Unmanned aerial vehicle accurate landing system and method based on visual positioning

Abstract

The invention relates to an unmanned aerial vehicle accurate landing system and method based on visual positioning. The system comprises a landing platform used for landing of an unmanned aerial vehicle; the positioning and guiding device comprises a plurality of independent controllable light-emitting systems arranged in a layered manner, an orientation sensing module used for sensing the position of the unmanned aerial vehicle, and a controller used for controlling controllable light-emitting units to generate dynamic visual guiding signals, and the orientation sensing module transmits sensed position information of the unmanned aerial vehicle to the controller; and the unmanned aerial vehicle end module is used for identifying the dynamic visual signals of the positioning and guiding device, calculating the space coordinates of the landing platform, and generating a landing path from the unmanned aerial vehicle to the landing platform based on the space coordinate information of the landing platform. The landing method uses the system to execute accurate landing of the unmanned aerial vehicle. Through implementation of the technical scheme, the problem of signal deficiency in a traditional scheme can be effectively solved, cargoes shield a bottom visual or laser sensor, and autonomous positioning and autonomous precise landing can be completed through visual information.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) flight technology, and more specifically to a visual positioning-based UAV precision landing system. Background Technology

[0002] With the widespread application of logistics drones in e-commerce delivery and remote area material transportation, the accuracy and reliability of their autonomous landing have become key technological bottlenecks. Existing mainstream landing systems mainly rely on GPS, inertial navigation, or multi-sensor fusion positioning, but these systems have revealed the following shortcomings in practical applications: First, existing systems cannot acquire drone pose information in remote areas without GPS signals (such as mountains and tunnels) or in environments with electromagnetic interference, leading to landing failures. Under complex lighting conditions (such as strong light and backlight) or weather conditions (such as rain, snow, and fog), the recognition accuracy of visual sensors drops significantly, with positioning errors reaching the meter level, which cannot meet the millimeter-level accuracy requirements of logistics drones.

[0003] Secondly, traditional solutions often employ a multi-sensor redundancy design of "GPS + vision + LiDAR" to improve positioning reliability, which increases the drone's payload and complicates its structure. When the drone is carrying cargo, the cargo can easily obstruct the bottom vision or LiDAR sensors, creating positioning blind spots and jeopardizing landing safety. For example, in cold chain logistics drones, the scenario of insulated boxes obstructing the bottom camera is common, and existing technologies are difficult to address.

[0004] Secondly, existing landing systems heavily rely on real-time communication between the drone and the platform (such as WiFi or 4G). When communication is interrupted or the signal is weak, the drone cannot receive pose updates from the platform, leading to autonomous landing failure. In addition, existing path planning algorithms (such as artificial potential field methods and genetic algorithms) have poor generalization ability and are prone to getting stuck in local optima or collision risks when faced with sudden obstacles or sudden changes in wind speed.

[0005] In summary, there is an urgent need for a vision-based positioning-based UAV precision landing system and method that can solve the problem of signal loss in traditional solutions and the obstruction of bottom visual or laser sensors by cargo, enabling autonomous positioning and precise landing through visual information. Summary of the Invention

[0006] The purpose of this invention is to provide a vision-based positioning-based drone precision landing system and method that can solve the problem of signal loss in traditional solutions and the problem of cargo obstructing the bottom visual or laser sensors, and can achieve autonomous positioning and precise landing through visual information.

[0007] The above objective is achieved through the following technical solution: a visual positioning-based precision landing system for unmanned aerial vehicles (UAVs), comprising: Landing platform: Used for landing drones; Positioning and guidance device: Located on one side of the landing platform, it includes multiple independent controllable light-emitting systems arranged in layers, an orientation perception module for sensing the position of the UAV, and a controller. The orientation perception module transmits the sensed UAV position information to the controller, and the controller is used to control the controllable light-emitting units to generate dynamic visual guidance signals. The UAV terminal module is used to identify the dynamic visual signals of the positioning and guidance device, calculate the spatial coordinates of the landing platform, and generate the landing path of the UAV to the landing platform based on the spatial coordinate information of the landing platform.

[0008] This invention is used for the precise landing of drones on a landing platform. In specific applications, the drone flies to a distance of 10m from the logistics landing platform, slowly descends to a position below 5m, and hovers. The drone's terminal module searches for the positioning guidance device, and the orientation perception module transmits the perceived drone position information to the controller. The controller controls the controllable light-emitting unit to generate dynamic visual guidance signals. Based on the identified dynamic visual signals from the positioning guidance device, the drone's terminal module determines its orientation and calculates the spatial coordinates of the landing platform. Based on the spatial coordinates of the landing platform, it uses a pre-trained model to generate a landing path from the drone to the landing platform and controls the drone to land on the landing platform.

