A method and system for low-altitude logistics distribution
By employing multimodal perception fusion and a multi-level security protection mechanism, the positioning accuracy and security issues in drone delivery have been resolved, enabling all-weather, high-precision, and low-risk drone delivery capabilities that meet the actual needs of urban building delivery.
Patent Information
- Application Number
- CN202611201387.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-10
- Publication Date
- 2026-11-17
AI Technical Summary
Existing drone delivery technologies suffer from insufficient positioning accuracy, poor environmental adaptability, and a lack of safety protection mechanisms, making it difficult to meet the high reliability and high safety requirements of low-altitude food delivery.
A multimodal perception fusion system is adopted, combining millimeter-wave radar, infrared thermal imager and visible light camera to achieve all-weather 3D environment modeling. Dynamic landing area identification and evaluation are performed through an improved semantic segmentation network. Centimeter-level relative positioning is achieved by combining extended Kalman filter algorithm. A dual obstacle avoidance system of forward obstacle avoidance radar and downward ultrasonic sensor is deployed. Real-time path planning is combined with deep reinforcement learning. Model predictive control and sliding mode variable structure controller are used for precise landing.
It achieves centimeter-level relative positioning accuracy, improves environmental adaptability, reduces collision risk, ensures high reliability and safety in all weather conditions, and has a success rate of over 99.5%.
Smart Images

Figure CN123117948A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of logistics technology, and in particular to a low-altitude logistics delivery method and system. Background Technology
[0002] With the rapid growth in demand for on-demand delivery and the increasing maturity of drone technology, utilizing drones for last-mile logistics delivery has become an important development direction for the industry. Drone delivery has significant advantages such as being unaffected by ground traffic congestion, short straight-line flight distance, and high delivery efficiency, demonstrating enormous application potential in scenarios such as food delivery, fresh food delivery, and urgent medical supplies delivery.
[0003] Currently, drone last-mile delivery mainly adopts the following technical solutions: 1. GPS / RTK-based fixed-coordinate landing method Most existing drone delivery systems rely on GPS or RTK differential positioning technology to deliver goods to pre-set fixed pickup points (such as users' balconies or the top platform of community parcel lockers). After the drone flies to the target point coordinates according to a pre-planned three-dimensional flight path, it descends with the assistance of visual or ultrasonic sensors to complete the delivery. This method has significant limitations in practical applications: GPS signals are easily blocked and reflected in urban canyons and high-density residential areas, resulting in large fluctuations in positioning accuracy. This makes it difficult for drones to accurately reach narrow pickup locations, posing a risk of landing on the wrong floor or colliding with obstacles. Furthermore, the fixed coordinate method cannot cope with dynamic changes such as moved pickup boxes or temporary obstructions, exhibiting poor adaptability.
[0004] 2. Parachute or rope descent deployment method Some solutions involve hovering drones over the delivery area and then gently lowering the goods to the ground using ropes or parachutes, without the drone landing directly. While this method reduces the risk of collisions between drones and people, rope descent is significantly affected by wind, resulting in large deviations in the landing position of the goods. It also poses safety hazards such as rope entanglement and accidental receipt of goods by others, leading to a low delivery success rate, and is particularly unsuitable for densely populated urban residential areas.
[0005] 3. QR code or visual signage guidance at the receiving point Another approach involves setting up visual markers such as QR codes, color blocks, or infrared beacons at delivery points, with drones equipped with cameras recognizing the markers and guiding them to land. This method works reasonably well under good lighting conditions, with clear and unobstructed markings. However, the reliability of visual recognition drops significantly at night, in rainy or foggy weather, or when the markings are damaged or obstructed, making it difficult to meet the needs of 24 / 7 delivery.
[0006] 4. Existing solutions provide insufficient protection for personal safety. Existing technological solutions generally lack effective preventative mechanisms against potential harm to recipients and surrounding personnel during the final landing or delivery phase of a drone. If a recipient or child suddenly approaches while the drone is nearing the ground or delivery container, the rapidly rotating rotors and descending cargo could cause personal injury. Existing systems lack the ability to perceive and proactively avoid such dynamic risks.
[0007] In summary, existing drone delivery solutions have significant shortcomings in terms of precise positioning, environmental adaptability, prevention of accidental pickup, and personal safety protection, making it difficult to meet the actual needs of low-altitude food delivery for high reliability and high safety. Summary of the Invention
[0008] The purpose of this invention is to provide a low-altitude logistics delivery method and system that can effectively solve the problems in the existing drone last-mile delivery technology, such as insufficient positioning accuracy, poor environmental adaptability, and lack of safety protection mechanisms.
