser beam is emitted outward and the reflected signal is received. The distance is determined by calculating the laser's time of flight. This occurs within a complete scan cycle. Within a given frame, the lidar generates a single frame containing a large collection of spatial points, i.e., a point cloud. At time [time value missing]... A single frame of point cloud data can be represented as: in, Indicates at time The collected single-frame point cloud data. It represents the total number of spatial points in the point cloud of that frame. It is a single point in the point cloud; it is a four-dimensional vector. It is the three-dimensional spatial coordinate of the point in the UAV's body coordinate system. It is the intensity of the laser beam reflected at that point.

[0068] After data acquisition, the onboard computing unit will immediately utilize the real-time pose information (including position) of the UAV provided by RTK-GNSS. and attitude quaternions ), the point cloud in the body coordinate system Convert to the global geographic coordinate system to obtain the global point cloud. .

[0069] Global point cloud Its core function is to perform high-precision 3D environmental structural modeling unaffected by lighting conditions. It can accurately depict the outlines of buildings and the branches of trees, and it has an extremely high detection rate for small, temperature-neutral obstacles such as overhead power lines and cables that are completely undetectable by visible light and infrared cameras.

[0070] During the data acquisition phase, the reconnaissance drone, through its heterogeneous sensor payload, acquired visible light information about the airspace to be explored ahead in parallel. ), thermal characteristic distribution information ( ) and three-dimensional spatial structure information ( ).

[0071] Next, this embodiment elaborates on the data fusion and risk map generation stages in the multimodal data fusion of the dynamic path planning method for multi-drone collaborative delivery.

[0072] To achieve deep fusion and accurate risk assessment of multimodal heterogeneous data (visible light, infrared, and lidar) in nighttime environments, this embodiment proposes an interactive fusion network (HIF-Net). Instead of performing a one-time parallel processing and fusion of the three modalities, the interactive fusion network uses a multi-stage hierarchical encoder to repeatedly guide and aggregate feature streams representing different physical information at multiple scales. Finally, long-distance dependency modeling is used in the risk synthesis path to refine features, thereby generating a high-precision dynamic risk map.

[0073] HIF-Net includes: an input preprocessing and information supplementation module, a hierarchical interactive encoder consisting of five stages (Stage 1 to Stage 5), and a risk synthesis path.

[0074] Input preprocessing and information supplementation Before the collected data is fed into the neural network, it needs to be preprocessed and supplemented with prior map information.

[0075] LiDAR point cloud rasterization: This involves rasterizing the acquired 3D point cloud in the global geographic coordinate system. It is converted into a multi-channel two-dimensional raster image through bird's-eye view (BEV) projection. .

[0076] Specifically, to reconnoiter the current location of the drone Define a rectangular area on the ground centered on the target (preferably). ), and divide it into fine grids (preferably, with a resolution of ), ).Will Each point in It falls into the corresponding grid, and multiple channels are generated based on statistical characteristics. Preferably, The three channels are encoded as follows: Height channel: the maximum height value of all points within the grid; Intensity channel: the average laser reflection intensity of all points within the grid. Density channel: The number of points within a grid.

[0077] This converts three-dimensional geometric information into a two-dimensional image format.

[0078] Additional map prior information: based on the current position of the reconnaissance drone. The map prior information for the corresponding BEV area is extracted from the high-precision 3D city map database stored onboard, and rendered as a multi-channel raster map. Preferred, The four channels are encoded using binary masks: Building Area Channel: Identifies the area occupied by a building. Road / Lane Line Channel: Identifies the area for motor vehicles. Sidewalk / Green Belt Channel: Identifies areas where pedestrian activity is generally permitted or where it is relatively safe. Static Obstacle Channel: Identifies the locations of small, fixed obstacles such as lampposts and fire hydrants.

[0079] As a form of strong prior knowledge, it assists the network in understanding the static structure of a scene. Real-time sensor data provides information about the current situation, but lacks prior knowledge about what should be present. High-precision maps serve as the carrier of this prior knowledge. At night, when sensor data is blurry (e.g., an indistinguishable dark area), map information can provide strong constraints for the network, informing it whether it is a park lawn or a driveway. Stitching map information as an additional channel with sensor images is the most direct and effective way to achieve the fusion of data and prior knowledge at the input layer.

[0080] The hierarchical interactive encoder contains three functionally decoupled parallel branches and performs deep interaction in stages 3, 4, and 5.

[0081] Perceiving the urban nighttime environment is a complex task fraught with severe information asymmetry and uncertainty. Visible light information is lacking, infrared information lacks structure, and lidar information lacks semantics. Traditional parallel extraction-single-step fusion architectures are prone to contaminating the final decision due to limitations in a particular modality of data (such as noise in dark areas of RGB images) or misjudgments. Hierarchical interactive fusion decomposes the fusion process into multiple different levels of abstraction (different depths of the network), allowing the network to fuse low-level features (such as edges and contours) in the early stages, mid-level features (such as object components) in the mid-stages, and high-level features (such as object categories and motion trends) in the deep stages. This progressive, shallow-to-deep fusion approach aligns better with the logic of intelligent agents cognizing complex phenomena, enabling the construction of a more robust and comprehensive understanding of the environment.