[0009] This invention solves the positioning problem when there is no GPS signal or the sensor is blocked by the dynamic guidance of the visual positioning and controllable light-emitting system, and improves the system's adaptability in complex environments, especially suitable for cargo-carrying scenarios of logistics drones.

[0010] As a preferred further technical solution, the positioning guidance device includes a columnar body, the independent controllable light-emitting system includes multiple sets of light-emitting units, the light-emitting units are arranged at equal heights on each side of the columnar body, the orientation sensing module is a radar array, used to acquire the altitude and orientation information of the UAV and transmit it to the controller, and the controller controls and adjusts the brightness or on / off state of the light-emitting units.

[0011] In practice, the positioning and guidance device consists of a cylindrical body made of aluminum alloy, a four-sided prism, erected on one side of the landing platform, with a height of 3-5 meters and a width of 0.5-1 meter, maintaining a safe distance from the landing platform. The orientation sensing module is a UAV detection millimeter radar array, fixed at the four corners of the top surface of the cylindrical body, with the sensing direction installed at 180° to the top diagonal. It includes four radars, with a sensing distance of 10 meters, an accuracy of 0.1 meters, a sensing angle range of >90°, and an angle accuracy of 0.5°. The controllable lighting system consists of 10-watt LED positioning lights, arranged in two equal rows on the four sides of the cylindrical body. The controller is located inside the cylindrical body. It receives the UAV position information transmitted by the radar and controls the corresponding positioning lights to turn on and off. Specifically, the row of positioning lights facing the UAV and closest to the UAV remains constant, while the other positioning lights are turned off to avoid interference and save energy.

[0012] As a preferred further technical solution, the light-emitting units on each side of the columnar main body emit different colors, and the space around the positioning guidance device is divided into multiple spatial regions based on the light-emitting colors. The UAV terminal module pre-stores the relative positional relationship between each spatial region and the landing platform. The UAV terminal module includes a deep learning neural network for recognizing the color and coordinate features of the light-emitting units, uses a positioning algorithm to calculate the spatial coordinates of the positioning guidance device relative to the UAV, and calculates the coordinates of the landing platform in the UAV coordinate system based on the determined spatial region.

[0013] In practical application, the positioning lights on the same side of the columnar body are the same color, with a total of four positioning light colors on all four sides. The diagonal of the positioning column divides the space into four regions A, B, C, and D, with the corresponding positioning light colors of the columnar body being A, B, C, and D. Using the columnar body as the origin and a coordinate system with north at the top and south at the bottom, region A is located to the west, which is the location of the positioning column; region B is located to the north; region C is located to the east; and region D is located to the south. By recognizing the positioning light colors, the UAV can determine its own location and then calculate the orientation of the landing platform based on the pre-stored relative positional relationships between each spatial region and the landing platform.

[0014] This invention links multi-color positioning lights with radar, enabling drones to determine their location without direct observation of the platform through color-area positioning. This solves the problem of sensor field of view obstruction when carrying cargo and reduces reliance on external positioning signals.

[0015] As a preferred further technical solution, the landing platform is provided with a reset actuator, which includes a slide rail, a sliding component, a drive component, a control unit, and a displacement monitoring sensor. The sliding component includes a vertically arranged first sliding component and a second sliding component. The drive component is used to drive the first sliding component and the second sliding component to move along the slide rail. The displacement monitoring sensor is used to detect the displacement information of the sliding component and feed back the positional deviation of the sliding component from the center of the landing platform to the control unit in real time. The control unit controls the drive component to drive the sliding component to push the UAV and / or logistics items to the center of the landing platform.

[0016] In practical applications, this mainly targets logistics drones carrying packages. The landing platform can be square, made of aluminum alloy, and fixed to the ground; the slide rail is made of non-magnetic aluminum alloy and symmetrically fixed to the four sides of the landing platform with screws; the drive component is a servo motor that can slide in the slide rail and drive the first and second sliding components to move along the slide rail, and the sliding components are made of lightweight aluminum alloy; the displacement monitoring sensor can be a laser rangefinder sensor, fixed in the middle of the slide rail, used to measure the distance moved by the first and second sliding components, with an accuracy within 0.1mm. The control unit can be an embedded PID controller.