[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A low-altitude logistics delivery method includes the following specific steps: Step 1: Constructing a multimodal perception fusion system: By integrating three heterogeneous sensors—millimeter-wave radar, infrared thermal imager, and visible light camera—on a UAV, a perception framework for multi-source data fusion is established to achieve all-weather 3D environmental modeling of the landing area; Step 2: Dynamic landing area identification and evaluation: Based on the multimodal perception fusion data, an improved semantic segmentation network is used to identify the landing area in real time, calculate the flatness, load-bearing capacity, and safe distance of the area, generate a comprehensive score, and dynamically select the optimal landing point; Step 3: High-precision relative positioning: fusing visual odometry and millimeter-wave radar... Radar ranging data is used to achieve centimeter-level relative positioning accuracy through an extended Kalman filter algorithm, correcting positioning errors caused by GPS signal obstruction or multipath effects; Step 4: Multi-level safety protection mechanism: A dual obstacle avoidance system consisting of forward-looking obstacle avoidance radar and downward-looking ultrasonic sensors is deployed, combined with a real-time path planning algorithm based on deep reinforcement learning, to automatically perform hovering, detour, or abort landing operations when personnel or obstacles are detected approaching; Step 5: Precise landing control: A sliding mode variable structure controller based on model predictive control is adopted, combined with a real-time wind speed and direction compensation algorithm, to control the UAV to complete a precise landing with a positioning deviation of less than or equal to 0.3 meters.
[0010] Preferably, in step 1, the millimeter-wave radar operates in the frequency band of 77 to 81 GHz, has an angular resolution of less than or equal to 2 degrees, and a ranging accuracy of less than or equal to 0.05 meters; the infrared thermal imager has a resolution of 640 x 512 pixels, operates in the wavelength band of 8 to 14 micrometers, and can identify heat source targets with a temperature difference greater than or equal to 0.05 degrees Celsius; the visible light camera uses a 20-megapixel global shutter sensor with a frame rate of 60 frames per second.
[0011] Preferably, the improved semantic segmentation network in step 2 is based on an encoder-decoder architecture. The encoder uses a 50-layer residual network as the backbone network, and the decoder introduces a pyramid pooling module to fuse multi-scale features, achieving a segmentation accuracy of greater than or equal to 95% for the landing area. The comprehensive scoring weights are: flatness 40%, load-bearing capacity 35%, and safety distance 25%.
[0012] Preferably, in step 3, the state vector of the extended Kalman filter algorithm includes 12 dimensions: UAV position, velocity, attitude angle, and angular velocity. The observations include the relative position output by the visual odometry, the distance information measured by the millimeter-wave radar, and the altitude data from the barometer. The filter update frequency is 100 Hz, and the relative positioning accuracy reaches the centimeter level.
[0013] Preferably, in step 4, the forward obstacle avoidance radar has a detection range of ≥30 meters, a horizontal field of view of ≥120 degrees, a vertical field of view of ≥60 degrees, and can detect obstacles with a minimum size of 0.05m x 0.05m x 0.05m; the ultrasonic sensor has an effective detection height range of 0.1 to 5 meters and a ranging accuracy of ≤0.01 meters; the deep reinforcement learning algorithm adopts an actor-critic architecture, with a state space containing 256 dimensions including radar point cloud, image features, and UAV state, and an action space of a 4-dimensional discrete action set; the reward function comprehensively considers obstacle avoidance success rate, path smoothness, and task completion efficiency.
[0014] Preferably, in step 5, the prediction time domain of the model predictive control is 0.5 seconds, the control time domain is 0.1 seconds, the optimization solution adopts a quadratic programming algorithm, and the solution period is less than or equal to 10 milliseconds; the switching gain of the sliding mode variable structure controller is adaptively adjusted according to the upper bound of the system uncertainty, and the convergence rate exponent is greater than or equal to 2; the wind speed compensation algorithm is based on a feedforward-feedback structure, the feedforward component is calculated in real time based on the wind speed sensor data, and the feedback component is compensated by the attitude loop controller.
[0015] Preferably, the method also includes the step of sending a confirmation command to the ground control station after the landing area is locked, the ground control station verifying the safety of the landing point and authorizing the landing, and the UAV remaining in a hovering state until it receives the authorization signal.
[0016] Preferably, it also includes a real-time safety monitoring step during the landing process: continuously monitoring the movement trajectory of personnel in the landing area, predicting the movement trend of personnel, and when there is a risk of the predicted trajectory intersecting with the landing path of the drone, the drone automatically increases its altitude to a safe altitude and reassesses the landing point.