[0082] Initial feature extraction (Stage 1-2): The hierarchical interactive encoder includes a geometric branch that receives a LiDAR bird's-eye view. As input, downsampling is performed only once in the first two stages (Stage 1, 2) to preserve accurate spatial geometric details, such as the precise edges of obstacles, to the maximum extent possible.

[0083] The hierarchical interactive encoder also includes a visual semantic branch, which receives the visible light image stitched along the channel dimension and the map prior. As input, multiple downsampling and dilated convolutions are performed in the first two stages to quickly extract high-level scene semantic information.

[0084] The hierarchical interactive encoder also includes a dynamic thermodynamic branch, which serves as the main branch and receives the stitched infrared thermal image and prior map data. It employs a standard backbone network structure (preferably using the initial layers of ResNet) to extract the most reliable dynamic heat source information and contextual relationships in nighttime scenes.

[0085] After Stage 2, the three branches output feature maps with different scales and semantic levels, denoted as... , , .

[0086] Multi-stage feature interaction and fusion (Stages 3, 4, 5): Starting from Stage 3, each of the three branches interacts with information at the end of each stage through a gated cross-attention fusion module. (The text then repeats itself, so the translation stops.) Each stage ( For example, the process is as follows: S4.a.1 Feature scale alignment, aligning the first... At the start of the phase, feature maps are extracted independently by the three branches. , , Unify to the feature map scale of the main branches. This involves... downsampling and The upsampling operation. The aligned feature map is denoted as... and .

[0087] S4.a.2 GCF Module Fusion: The GCF module uses dynamic thermodynamic features as a gating mechanism to dynamically adjust the fusion weights of geometric and semantic features.

[0088] First, generate the gating signal. : This gating signal This reflects which regions contain dynamic heat sources requiring high attention at the current scale. Subsequently, this gating signal was used to guide the fusion of two other modalities: in, It represents a standard processing unit that includes convolution (Conv), batch normalization (BatchNorm), and activation function (ReLU). This indicates element-wise multiplication. This indicates element-wise addition.

[0089] In areas with significant heat sources ( (Higher values) result in a fusion outcome that prioritizes high-precision geometric features; in the background region ( If the value is low, it is more focused on the semantic features of the scene.

[0090] Simple feature addition or concatenation fails to capture the varying importance of different information in different scenarios. The GCF module utilizes the most reliable thermal imaging features to generate an attention gating mechanism, essentially a dynamic, content-based feature selection mechanism. It teaches the network that when the thermal imaging branch indicates the presence of a pedestrian, the fusion unit assigns higher weights to geometric features (LiDAR) that accurately describe the pedestrian's outline; conversely, in open background areas, it can place greater trust in the visual semantic branch's overall assessment of the scene.

[0091] S4.a.3 Feature information back-annotation: Features obtained after fusion It's not just the output of this stage; it's also injected back into the main branch as the... The input to the stage dynamic thermodynamic branch. The back-injection process enables the main branch to fully utilize the knowledge that has been integrated with the other two modal information in subsequent feature extraction, thereby achieving progressive information enhancement.

[0092] After layers of interaction in Stages 3, 4, and 5, the encoder finally outputs a set of deeply fused feature maps from different scales. .

[0093] By repeatedly verifying and fusing information from different sensors across multiple stages, the probability of overall judgment errors caused by the failure of a single sensor under specific conditions (such as RGB overexposure due to strong car headlights or infrared missed detection due to low-temperature objects) is greatly reduced. This is crucial for the safety of cargo-carrying drones. The backfeeding and re-fusion of information across layers allows the network to use the fusion results of one stage to guide feature extraction in the next stage, essentially enabling the network to continuously self-correct and focus, thereby obtaining a more refined and accurate feature representation than a single fusion.

[0094] The risk synthesis path and feature purification: the multi-scale fusion features extracted by the encoder need to be gradually upsampled and aggregated through a path similar to the decoder, and finally generate a risk map with the same resolution as the input.

[0095] Stepwise upsampling and feature aggregation: Risk synthesis path starts from the output of the deepest layer of the encoder Initially, spatial resolution is gradually restored through an upsampling module. After each upsampling, the data is concatenated with the fused features from the corresponding stage of the encoder.

[0096] To address the feature redundancy issue caused by simple splicing and enhance the network's understanding of the global scene, before each splicing operation, the features (i.e., features passed from the encoder) are re-evaluated. and The feature map first needs to be refined using a Spatial Long-Range Correlation (SLC) module. The role of the SLC module is to dynamically establish long-range dependencies between different spatial locations in the feature map.

[0097] The SLC operator is internally implemented as an efficient sparse self-attention computation based on region partitioning and routing mechanisms. It computes correlations at a coarse-grained region level on the input feature map, filtering out a large number of irrelevant regions. Then, it performs fine-grained, pixel-by-pixel attention computation between the few highly correlated regions. The specific process includes... S4.b.1: Region Division and Feature Projection Given a feature map passed from the encoder skip connection (For example ), dividing it spatially into There are 3 non-overlapping regions, each with a size of 1. .therefore, .

[0098] The input feature map is processed through three independent linear mapping layers. Projection generates three tensors: query, key, and value, denoted as . : in, It is a projection matrix.