[0017] The first and second sliding components of the sliding assembly form a grid-like sliding reset mechanism for precise reset. When the drone lands above the landing platform, the logistics landing platform uses a reset actuator to push the logistics component or drone beneath it, returning the logistics component and drone to the exact center of the logistics landing platform, after which the sliding bar resets.

[0018] To achieve the above objectives, the present invention also provides a visual positioning-based method for precise landing of unmanned aerial vehicles (UAVs), which is executed using any of the visual positioning-based UAV precise landing systems described above, and includes the following steps: S1, Dynamic visual guidance signal generation: The orientation perception module identifies the UAV's position information and transmits it to the controller. The controller controls the controllable light-emitting unit to generate a dynamic visual guidance signal based on the UAV's position information. S2, Spatial coordinate calculation: The UAV module determines the orientation and coordinates of the logistics landing platform relative to the positioning guidance device based on the dynamic visual guidance signal, and calculates the spatial coordinates of the landing platform by combining the regional positioning algorithm. S3, Path Planning and Landing: Using information including the spatial coordinates of the landing platform as input, calculate the control variables of the UAV, plan the landing path, and control the UAV to land on the landing platform.

[0019] As a preferred further technical solution, step S3 is further included as a sliding reset step: after the UAV lands above the landing platform, the reset actuator pushes the logistics component and / or the UAV to return the logistics component and / or the UAV to the middle of the landing platform, and then the sliding component reset is performed.

[0020] In practical applications, once the drone lands above the landing platform, the system immediately triggers a reset command: the high-frequency sampling rate of the laser rangefinder sensor monitors the displacement data of the first slide bar (lateral slide bar) and the second slide bar (longitudinal slide bar) in real time, generating a dynamic position deviation signal relative to the geometric center of the landing platform; this signal is input to the embedded PID controller to calculate the servo motor control quantity in real time; the servo motor performs the reset in two stages—first, it drives the lateral slide bar along the X-axis to push the logistics item to a deviation of ≤0.5mm, and after the electromagnetic lock of the lateral slide bar is pre-locked, it then starts the longitudinal slide bar along the Y-axis to complete the reset with the same precision; the entire process uses a closed-loop feedback mechanism to dynamically adjust the slide bar speed. After the reset is completed, the electromagnetic lock built into the slide bar assembly is automatically activated, applying a strong 200N adsorption force to fix the cargo and resist the airflow disturbance during takeoff.

[0021] As a preferred further technical solution, in step S1, the dynamic visual guidance signal is based on the altitude and orientation of the UAV, keeping the group of light-emitting units closest to the UAV constantly lit, while turning off or reducing the brightness of the remaining light-emitting units; the area positioning algorithm divides the space into different areas based on the color of the light-emitting units; in step S2, the UAV-side module pre-stores the position mapping relationship between each area and the landing platform; the UAV-side module identifies the color and coordinate characteristics of the light-emitting units; uses the positioning algorithm to calculate the spatial coordinates of the positioning guidance device relative to the UAV; determines the spatial area based on the color of the light-emitting units; calls the pre-stored position mapping relationship between each area and the landing platform; and calculates the coordinates of the landing platform in the UAV coordinate system through a coordinate transformation matrix.

[0022] As described above, when the orientation perception module transmits the perceived UAV position information to the controller, the row of positioning lights that are facing the UAV and are closest to the UAV at the same altitude will remain on, while the remaining positioning lights will be turned off or have their brightness reduced. After the UAV recognizes the color and coordinate characteristics of the positioning lights, it determines the position of the landing platform based on the pre-stored position mapping relationship between each area and the landing platform. It then uses a positioning algorithm to calculate the spatial coordinates of the positioning guidance device relative to the UAV, calculates the coordinates of the landing platform in the UAV coordinate system through a coordinate transformation matrix, and generates control commands to guide the landing.

[0023] As a preferred further technical solution, in step S3, the spatial coordinates of the landing platform, the UAV status information and environmental obstacle information are input into the reinforcement learning model, and the motion control vector and acceleration control vector of the UAV are output and updated step by step to guide the UAV to land on the landing platform.

[0024] After determining the coordinates of the landing platform in the drone's coordinate system, the logistics drone will transfer the coordinates of the logistics landing platform to the system. drone location UAV end module line-of-sight direction vector Environmental information The drone's current velocity vector The current acceleration vector of the drone As input, the control variables for the UAV at the next moment, including the velocity vector, are obtained through a pre-trained actor-critic (T-Actor-Critic) reinforcement learning model. and acceleration vector The next moment, the drone updates the system coordinates of the logistics landing platform, obtains motion planning again through reinforcement learning algorithm, and controls the drone to gradually fly above the logistics landing platform.