[0017] Preferably, when the method is applied to urban building delivery scenarios, the drone's body weight is less than or equal to 25 kg, its maximum payload is 5 kg, its maximum flight speed is 15 m / s, its endurance is greater than or equal to 30 minutes, and it can perform delivery tasks normally in winds below level 6.
[0018] The beneficial effects of this invention are: 1. High-precision positioning and landing capability. Breaking through the bottleneck of positioning accuracy: By combining multimodal perception fusion with extended Kalman filtering algorithm, centimeter-level relative positioning accuracy is achieved, completely eliminating dependence on GPS signals, effectively solving the positioning drift problem in urban canyon environments, with a landing position deviation of less than or equal to 0.3 meters, and an accuracy improvement of more than 80% compared with traditional RTK positioning methods; Strong adaptability to complex environments: millimeter-wave radar has strong ability to penetrate rain and fog, infrared thermal imager can work normally at night, and visible light camera provides high-resolution visual information during the day. The three sensors complement each other and are redundant, enabling uninterrupted delivery capabilities 24 hours a day.
[0019] 2. High-security guarantee mechanism Multi-layered protection system: The dual obstacle avoidance system, consisting of forward-looking obstacle avoidance radar and downward-looking ultrasonic sensors, combined with real-time path planning using deep reinforcement learning, can respond and perform avoidance actions within 0.5 seconds in sudden situations such as people suddenly approaching or obstacles appearing, reducing the risk of collision by more than 95%. Intelligent risk prediction: By continuously monitoring the movement trajectory of people in the landing area and predicting movement trends, it can proactively increase altitude or reselect landing point before path conflicts occur, upgrading from passive avoidance to proactive prevention.
[0020] 3. High reliability performance High system redundancy: The three heterogeneous sensors are mutually redundant, and the failure of any one sensor will not affect the basic functions of the system. The safety protection mechanism adopts a dual design, and a single failure will not cause the system to completely lose its protection capability. High control precision: The landing controller, based on a combination of model predictive control and sliding mode variable structure control, and with the wind speed compensation algorithm, can maintain a stable landing even under level 4 wind conditions, with a success rate of greater than or equal to 99.5%, which is more than 10 percentage points higher than the traditional PID controller. Attached Figure Description
[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the low-altitude logistics delivery method proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework for achieving high-precision relative positioning by combining multimodal perception fusion and extended Kalman filtering in this invention. Detailed Implementation
[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0023] Specific implementation examples are given below. Example
[0024] This embodiment focuses on an urban building delivery scenario, detailing the specific implementation of the low-altitude logistics delivery method and system of the present invention. In this scenario, the drone needs to transport goods from a delivery point to a receiving box in the target building, completing precise landing and cargo delivery. The overall system consists of five core modules: a multimodal perception fusion subsystem, a dynamic landing area identification and evaluation subsystem, a high-precision relative positioning subsystem, a multi-level safety protection subsystem, and a precise landing control subsystem. These subsystems are interconnected via a high-speed data bus, forming a complete closed-loop control system.
[0025] At the system hardware architecture level, the drone platform adopts a hexacopter configuration, with the fuselage made of carbon fiber composite material, and the overall weight controlled within 25 kg. It has a maximum payload capacity of 5 kg, a maximum flight speed of 15 m / s, and an endurance of no less than 30 minutes, enabling it to perform delivery tasks normally in winds below level 6. This drone platform serves as the physical carrier of the system, providing installation space and power support for various sensors and control modules.
[0026] The multimodal perception fusion subsystem is the foundational module for all-weather environmental perception. Its hardware comprises three heterogeneous sensors: a millimeter-wave radar, an infrared thermal imager, and a visible light camera. The millimeter-wave radar operates at 77-81 GHz, with an angular resolution of ≤2 degrees and a ranging accuracy of ≤0.05 meters. Mounted at the forward position of the UAV, it acquires distance and angle information for obstacles ahead. The infrared thermal imager has a resolution of 640 x 512 pixels, operates at 8-14 micrometers, and can identify heat sources with a temperature difference greater than or equal to 0.05 degrees Celsius. Mounted on the bottom of the UAV, it detects targets at night or in low-visibility environments. The visible light camera uses a 20-megapixel global shutter sensor with a frame rate of 60 frames per second. Mounted downwards on the bottom of the UAV, it acquires high-resolution visual information. All three sensors achieve time-synchronized data acquisition via a high-speed data acquisition card, with a uniform sampling frequency of 100 Hz. Data is transmitted via gigabit Ethernet to the onboard edge computing unit for processing.