[0099] S4.b.2: Construction of Regional Sparse Routing Graphs To avoid the enormous computational cost of calculating attention globally pixel by pixel, a sparse connection relationship is constructed at the region level. The pixel features within each region are then averaged using pooling to obtain region-level query and key representations. .

[0100] Calculate the correlation matrix between regions : matrix Each element in Represents the first The query range and the first Semantic similarity of key regions.

[0101] Next, the correlation matrix Perform a top-k operation on each row to filter out the best results for each query range. The most relevant key regions. This process generates a sparse routing index graph. : in, The operator returns the largest value in each row. Column index of each element. This is a hyperparameter representing route sparsity, much smaller than the total number of regions. (Preferred) Sparsity operations enable attention computation to focus on a few highly correlated regions (e.g., a vehicle region and several regions along its path ahead), while ignoring a large number of irrelevant background regions (e.g., building walls), thereby greatly reducing the computational load on the airborne computing unit while maintaining the ability to model long-range dependencies.

[0102] S4.b.3 Key Information Aggregation Based on Routing Graph Using the generated routing index graph From the original, pixel-by-pixel key Sum In the tensor, a sparse set of keys and values ​​is aggregated for each query region. This process can be implemented using a Gather operation: After the Gather operation, the attention calculation for each query pixel is no longer limited to the entire feature map, but is instead confined to the area specified by the routing graph. The pixels within a region enable dynamic sparse attention.

[0103] S4.b.4 Fine-grained attention calculation and local information enhancement Standard multi-head self-attention computation is performed on top of the sparse keys and values. For simplicity, a single-head example is used here: in It is the dimension of the key vector, used for scaling.

[0104] Simultaneously, to compensate for the local detail information that may be lost due to sparse attention, a Local Convolution Enhancement (LCA) pathway is set up in parallel. This pathway enhances the original value tensor. Perform a depthwise separable convolution to efficiently capture the local spatial context: Final output of the SLC module It is the sum of the results of fine-grained attention calculation and local convolution enhancement: Final output (Right now This refers to a refined feature map that contains long-range spatial dependency information. It is fed into the generation of the risk synthesis path.

[0105] Dynamic Risk Map Generation: The final output of the risk synthesis path is a high-dimensional feature map with the same resolution as the input. This feature map is then used to generate a header (composed of a...) from the risk map. The system consists of a convolutional layer and a sigmoid activation function, generating the final two-dimensional dynamic risk weight map. .

[0106] Next, this embodiment elaborates on the spatial risk assessment stage in the dynamic path planning method for multi-UAV collaborative delivery. This stage is executed in parallel with the ground risk assessment and is also handled by the onboard computing unit of the reconnaissance UAV. Through indirect extrapolation, the flight status data of the reconnaissance UAV is analyzed in real time to calculate the microclimate and turbulence field intensity of its current spatial location, generating quantified three-dimensional spatial airflow risk data.

[0107] Theoretical basis and data sources for space risk assessment The space risk assessment method proposed in this embodiment is based on the fact that, in order to maintain its preset attitude (such as hovering or uniform straight flight), the flight control system of an unmanned aerial vehicle (UAV) must continuously resist aerodynamic disturbances from the external environment. These disturbances are mainly caused by wind fields (microclimates) and turbulence. Therefore, the control efforts made by the flight control system to maintain stability, as well as the minute attitude changes of the fuselage under disturbances, directly reflect the characteristics of the external airflow environment.

[0108] This method does not add any additional weather sensors. Instead, it utilizes the existing high-frequency output data source within the reconnaissance UAV's flight control system: flight motor control commands are sent from the flight controller (preferably CUAV Nora+) to each motor speed controller via pulse-width modulation or digital signals. These commands determine the rotational speed of each rotor, and their update frequency is extremely high (preferably consistent with the flight controller's main cycle frequency). Inertial measurement unit (IMU) data consists of the fuselage's three-axis angular velocity and three-axis linear acceleration measured by onboard IMU sensors. This data is also recorded at a high frequency (preferably up to 1 kHz).

[0109] The system synchronously samples and analyzes the two types of data at a frequency of 100Hz.

[0110] Turbulence intensity solution based on control moment analysis The overall intensity of turbulence is quantified by analyzing the differential control torque output by the flight control system to counteract airflow disturbances. For a quadcopter reconnaissance UAV, the lift generated by its four motors is as follows: The flight control system generates the desired total lift by changing the magnitude of these four lift forces. And the control torques around the fuselage along the x, y, and z axes. .

[0111] In ideal still air, the outputs of the four motors should be approximately equal and constant to maintain hovering. When turbulence is present, the flight control system will adjust the outputs of each motor at high frequency and drastically to maintain attitude, resulting in violent fluctuations in control torque.

[0112] S5.a.1 Control torque signal extraction, at time... The system records the normalized throttle commands from the four motors. ( The lift generated by a single motor. and its throttle command Forming a quadratic relationship: in, It is the lift coefficient of the rotor ( ), It is the maximum angular velocity of the motor ( ), The equivalent coefficient (m) is used to characterize the proportional relationship between the anti-torsional torque and the lift of the motor.

[0113] Control torques around the fuselage x-axis (roll), y-axis (pitch), and z-axis (yaw) Lifting force can be generated by each motor and its lever arm vector relative to the fuselage center of gravity The calculation is as follows. For a standard "X"-shaped quadrotor configuration, the calculation formula is: in, It is the distance from the motor shaft to the center of the machine body. It is the motor's counter-torque coefficient.