[0025] Thus, the T-Actor-Critic reinforcement learning model is used to generate speed and acceleration control variables, enabling fully autonomous landing. Even when communication is interrupted, the operation can still be completed through visual positioning, improving path planning efficiency and possessing strong generalization ability.

[0026] As a preferred embodiment, a further technical solution is that the reinforcement learning model includes a main decision channel and a confidence evaluation channel. The main decision channel is used to generate the motion control vector and acceleration control vector of the UAV. The confidence evaluation channel is equipped with a visual positioning quality evaluation module and is used to evaluate the visual positioning quality in real time. When the visual positioning confidence is greater than or equal to a predetermined threshold, the main decision channel action is directly output. When the visual positioning confidence is less than the predetermined threshold, the main decision channel action and the conservative safety action are fused according to the confidence weight.

[0027] This invention introduces a dual-channel mechanism: a main decision-making channel and a confidence evaluation channel. The main decision-making channel, as described above, is the conventional Actor-Critic structure, responsible for generating flight commands under normal circumstances. The confidence evaluation channel is implemented using a Long Short-Term Memory (LSTM) network. The LSTM network can process time-series data, taking into account image features from multiple consecutive frames, the number of location lights identified, image blurriness, and the angular velocity of the UAV's inertial measurement unit (IMU). This information comprehensively reflects the quality of visual positioning and the stability of the UAV's motion state. After training, the LSTM network outputs a confidence score between 0 and 1, representing the reliability of the current visual positioning information. When the confidence score is higher than a set threshold, the model primarily relies on the actions generated by the main decision-making channel for flight control; when the confidence score is lower than the threshold, the model will fuse the main channel actions with conservative safety actions generated based on Model Predictive Control (MPC) according to certain weights, thereby ensuring that the UAV can still fly safely and attempt to find a landing path even when visual positioning is inaccurate or interfered with.

[0028] The main channel of this invention generates regular actions based on the Actor-Critic algorithm, while the confidence evaluation channel calculates a visual positioning quality score using an LSTM network. If the score is less than 0.7, the main channel actions are mixed with the safety actions generated by the MPC model according to the confidence weight. This dual-channel mechanism addresses scenarios such as cargo obstruction and sudden changes in lighting, preventing drone loss of control and improving the robustness and safety of path planning.

[0029] As a preferred further technical solution, the reinforcement learning model is a confidence-based contingent actor-evaluation model, which includes a reward function. For UAV path planning, the reward function is designed as a weighted sum of multiple reward items. The target approach reward, used to encourage the drone to approach the landing platform, is calculated using the following formula:

[0030] in, Location of the landing platform Location of the drone. This represents the unit direction vector of the UAV module relative to the target. Rewards for approaching the goal The coefficient is greater than 0; To ensure the smoothness of the logistics drone's trajectory planning, a trajectory smoothness reward is added, using the difference between the acceleration control vectors from the previous and current time steps as the cost. The trajectory smoothness reward, used to penalize abrupt acceleration changes, is calculated using the following formula:

[0031] in, for The acceleration vector of the drone at any time. for The acceleration vector of the drone at any time. This represents the unit direction vector of the UAV module relative to the target. For trajectory smoothing reward, The coefficient is greater than 0; To avoid drone collisions, an obstacle avoidance reward / penalty system is implemented. A collision penalty is applied when a collision occurs, while the reward increases the further the drone is from the obstacle when no collision occurs. The obstacle avoidance reward encourages the drone to maintain a safe distance, and its calculation formula is as follows:

[0032] in, For trajectory smoothing reward, As a penalty for collision, All are coefficients, and all are greater than 0. Location of the drone. This is the location information of the obstacle.

[0033] To ensure the positioning post remains within the drone's line of sight as much as possible, a line-of-sight alignment bonus is added. This bonus is calculated by taking the inner product of the drone's optical axis and the line connecting the drone and the positioning post. The line-of-sight alignment bonus guides the line of sight towards the landing platform, and its calculation formula is as follows:

[0034] in, For trajectory smoothing reward, The coefficient is greater than 0; The unit direction of the UAV's line of sight; The formula for calculating the comprehensive reward is:

[0035] in, For comprehensive rewards, These are the weighting coefficients for each item, and they are greater than 0.