[0027] At the data fusion level, the multimodal perception fusion subsystem establishes a perception framework for multi-source data fusion. First, image data acquired by a visible light camera undergoes distortion correction and color restoration by an ISP image signal processor before being fed into a subsequent semantic segmentation network for processing. Raw point cloud data acquired by millimeter-wave radar is filtered, denoised, and clustered to form 3D spatial distribution information of obstacles. Thermal image data acquired by an infrared thermal imager, after temperature calibration and noise suppression, is used to extract the location and contour information of heat source targets. After aligning the data from the three sensors in the temporal dimension, spatial fusion is performed according to a weighted fusion strategy, ultimately generating a comprehensive environmental perception result containing obstacle location, material properties, and motion state, enabling all-weather 3D environmental modeling of the landing area.
[0028] The core hardware of the dynamic landing area identification and evaluation subsystem is an onboard high-performance GPU computing platform, equipped with a graphics processor with at least 8GB of video memory, used to run an improved semantic segmentation network. This network is based on an encoder-decoder architecture. The encoder uses a 50-layer residual network as its backbone, effectively mitigating the gradient vanishing problem in deep networks through residual connections, enabling multi-level feature extraction. The decoder introduces a pyramid pooling module, fusing global contextual information and local detailed features through pooling operations at different scales, significantly improving the segmentation accuracy of complex landing areas, achieving a segmentation accuracy of greater than or equal to 95%. This subsystem first receives 3D environmental data output from the multimodal perception fusion subsystem, extracting potentially flat areas as candidate landing areas. Then, based on the multimodal perception fusion data, each candidate landing area undergoes multi-dimensional evaluation: flatness evaluation determines surface flatness by analyzing the region's elevation variance; load-bearing capacity evaluation combines material information detected by an infrared thermal imager and structural features identified by a visible light camera for comprehensive judgment; and safety distance evaluation calculates the minimum distance between the candidate area and surrounding obstacles and personnel activity areas. Based on the above evaluation results, the system calculates a comprehensive score with a weighting of 40% for flatness, 35% for load-bearing capacity, and 25% for safety distance, and dynamically selects the optimal landing point with the highest comprehensive score.
[0029] The high-precision relative positioning subsystem comprises a visual odometry module, a millimeter-wave radar ranging module, a barometer module, and an extended Kalman filter processing unit. The visual odometry module, consisting of a visible light camera and a matching image processing chip, calculates relative position changes through feature point matching and triangulation of consecutive frames. The millimeter-wave radar ranging module utilizes the high-precision ranging capability of a 77-81 GHz radar to obtain the precise distance between the UAV and the landing area. The barometer module measures atmospheric pressure and converts it into altitude information, providing an absolute altitude reference. The extended Kalman filter processing unit, implemented using a high-performance DSP chip, has a state vector containing 12 dimensions: UAV position, velocity, attitude angle, and angular velocity. Observations include the relative position output from the visual odometry, the distance information measured by the millimeter-wave radar, and the barometer altitude data. The filter update frequency is 100 Hz. This subsystem achieves centimeter-level relative positioning accuracy by fusing visual odometry and millimeter-wave radar ranging data and utilizing the extended Kalman filter algorithm, effectively correcting positioning errors caused by GPS signal obstruction or multipath effects. The extended Kalman filter algorithm employs two alternating phases: prediction and update. In the prediction phase, the current state is estimated based on the UAV dynamics model, and in the update phase, the predicted values are corrected based on the observations of each sensor. Finally, a high-precision estimate of the relative position and velocity is output.
[0030] The hardware deployment of the multi-level safety protection subsystem includes a dual obstacle avoidance system consisting of a forward-looking obstacle avoidance radar and a downward-looking ultrasonic sensor. The forward-looking obstacle avoidance radar has a detection range of ≥30 meters, a horizontal field of view of ≥120 degrees, a vertical field of view of ≥60 degrees, and can detect obstacles as small as 0.05m x 0.05m x 0.05m. It is installed at the forward position of the UAV to detect obstacles ahead of the flight path. The downward-looking ultrasonic sensor has an effective detection height range of 0.1 to 5 meters and a ranging accuracy of ≤0.01 meters. It is installed at the bottom of the UAV facing downwards to detect near-ground obstacles and protrusions on the landing area surface. The core algorithm of this subsystem employs a real-time path planning algorithm based on deep reinforcement learning, specifically implemented using an actor-commentator architecture. The state space comprises 256 dimensions, including radar point cloud, image features, and UAV state. The action space is a 4-dimensional discrete action set, including four actions: maintaining the current track, slight detour, large detour, and emergency hovering. The reward function comprehensively considers obstacle avoidance success rate, path smoothness, and task completion efficiency, obtaining the optimal strategy through extensive simulation training. When the forward-looking obstacle avoidance radar or downward-looking ultrasonic sensor detects a person approaching or an obstacle appearing, the deep reinforcement learning algorithm completes the response assessment within 0.5 seconds and automatically executes hovering, detour, or abort landing operations. In addition, this subsystem also includes the function of sending a confirmation command to the ground control station after the landing area is locked. The ground control station verifies the safety of the landing point and authorizes landing, and the UAV remains hovering until it receives the authorization signal.