[0114] S5.a.2 Definition and Calculation of Turbulence Intensity Index This embodiment defines a turbulence intensity index (TII). This index is used to quantify the intensity of turbulence at a current spatial point. It is based on the total change or energy of the control torque signal over a time window. Specifically, the system operates within a sliding time window. (Preferred) Within seconds, calculate the control torque vector. Total variance: in, It is a time window The average vector of the internal control torque, i.e. .

[0115] The larger the value, the more violent the fluctuation of the control torque exerted by the flight control system to maintain attitude stability, thus indirectly proving that the turbulence intensity at that spatial location is higher.

[0116] Wind field direction and characteristics calculation based on IMU data analysis This method analyzes specific frequency components in IMU data to further infer the direction and characteristics of the wind field (such as whether it is a continuous crosswind or a high-frequency random disturbance).

[0117] Step 5.b.1: Attitude error signal extraction Even with flight control compensation, strong airflow can still cause minor attitude errors on the aircraft. The IMU measures the three-axis angular velocities. Angular velocity as desired by the flight control system The difference between them is the angular velocity error. This error vector is mainly caused by external airflow disturbances.

[0118] Step 5.b.2: Define wind field direction characteristics and calculate This embodiment defines a wind field direction feature (WDF) vector. This feature is used to characterize the dominant direction of persistent wind fields. It is extracted by low-pass filtering the angular velocity error signal. Persistent crosswinds generate a persistent, low-frequency angular velocity error on the corresponding fuselage axis (e.g., the roll axis).

[0119] in, For a low-pass filter (preferably a second-order Butterworth filter), its cutoff frequency is... Set it to a lower (preferred) Hz).

[0120] It should be noted that, Units are In this embodiment, The direction of the vector approximately reflects the direction of the continuous aerodynamic torque acting on the UAV, thus effectively characterizing the prevailing wind direction. For example, a significantly positive vector... The x-component indicates the presence of a continuous wind blowing from right to left.

[0121] Step 5.b.3: Calculate the energy of high-frequency disturbances To distinguish between persistent wind and high-frequency turbulence, the system also performs high-pass filtering on the angular velocity error signal and calculates its energy.

[0122] in, For a high-pass filter, its cutoff frequency is... Set to a higher (preferred) Hz). It represents the energy of high-frequency disturbances. The higher the value, the more unstable the ambient airflow is and the more small-scale eddies there are.

[0123] Generate three-dimensional spatial airflow risk data At any moment Reconnaissance drones in position At that location, a data packet containing airflow risk information about that spatial point is generated. This data packet contains at least the following: Turbulence Intensity Index This represents the intensity of the overall airflow disturbance. Wind field directional characteristics. Represents the dominant direction and intensity of the persistent wind field. High-frequency disturbance energy. It represents turbulence and instability in airflow.

[0124] Next, this embodiment elaborates on the acoustic impact assessment stage in the dynamic path planning method for multi-UAV collaborative delivery. This stage is performed in parallel with the ground risk assessment and space risk assessment, and is also handled by the onboard computing unit of the reconnaissance UAV.

[0125] The acoustic impact assessment method proposed in this embodiment follows the analysis framework of sound source-propagation path-receiving point. The entire process is decomposed into four modules: UAV sound source modeling, environmental background noise perception, sound propagation path attenuation calculation, and the final generation of the Acoustic Impact Index (AII) and impact map. This method aims to calculate the actual auditory impact of UAV flight on preset noise-sensitive points, rather than just the noise level of the UAV itself.

[0126] The noise of a drone is not a constant value; it is closely related to its flight state (especially motor speed). Therefore, it is necessary to establish a model that can characterize its noise generation capability (i.e., sound power).

[0127] Preferably, during the UAV design phase, its noise characteristics under different throttle commands have been calibrated in an anechoic chamber. (Sound power level of the reconnaissance UAV) (Unit: dB) can be modeled as a function of the average normalized motor command. Functions: in, It is for performing a specific flight maneuver (such as hovering, speed) During flight, the flight control system outputs the average normalized throttle command to the four motors, with a value range of [value range missing]. . These are the coefficients of the calibrated UAV acoustic model, where and The unit is dB. It is a dimensionless quantity.

[0128] Furthermore, the noise radiation from a drone is not omnidirectionally uniform. This embodiment introduces a directional factor. (Unit: dB), used to describe different directional angles The difference in sound pressure level.

[0129] Meanwhile, the impact of noise is relative. In a noisy environment, the sound of a drone is easily masked; however, in a quiet night, the same sound will be extremely jarring. Therefore, the system needs to perceive the background noise level of the current environment in real time.

[0130] Preferably, at the start of the mission or during each hovering perception phase, the reconnaissance UAV uses its onboard miniature microphone to sample ambient noise at the moment its motors stop or idle. The collected sound pressure signal is filtered through an A-weighted network to simulate the frequency response characteristics of the human ear. The final A-weighted background noise sound pressure level of the current environment is then obtained. (Unit: dBA).

[0131] From the sound source (reconnaissance drone) to any receiving point (e.g., a residential building window) During sound propagation, sound energy attenuates due to various factors. In complex urban environments, the main attenuation factors include: geometrical diffusion attenuation (GDA). ): The decrease in energy density caused by the diffusion of sound waves in three-dimensional space.

[0132] in, It is the straight-line distance between the drone and the window. For the reference distance, constant term 11 is the spherical divergence constant of the point sound source, i.e. .

[0133] Atmospheric absorption attenuation ( Sound energy absorption caused by factors such as the viscosity of air molecules is particularly noticeable for high-frequency noise.

[0134] in, It is the atmospheric sound absorption coefficient (unit: dB / m). This value is related to air temperature and humidity. It can be obtained by referring to the table built into the ISO 9613-1 standard after the airborne temperature and humidity sensor provides the data.

[0135] Obstacle diffraction attenuation ( When there are obstacles such as buildings between the drone and the receiving point, the sound waves will be blocked.

[0136] The system utilizes an onboard high-precision 3D city map database and real-time LiDAR point cloud data to determine the drone's location. With any window position Is there a direct sound path between them?

[0137] If LOS exists, then .

[0138] If there is no path of sight (LOS) (i.e., the path is blocked by a building), sound diffraction will occur. This embodiment uses a simplified model based on path difference, setting a fixed, relatively high attenuation value. Preferably, dBA represents an effective sound barrier.

[0139] Generate acoustic impact index and noise impact map Calculate at time Reconnaissance drones in position , in state During flight, focus on specific windows The A-weighted sound pressure level generated at the location : The level of intrusion felt by residents is not the absolute sound pressure level, but rather the degree to which the drone noise exceeds the background noise. Therefore, this embodiment defines an acoustic impact index. The calculation formula is as follows: The physical meaning of this value is the number of decibels by which the drone noise exceeds the background noise at the receiving point. If the value is 0, it means that the drone noise is completely masked by the background noise and has no impact; if the value is 10, it means that the noise is significantly more than 10 dBA than the background noise and may attract attention.

[0140] A noise impact map is generated based on the above model. include: From a high-precision 3D map, all building facades marked as residential within a certain radius (preferably 150 meters) around the reconnaissance drone are identified, and the center points of windows on these facades are extracted to form a set of noise-sensitive points. .

[0141] For each candidate flight location to be evaluated Calculate the position relative to all sensitive points. The generated value.

[0142] Candidate flight positions The two-dimensional ground grid in the noise impact map The value in is set to the maximum value that will affect all windows: in, express The projection position on a two-dimensional raster map.

[0143] The final noise impact map Each grid cell represents the maximum potential noise disturbance level that a drone flying over that grid cell would cause to nearby residents.

[0144] Next, this embodiment details the process of generating an instant safe flight corridor in the dynamic path planning method for multi-UAV collaborative delivery. Multi-source heterogeneous risk information is normalized and fused to construct a unified multi-dimensional risk cost field. Based on this, a time-limited, fully safety-verified three-dimensional flight space (i.e., a "safety bubble cube") and the next flight path instructions are planned and generated in real time for the rear transport UAVs.

[0145] The system is monitoring the current location of the reconnaissance drone. At that location, the spatial airflow risk data packet (containing...) is calculated in real time. , , To integrate it with other two-dimensional risk maps, these point-like, time-series data must be expanded into a two-dimensional spatial risk map.

[0146] S7.a.1: Three-dimensional airflow risk voxel mapping The reconnaissance drone maintains a dynamically updated 3D airflow risk voxel grid centered on itself in memory. This grid covers the airspace to be explored ahead of it. As the reconnaissance drone flies, it continuously updates its position at each location. ( The calculated spatial airflow risk data is then populated into the voxels along the trajectory. Each voxel stores a comprehensive airflow risk value. Preferred, It is a weighted sum of the turbulence intensity index and the energy of high-frequency disturbances: in, and These are the calibrated turbulence intensity index and the maximum reference upper bound of high-frequency disturbance energy, respectively. and The weighting coefficients are preferably satisfied. .

[0147] For the area ahead where the reconnaissance drone has not yet arrived, it is assumed that the wind field has a certain continuity within a short distance (preferably 10-20 meters). The calculated wind field direction characteristics are then utilized. The airflow risk value at the current location. It propagates forward along the direction of the wind field, accompanied by a certain degree of attenuation.

[0148] S7.b.2: Generate a two-dimensional spatial airflow risk map. To integrate it with other two-dimensional maps, the three-dimensional airflow risk voxel grid needs to be projected into a two-dimensional spatial airflow risk map. Considering that the transport drone must safely traverse the entire vertical space above its path during flight, this embodiment employs a maximum projection strategy: in, It represents a coordinate on a two-dimensional raster map. The operator represents the maximum airflow risk value among all voxels (i.e., a vertical column) above the two-dimensional coordinate. It ensures that the risk value of any point on the two-dimensional graph represents the most dangerous airflow situation in the entire airspace column above it, which is a conservative and safe design strategy.