[0031] The hardware of the precision landing control subsystem includes a model predictive controller, a sliding mode variable structure controller, a wind speed sensor, and a motor drive unit. The model predictive control has a prediction time domain of 0.5 seconds and a control time domain of 0.1 seconds. The optimization solution uses a quadratic programming algorithm with a solution period of less than or equal to 10 milliseconds, enabling real-time control with limited computational resources. The switching gain of the sliding mode variable structure controller is adaptively adjusted according to the upper bound of the system uncertainty, with a convergence rate exponent greater than or equal to 2, ensuring that the system state reaches the sliding surface within a finite time and remains stable. The wind speed sensor is installed at the end of the UAV arm to measure wind speed and direction data in real time. The motor drive unit uses an electronic speed controller with a response time of less than or equal to 1 millisecond and is used to execute the controller's output commands.
[0032] This subsystem employs a sliding mode variable structure controller based on model predictive control, combined with a real-time wind speed and direction compensation algorithm to achieve precise landing. The control structure uses a feedforward-feedback hybrid approach: the feedforward component calculates the wind disturbance component on the UAV in real time based on wind speed sensor data and compensates for it in the control input; the feedback component is calculated by the attitude loop controller based on position and velocity errors. Model predictive control generates the desired acceleration command, and the sliding mode variable structure controller calculates the motor speed adjustment based on the desired acceleration. The combined effect of these two components enables the UAV to achieve a precise landing with a positioning deviation of less than or equal to 0.3 meters. Even under wind speed conditions of level 4, the system maintains stable landing with a success rate of greater than or equal to 99.5%.
[0033] At the methodological level, the low-altitude logistics delivery method of this embodiment is executed sequentially according to the following steps: First, the drone takes off from the delivery point and flies towards the target building along a preset route. During the flight, the multimodal perception fusion subsystem continuously collects environmental data and constructs a three-dimensional spatial model of the flight path. The forward obstacle avoidance radar of the multi-level safety protection subsystem continuously scans obstacles within a 30-meter range ahead of the flight path, and the deep reinforcement learning algorithm assesses collision risks in real time and generates avoidance commands when necessary.
[0034] Secondly, when the drone flies near the target building, it enters the landing area search phase. The drone hovers around the building and descends, while the multimodal perception fusion subsystem scans for possible landing areas. The dynamic landing area identification and evaluation subsystem receives the fused environmental data, runs an improved semantic segmentation network to identify landing areas that meet the criteria, and performs a comprehensive evaluation of each candidate area based on flatness, load-bearing capacity, and safety distance, ultimately selecting the optimal landing point.
[0035] Then, the drone flies to the selected landing area and enters the high-precision relative positioning phase. The high-precision relative positioning subsystem activates the extended Kalman filter algorithm, fusing ranging data from visual odometry and millimeter-wave radar to calculate the centimeter-level relative position between the drone and the landing area in real time. Simultaneously, the multimodal perception fusion subsystem continuously updates the 3D model of the landing area, monitoring for personnel or other dynamic obstacles entering the area.
[0036] During descent, a multi-layered safety protection subsystem continuously plays a protective role. Downward-looking ultrasonic sensors monitor obstacles within a 5-meter range below the drone in real time; if an abnormal approach to the target is detected, an emergency hovering or detour maneuver is triggered. This subsystem also includes real-time safety monitoring during descent: continuously monitoring the movement trajectories of personnel within the descent area, predicting personnel movement trends through time-series data analysis, and automatically increasing the drone's altitude to a safe level and reassessing the landing point if the predicted trajectory intersects with the drone's descent path.