[0149] To integrate them into a unified cost map, it is necessary to... and Dimensionless processing is performed. Preferably, the maximum value normalization method is used: in, and It is a predictable upper limit value used to normalize the values ​​of the two graphs to an approximate value. interval The three normalized risk maps are then combined into a final comprehensive cost map through a weighted summation. : in, These are weighting coefficients representing the importance of different risk sources, and the sum of the three is 1. In the nighttime delivery scenario of this invention, ground safety (falling object prevention) is the highest priority, followed by flight stability, and finally noise impact. Therefore, the preferred weights can be set as follows: .

[0150] Generation of safety bubble cubes The safety bubble cube is a three-dimensional volume within which the system verifies its safety and prepares to guide the carrier drone into. Its generation process is as follows: Two-dimensional safe region search: in the comprehensive cost map Above, to reconnoiter the current location of the drone. Starting from the two-dimensional projection point, a region growing algorithm is executed forward. This algorithm, starting from the initial point, continuously grows the surrounding regions that meet the cost threshold. (Preferred) Connected grid points are incorporated into a set until no new points can be added. The area covered by this set is the two-dimensional safe flight zone. .

[0151] 3D Vertical Expansion and Obstacle Constraints: Transforming 2D Safe Zones The system expands vertically (Z-axis) to form an initial three-dimensional space. However, this expansion must be strictly constrained by aerial obstacles. The system utilizes the latest global lidar point cloud data acquired through data acquisition. Adjustments are made to this three-dimensional space.

[0152] Specifically, for The maximum safe flight altitude for each two-dimensional grid within it. It is set to the height of the first LiDAR point cloud in the vertical space above the grid, minus a safety margin. (Preferred, 3 meters). The final generated safety bubble cube It is a bottom surface However, the top surface of the three-dimensional safety space is uneven due to obstacles in the air (such as power lines and tree branches).

[0153] The next flight path command is generated within the already generated safety bubble cube. The final step is to plan the next segment of a specific flight path for the drone within this safe space.

[0154] Preferably, a fast random tree algorithm is used. This is based on the current location of the transport drone. Starting from, with The target point is the projection of the geometric center or the point furthest in front of it. Within the boundary constraints, quickly search for a smooth, collision-free flight path segment. .

[0155] This path segment Encoded as a series of compact waypoints, along with their corresponding safety bubble cubes. The valid timestamp, together with the instructions for the next flight path, is sent to the carrier drone behind via an encrypted communication link.

[0156] Next, this embodiment elaborates on the closed-loop execution logic of the entire collaborative flight framework in the dynamic path planning method for multi-drone collaborative delivery.

[0157] The reconnaissance drone has successfully generated a safety bubble cube for the next flight segment. and internal flight path segments This step will describe in detail the process of encapsulating, issuing, and executing this security instruction.

[0158] The safe flight command package is constructed and issued by the onboard computing unit of the reconnaissance UAV. The planning results are encapsulated into a time-sensitive safe flight command package. This data packet is digitally signed to ensure its integrity and provenance, and its data structure includes at least: path segment data. This refers to a series (preferably 5-10) dense 3D waypoints that define the precise trajectory of the transport drone in the next phase. (Safety boundary data) This involves mathematically describing the geometric boundaries of the safety bubble cube (preferably using vertex coordinates) as the absolute safety envelope for the UAV's flight. (Generate timestamps.) The precise moment when the instruction packet was generated. Effective time window. The effective duration of the command packet (preferably 5-10 seconds). This ensures that the environmental information upon which the drone operates is always up-to-date, avoiding the risks caused by outdated information due to dynamic environmental changes.

[0159] The command packet was sent from the reconnaissance drone to the carrier drone via a 5G communication link.

[0160] Verify the command and execution of the transport drone. After receiving the SFCP, the UAV's flight control system first executes a rigorous verification procedure: checking the current time. Does it meet the requirements? If the instruction has expired, the instruction packet is discarded, and a retransmission request is sent to the reconnaissance drone, while the drone remains hovering within its current safe zone. Verification and digital signatures confirm that the data packet has not been tampered with or corrupted during transmission.

[0161] After successful verification, the carrier drone began to precisely track and execute the path segment. Throughout the execution, the carrier drone performs a safety envelope self-check task at a high frequency (preferably 50Hz): in, This represents the current execution time. If... The risk response mechanism will be triggered immediately if the carrier drone deviates from the designated safety bubble boundary due to external interference (such as a sudden gust of wind) or its own tracking error at any time.

[0162] As the carrier drone flies into the current bubble Simultaneously, the reconnaissance drone has begun to move forward, surveying the area and generating the next safety bubble. This alternating and synchronized operation of the carrier aircraft entering and the reconnaissance aircraft leading creates the unique leapfrog safety corridor rolling advancement mechanism of this invention.

[0163] The entire delivery task is accomplished by repeatedly executing a cycle of perception, planning, and execution.

[0164] The loop condition for this iterative process is: the current position of the carrier drone. and the final target point The distance is greater than the preset target tolerance. (Preferred, 5 meters).

[0165] In each iteration, the system state changes from the first iteration to the second iteration. The next iteration evolved to the [number]th The iteration process is as follows: state The carrier drone is safely flying over the area generated by the reconnaissance drone. A safety bubble .

[0166] survey The reconnaissance drone has moved to a distance in front of the carrier drone. At this point, they begin to perceive, assess, and plan for the unknown area ahead.

[0167] generate The reconnaissance drone successfully generated the first... A safety bubble and path segment .

[0168] Issued : Reconnaissance drones develop and issue new Safe Flight Command Packages (SFCPs) .

[0169] Entering the state : Transporting drones at SFCP Successfully received and verified SFCP before expiration. And seamlessly connected, it began to fly. .

[0170] This cycle repeats continuously until the transport drone reaches the target point. Nearby, the system transitioned to the final landing procedure.

[0171] Risk Response and Dynamic Replanning Mechanisms: Triggering Conditions for Risk Replanning The occurrence of any of the following conditions will immediately interrupt the current iteration loop and trigger the risk replanning mechanism: Safe corridor generation failed: Overall cost map of the area in front of the reconnaissance drone The above cannot find a value that meets the cost threshold. A sufficiently large connected region This indicates that there are unacceptable ground, space, or acoustic risks ahead.

[0172] Communication link timeout: The carrier drone is in its current Safe Flight Command Packet (SFCP) Valid time Unable to successfully receive and verify the next instruction packet SFCP before exhaustion. .

[0173] Safety envelope deviation: Safety envelope self-check for unmanned aerial vehicles Returns 0.

[0174] Once a replanning is triggered, the system strictly follows the protocol below: Emergency braking of the carrier drone: The carrier drone immediately stops executing the current path segment. It then enters a fixed-point hovering state, maintaining the currently valid safety bubble. The geometric center location. Reconnaissance UAV withdrawal and lateral detection: The reconnaissance UAV ceases forward reconnaissance and, based on the global reference flight path... Move to the side (left or right) a safe distance. (Preferred distance: 20 meters), while simultaneously retreating to maintain a safe distance from the carrier drone. The location is determined. An alternative path search is initiated: the reconnaissance drone, from its new lateral position, restarts the entire environmental perception and path generation process, attempting to bypass previously encountered risk areas from a new direction and generate a completely new safety bubble. A retry and reporting mechanism is set: the system sets a maximum number of retries. (Preferred, 3 times) and maximum replanning time If, within the restrictions, the reconnaissance drone still cannot find a viable safe corridor, the current path is considered partially blocked. In this case, the system will report to the ground: sending a path blockage alert to the collaborative mission management platform via the communication link, requesting higher-level decision-making or manual intervention. The system will execute the established emergency plan, and according to preset strategies, can choose to return along the verified safe path or proceed to the nearest preset emergency landing point.

[0175] This embodiment also discloses a dynamic path planning system for multi-drone collaborative delivery, the system comprising: Formation Deployment Module: Deploy a reconnaissance UAV in front of a carrier UAV to form a coordinated formation and begin the mission along the global reference flight path; Weight generation module: The reconnaissance UAV performs advance environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. Path planning module: It integrates the ground risk weight map, the space airflow risk and noise impact map to generate a comprehensive cost map, and combines the real-time perceived air obstacle information to generate a time-sensitive three-dimensional safety bubble cube in the low-cost area of ​​the comprehensive cost map, and plans the next flight path segment within the safety bubble cube. Iterative optimization module: The reconnaissance UAV sends the flight path segment and the safety bubble cube to the carrier UAV for flight execution, while the reconnaissance UAV moves forward to repeat the leading environment perception and safety bubble cube generation steps for the next flight segment; Anomaly planning module: The above steps are executed repeatedly until the mission is completed. If an unacceptable risk is detected ahead or the flight deviates from the safe envelope at any stage, a dynamic replanning mechanism is triggered, in which the carrier UAV hovers and the reconnaissance UAV searches for an alternative path.

[0176] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the aforementioned dynamic path planning method for multi-drone collaborative delivery.

[0177] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0178] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0179] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A dynamic path planning method for multi-drone collaborative delivery, characterized in that, The method includes: Deploy a reconnaissance drone in front of a carrier drone to form a coordinated formation and begin the mission along the global reference flight path; The reconnaissance drone conducts advance environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. By integrating ground risk weight map, space airflow risk and noise impact map to generate comprehensive cost map, and combining real-time perceived air obstacle information, a time-sensitive three-dimensional safety bubble cube is generated in the low-cost area of ​​the comprehensive cost map, and the next flight path segment is planned within the safety bubble cube. The reconnaissance drone sends the flight path segment and safety bubble cube to the carrier drone for flight execution. At the same time, the reconnaissance drone moves forward to repeat the leading environmental perception and safety bubble cube generation steps for the next flight segment. The above steps are repeated until the mission is completed. If at any stage an unacceptable risk is detected ahead or the flight deviates from the safe envelope, a dynamic replanning mechanism is triggered, in which the carrier drone hovers and the reconnaissance drone searches for an alternative path.

2. The dynamic path planning method for multi-UAV collaborative delivery according to claim 1, characterized in that: The method of fusing multimodal sensor data to generate a ground risk weight map specifically includes: performing multi-stage parallel feature extraction and interactive fusion of lidar data, visible light images, and infrared thermal imaging data through a hierarchical interactive encoder containing geometric structure branches, visual semantic branches, and dynamic thermodynamic branches. In at least one fusion stage, the dynamic thermodynamic branch extracts dynamic thermodynamic features. Used to generate gating signals Geometric features that guide the branches of the geometric structure Semantic features of visual semantic branches The fusion is performed, and the calculation method is as follows: ; in, The fusion features output at this stage As a standard processing unit, For element-wise multiplication, To achieve element-wise addition, before performing the ⊕ addition operation, the feature tensors of each path have been aligned to the same number of channels and spatial dimensions through the network structure settings; and a ground risk weight map is generated based on the final fused features. .