[0037] Finally, when the drone descends to a height of 5 meters above the landing area and all safety conditions are met, the system sends a landing area lock confirmation command to the ground control station. Upon receiving the confirmation, the ground control station verifies the safety of the landing point, including confirming the absence of personnel activity and whether obstacle conditions have changed. After successful verification, the ground control station sends an authorized landing signal to the drone. Upon receiving the authorization signal, the precise landing control subsystem initiates the landing control program. The model prediction controller calculates the desired acceleration based on the current position and velocity errors, the sliding mode variable structure controller converts the desired acceleration into motor speed adjustment commands, and the wind speed compensation algorithm compensates for wind disturbances in real time. Under the control system, the drone lands smoothly at the target location with a positioning deviation of less than or equal to 0.3 meters, completes cargo delivery, and returns to the delivery point along the original route.
[0038] Throughout the delivery process, all subsystems work collaboratively to form a complete closed-loop control system. The multimodal perception fusion subsystem provides the foundation for environmental perception, the dynamic landing area identification and evaluation subsystem determines the optimal landing point, the high-precision relative positioning subsystem provides accurate location information, the multi-level safety protection subsystem ensures operational safety, and the precision landing control subsystem achieves accurate landing. The functions of each subsystem support each other, and the data is cross-checked to ensure the reliable and efficient completion of the entire delivery task.
[0039] In this embodiment, a combination of multimodal perception fusion and extended Kalman filtering achieves centimeter-level relative positioning accuracy, completely eliminating reliance on GPS signals and effectively solving the positioning drift problem in urban canyon environments. Millimeter-wave radar has strong rain and fog penetration capabilities, infrared thermal imagers can operate normally at night, and visible light cameras provide high-resolution visual information during the day. These three sensors complement each other, enabling 24 / 7 uninterrupted delivery capabilities. A dual obstacle avoidance system consisting of forward-looking obstacle avoidance radar and downward-looking ultrasonic sensors, combined with deep reinforcement learning real-time path planning, can quickly respond and execute avoidance actions in sudden situations such as sudden approach of personnel or the appearance of obstacles. A landing controller based on a combination of model predictive control and sliding mode variable structure control, along with a wind speed compensation algorithm, maintains stable landing even under complex weather conditions.
[0040] Through the comprehensive application of the above-mentioned technical means, the low-altitude logistics delivery method and system of the present invention have achieved the expected results in terms of positioning accuracy, environmental adaptability and safety, and can meet the actual application needs of urban building delivery scenarios. Example
[0041] Building upon Example 1, this example further optimizes the system's safety protection mechanism and landing control strategy for delivery scenarios in high-rise buildings within complex urban environments. In this scenario, the target delivery box is typically located at a designated position on the building's exterior facade, with limited landing space and potential obstacles such as protrusions on the building surface and air conditioning units, placing higher demands on the system's obstacle detection accuracy and path planning capabilities.
[0042] At the system hardware level, this embodiment adds a forward-looking high-precision laser scanner and side obstacle avoidance sensors to the first embodiment. The forward-looking high-precision laser scanner uses the time-of-flight principle, achieving an angular resolution of 0.1 degrees and a detection range of 50 meters, and is used to acquire detailed three-dimensional structural information of the landing area and surrounding obstacles. The side obstacle avoidance sensors use low-cost millimeter-wave radar, installed on the left and right sides of the UAV, to monitor obstacle conditions during lateral flight.
[0043] In terms of dynamic landing zone identification and evaluation, the improved semantic segmentation network in this embodiment enhances the ability to identify building facade features. The network training dataset contains a large number of building facade images with different materials, colors, and structures, enabling the network to accurately distinguish landing areas from non-landing areas such as air conditioner outdoor units and balcony protrusions. Simultaneously, the evaluation algorithm adds the function of detecting attachments on building surfaces. By analyzing the 3D point cloud data acquired by the laser scanner, it identifies protrusions with a height exceeding 0.2 meters and excludes them from the candidate landing areas.
[0044] In terms of safety protection, this embodiment enhances the personnel activity prediction function. By fusing continuous frame images from a visible light camera and point cloud data from millimeter-wave radar, the system can not only detect the current position of personnel, but also predict their movement trajectory within the next 3 seconds based on time-series data and motion models. When there is a risk of the predicted trajectory intersecting with the landing path of the drone, the system triggers an avoidance action 0.5 seconds in advance, reducing the collision risk by more than 95%.
[0045] Regarding landing control, this embodiment has been specifically optimized for the complex wind field environment near the building facade. The wind speed compensation algorithm incorporates a building shading effect model, dynamically adjusting the feedforward compensation amount based on the relative position of the UAV and the building. When the UAV enters the vicinity of the windward side of the building, the feedforward compensation amount increases; when the UAV enters the vortex region on the leeward side of the building, the feedforward compensation amount decreases and the weight of feedback control is increased to ensure the stability of the control system.
[0046] The workflow of this embodiment is basically the same as that of Embodiment 1. The difference lies in the addition of building facade feature recognition and obstacle detection during the landing area search phase, the addition of motion prediction-based personnel avoidance function during landing, and the addition of adaptive adjustments to the landing control algorithm for the wind field environment near the building. Through these improvements, the reliability and safety of the system in high-rise building delivery scenarios are further enhanced. Example
[0047] This embodiment addresses delivery scenarios in remote areas or under special terrain conditions, such as mountainous regions, islands, or temporary take-off and landing points. These scenarios are characterized by significant terrain undulations, a lack of accurately mapped data, and the potential presence of temporary obstacles such as trees or vehicles, placing higher demands on the system's environmental adaptability and autonomous decision-making capabilities.
[0048] At the system hardware level, this embodiment adds a terrain-aided navigation module and an emergency communication module. The terrain-aided navigation module includes a forward-looking terrain radar and a terrain database. By comparing real-time detected terrain data with pre-stored terrain information, it improves navigation accuracy in areas with weak or unavailable GPS signals. The emergency communication module adopts a redundant design, integrating satellite communication and a self-rescue radio to ensure that basic control functions can be maintained even if the communication link is interrupted.
[0049] In terms of dynamic landing area identification and evaluation, the evaluation algorithm in this embodiment adds the evaluation dimensions of ground slope and load uniformity. By analyzing the terrain data acquired by millimeter-wave radar and laser scanner, the slope variance and undulation of candidate areas are calculated to filter out flat areas suitable for UAV landing. For areas with a slope exceeding 15 degrees or with obvious unevenness, the system automatically reduces their comprehensive score weight or excludes them.
[0050] In terms of safety, this embodiment adds the ability to identify wild animals and livestock. By improving the target detection algorithm, the system can identify common ground animals and classify them as obstacles. When an animal is detected approaching the landing area, the system implements the same avoidance strategy as when personnel approach, ensuring that the delivery operation does not disturb the local ecosystem.
[0051] This embodiment adds terrain data download and landing area pre-screening steps before takeoff. Before takeoff, the UAV obtains the terrain database of the target area from the ground control station and preliminarily screens possible safe landing areas based on the terrain data during the flight path planning stage. During flight, the system continuously updates the assessment results of the landing areas and dynamically adjusts the landing strategy based on the latest environmental data.
[0052] Through the above technical solutions, the system of this embodiment can adapt to the delivery needs of remote areas and special terrain conditions, expand the application scope of low-altitude logistics delivery technology, and provide technical support for a wider range of logistics delivery scenarios.
[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A low-altitude logistics delivery method, characterized in that, Includes the following steps: Step 1, Construct a multimodal perception fusion system: By integrating three heterogeneous sensors—millimeter-wave radar, infrared thermal imager, and visible light camera—on the UAV, a perception framework for multi-source data fusion is established to achieve all-weather 3D environmental modeling of the landing area. Step 2, Dynamic Landing Area Identification and Evaluation: Based on multimodal perception fusion data, an improved semantic segmentation network is used to identify the landing area in real time, calculate the flatness, load-bearing capacity and safety distance of the area, generate a comprehensive score and dynamically select the optimal landing point; Step 3, High-precision relative positioning: By fusing visual odometry and millimeter-wave radar ranging data, centimeter-level relative positioning accuracy is achieved through extended Kalman filtering algorithm, correcting positioning errors caused by GPS signal blockage or multipath effects. Step 4, Multi-level safety protection mechanism: Deploy a dual obstacle avoidance system consisting of forward obstacle avoidance radar and downward-looking ultrasonic sensors, combined with a real-time path planning algorithm based on deep reinforcement learning, to automatically perform hovering, detour or abort landing operations when personnel or obstacles are detected approaching. Step 5, Precise Landing Control: A sliding mode variable structure controller based on model predictive control is adopted, combined with a real-time wind speed and direction compensation algorithm, to control the UAV to complete a precise landing with a positioning deviation of less than or equal to 0.3 meters.