3. The dynamic path planning method for multi-UAV collaborative delivery according to claim 2, characterized in that: Analyzing its own flight status data to calculate space airflow risks specifically includes: calculating the real-time control torque vector based on the flight control motor control commands of the reconnaissance UAV. and in the sliding time window The total variance is calculated internally to obtain the turbulence intensity index. : ; Furthermore, the angular velocity error was calculated based on the inertial measurement unit data of the reconnaissance drone. The wind field direction characteristics were obtained by low-pass filtering. : ; in, For the average control torque, It is a low-pass filter. The cutoff frequency is used; and a space airflow risk map is constructed based on the turbulence intensity index.

4. The dynamic path planning method for multi-UAV collaborative delivery according to claim 3, characterized in that: The assessment and generation of noise impact maps specifically includes: calculating the noise impact of reconnaissance drones at specific receiving points. The sound pressure level generated at the location : ; And based on the perceived ambient background noise sound pressure level Calculate the acoustic influence index : ; in, For sound power level, As a directional factor, , , These are geometric diffusion, atmospheric absorption, and obstacle diffraction attenuation, respectively; and a noise impact map is generated based on the acoustic impact index calculated from all noise-sensitive points. .

5. The dynamic path planning method for multi-UAV collaborative delivery according to claim 4, characterized in that: The process of integrating ground risk weight maps, space airflow risk and noise impact maps to generate a comprehensive cost map specifically includes: Space airflow risk map Noise impact diagram After normalization, we get and ; By analyzing the ground risk weight map and after normalization and By performing a weighted summation, we obtain the comprehensive cost diagram. : ; in, This is a preset risk weighting coefficient.

6. The dynamic path planning method for multi-UAV collaborative delivery according to claim 5, characterized in that: The steps for generating a three-dimensional safety bubble cube include: In the comprehensive cost diagram Above, search for results that meet the preset cost threshold. Two-dimensional safe flight area ; Two-dimensional safe flight area Extending vertically and utilizing global point clouds acquired by lidar. Constrain the expanded three-dimensional space, which exists in any ground grid. Maximum safe height satisfy: ； in, This refers to the set of lidar points located in the vertical space above the grid. Let be the perpendicular coordinates of the point. To ensure a safety margin, a safety bubble cube is generated. .

7. The dynamic path planning method for multi-UAV collaborative delivery according to claim 6, characterized in that: The flight of the transport drone also includes: continuously checking its current position as the transport drone flies along the flight path segment. Is it located in the safety bubble cube? Within the boundaries, and examine the current moment. Is it within the effective time window of the safety bubble cube? Internally, this means continuously verifying the following conditions: ； in, This is the generation timestamp of the safety bubble cube; if any condition is not met, a dynamic replanning mechanism is triggered.

8. The dynamic path planning method for multi-UAV collaborative delivery according to claim 7, characterized in that: During the deployment process, the global reference flight path is used. Through a preprocessed compliance subgraph The A* search algorithm is used to solve for minimizing the overall toll cost. What was obtained The calculation method is as follows: ； in, For flight segment distance, For the maximum possible distance, Let be the static risk potential field function. The maximum risk value, and These are the weighting coefficients.

9. The dynamic path planning method for multi-UAV collaborative delivery according to claim 1, characterized in that: During the deployment process, the reconnaissance drone flies in front of the carrier drone to form a coordinated formation, which is achieved by controlling the reconnaissance drone to reach an instantaneous target position vector. This is achieved through a calculation method as follows: ； in, This is the real-time position vector of the transport drone. To provide global reference routes The projection operator, The preset safe leading distance, This is the unit tangent vector of the global reference route at this projection point.

10. A dynamic path planning system for multi-drone collaborative delivery, characterized in that: The system includes: Formation Deployment Module: Deploy a reconnaissance UAV in front of a carrier UAV to form a coordinated formation and begin the mission along the global reference flight path; Weight generation module: The reconnaissance UAV conducts preliminary environmental perception of the airspace to be explored ahead. The perception includes: fusing multimodal sensor data to generate a ground risk weight map, analyzing its own flight status data to calculate the space airflow risk, and assessing and generating a noise impact map. Path planning module: It integrates the ground risk weight map, the space airflow risk and noise impact map to generate a comprehensive cost map, and combines real-time perceived air obstacle information to generate a time-sensitive three-dimensional safety bubble cube in the low-cost area of ​​the comprehensive cost map, and plans the next flight path segment within the safety bubble cube. Iterative optimization module: The reconnaissance UAV sends the flight path segment and safety bubble cube to the carrier UAV for flight execution. At the same time, the reconnaissance UAV moves forward to repeat the leading environmental perception and safety bubble cube generation steps for the next flight segment. Anomaly planning module: The above steps are executed repeatedly until the mission is completed. If an unacceptable risk is detected ahead or the flight deviates from the safety envelope at any stage, a dynamic replanning mechanism is triggered, in which the carrier UAV hovers and the reconnaissance UAV searches for an alternative path.

Images