2. The method according to claim 1, characterized in that, In step 1, the millimeter-wave radar operates at a frequency of 77 to 81 GHz, has an angular resolution of less than or equal to 2 degrees, and a ranging accuracy of less than or equal to 0.05 meters; the infrared thermal imager has a resolution of 640 x 512 pixels, operates at a wavelength of 8 to 14 micrometers, and can identify heat source targets with a temperature difference greater than or equal to 0.05 degrees Celsius. The visible light camera uses a 20-megapixel global shutter sensor with a frame rate of 60 frames per second.
3. The method according to claim 1, characterized in that, The improved semantic segmentation network in step 2 is based on an encoder-decoder architecture. The encoder uses a 50-layer residual network as the backbone network, and the decoder introduces a pyramid pooling module to fuse multi-scale features, achieving a segmentation accuracy of greater than or equal to 95% for the landing area. The comprehensive scoring weights are: flatness 40%, load-bearing capacity 35%, and safety distance 25%.
4. The method according to claim 1, characterized in that, In step 3, the state vector of the extended Kalman filter algorithm includes 12 dimensions: UAV position, velocity, attitude angle, and angular velocity. The observations include the relative position output by the visual odometry, the distance information measured by the millimeter-wave radar, and the altitude data from the barometer. The filter update frequency is 100 Hz, and the relative positioning accuracy reaches the centimeter level.
5. The method according to claim 1, characterized in that, In step 4, the forward obstacle avoidance radar has a detection range of ≥30 meters, a horizontal field of view of ≥120 degrees, a vertical field of view of ≥60 degrees, and can detect obstacles with a minimum size of 0.05m x 0.05m x 0.05m. The ultrasonic sensor has an effective detection height range of 0.1 to 5 meters and a ranging accuracy of ≤0.01 meters. The deep reinforcement learning algorithm adopts an actor-critic architecture, with a state space containing 256 dimensions including radar point cloud, image features, and UAV state, and an action space of a 4-dimensional discrete action set.
6. The method according to claim 1, characterized in that, In step 5, the prediction time domain of the model predictive control is 0.5 seconds, the control time domain is 0.1 seconds, and the optimization solution adopts the quadratic programming algorithm with a solution period of less than or equal to 10 milliseconds. The switching gain of the sliding mode variable structure controller is adaptively adjusted according to the upper bound of the system uncertainty, and the approach rate exponent is greater than or equal to 2. The wind speed compensation algorithm is based on a feedforward-feedback structure. The feedforward component is calculated in real time based on wind speed sensor data, and the feedback component is compensated by the attitude loop controller.
7. The method according to claim 1, characterized in that, It also includes the step of sending a confirmation command to the ground control station after the landing area is locked, the ground control station verifying the safety of the landing point and authorizing the landing, and the drone remaining in a hovering state until it receives the authorization signal.
8. The method according to claim 1, characterized in that, It also includes real-time safety monitoring steps during the landing process: continuously monitoring the movement trajectory of personnel in the landing area, predicting the movement trend of personnel, and when there is a risk of the predicted trajectory intersecting with the drone's landing path, the drone automatically increases its altitude to a safe altitude and reassesses the landing point.
9. The method according to any one of claims 1 to 8, characterized in that, When the method is applied to urban building delivery scenarios, the drone's body weight is less than or equal to 25 kg, the maximum payload is 5 kg, the maximum flight speed is 15 m / s, the endurance is greater than or equal to 30 minutes, and it can perform delivery tasks normally in winds below level 6.
10. A low-altitude logistics distribution system, characterized in that, include: The multimodal perception fusion subsystem is used to establish a perception framework for multi-source data fusion through three heterogeneous sensors: millimeter-wave radar, infrared thermal imager, and visible light camera, so as to realize all-weather three-dimensional environmental modeling of the landing area. The dynamic landing area identification and evaluation subsystem is used to identify landing areas in real time based on multimodal perception fusion data and an improved semantic segmentation network, calculate the flatness, load-bearing capacity and safety distance of the area, generate a comprehensive score and dynamically select the optimal landing point; A high-precision relative positioning subsystem is used to fuse visual odometry and millimeter-wave radar ranging data, and achieve centimeter-level relative positioning accuracy through an extended Kalman filter algorithm; The multi-level safety protection subsystem is used to deploy a dual obstacle avoidance system consisting of forward-looking obstacle avoidance radar and downward-looking ultrasonic sensors. Combined with a real-time path planning algorithm based on deep reinforcement learning, it automatically performs hovering, detour, or abort landing operations when it detects people approaching or obstacles. The precision landing control subsystem uses a sliding mode variable structure controller based on model predictive control, combined with a real-time wind speed and direction compensation algorithm, to control the UAV to complete a precise landing with a positioning deviation of less than or equal to 0.3 meters.