[0036] The reward function, as the mathematical expression of the optimization objective, drives the iterative learning of the Actor-Critic network, enabling the drone's action strategy to gradually converge from random exploration to the optimal path planning strategy. In practical applications, the optimization objective is defined through the reward function construction step, and the reinforcement learning decision step iteratively updates the strategy based on the reward function, causing the drone's actions to tend towards maximizing the cumulative reward. Furthermore, the weight coefficients can be adjusted... Balancing the priority of various rewards, such as increasing them in complex obstacle environments. (Obstacle avoidance weight) is increased in areas where visual positioning is weak. (Focus your gaze on the weights.)

[0037] Compared with existing technologies, the implementation of the technical solution of this invention has the following technical advantages: Enhanced autonomous and precise landing capability: A multi-layered and multi-color controllable light-emitting system (positioning light) provides guidance, and combined with positioning algorithms, the strong correlation between the UAV and the landing platform is transformed into a weak correlation, enabling the UAV to complete autonomous positioning through visual information. This solves the problem that traditional landing systems cannot complete precise landings in environments with no signal.

[0038] By maximizing the use of the drone's visual information, the drone can achieve visual positioning capabilities without relying on traditional bottom sensors when carrying cargo. Compared to traditional methods, the drone structure is simple and lightweight.

[0039] Reinforcement learning path planning: Based on an improved reinforcement learning model, the path planning efficiency of logistics drones is greatly accelerated, and it also has strong generalization ability to cope with various scenarios.

[0040] Precision Reset System: The dynamic collaborative control sliding component works in conjunction with the laser positioning sensor and the drive motor to enable the drone to quickly return to its original position after landing, reducing in-flight adjustment time and improving safety. Attached Figure Description

[0041] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0042] Figure 1 This is a schematic diagram showing the arrangement of the landing platform and positioning guidance device of a vision-based unmanned aerial vehicle (UAV) precision landing system according to one embodiment of the present invention. Figure 2 This is a schematic diagram of the landing platform according to one embodiment of the present invention; Figure 3 This is a schematic diagram of the arrangement structure of a positioning and guiding device according to one embodiment of the present invention; Figure 4 This is a flowchart illustrating a vision-based positioning-based precision landing system for unmanned aerial vehicles according to one embodiment of the present invention. Figure 5 A schematic diagram of the spatial area surrounding a positioning guide device for determining the color of a light-emitting unit according to one embodiment of the present invention. Figure 6 This is a schematic diagram showing the spatial region where the color of the light-emitting unit is determined according to one embodiment of the present invention and the positional mapping relationship between the landing platform and the location of the light-emitting unit. Figure 7 This is a schematic diagram illustrating the calculation principle of the triangulation method according to one embodiment of the present invention.

[0043] In the picture: 1. Landing platform 2. Positioning and guidance device 3. Landing marker 4. First sliding bar 5 Second slide bar 6 Drive component 7 Displacement monitoring sensor 8 Slide rail 9 Controllable light emission system 10 Orientation sensing module 11 Columnar body 12 Unmanned aerial vehicle Detailed Implementation

[0044] The present invention will now be described in detail with reference to the accompanying drawings. This description is merely illustrative and explanatory, and should not be construed as limiting the scope of protection of the present invention. Furthermore, those skilled in the art can combine the features in the embodiments described herein and in different embodiments accordingly based on the description in this document.

[0045] The embodiments of the present invention are as follows, with reference to Figures 1-6 A visual positioning-based precision landing system for UAVs includes: Landing Platform 1: Used for landing UAV 12; Positioning and guidance device 2: Located on one side of the landing platform 1, it includes multiple independent controllable light-emitting systems 9 arranged in layers, a position sensing module 10 for sensing the position of the UAV 12, and a controller. The position sensing module 10 transmits the sensed position information of the UAV 12 to the controller, and the controller is used to control the controllable light-emitting units to generate dynamic visual guidance signals. The UAV terminal module is used to identify the dynamic visual signals of the positioning and guidance device 2, calculate the spatial coordinates of the landing platform 1, and generate the landing path of the UAV 12 to the landing platform 1 based on the spatial coordinate information of the landing platform 1.

[0046] This invention is used for the precise landing of a drone 12 on a landing platform 1. In specific application, the drone 12 flies to a range of 10m from the logistics landing platform 1, slowly descends to a position below 5m, and hovers. The drone-end module searches for the positioning guidance device 2, and the orientation perception module 10 transmits the perceived position information of the drone 12 to the controller. The controller is used to control the controllable light-emitting unit to generate dynamic visual guidance signals. The drone-end module determines the orientation and calculates the spatial coordinates of the landing platform 1 based on the identified dynamic visual signals from the positioning guidance device 2. Based on the spatial coordinate information of the landing platform 1, it uses a pre-trained model to generate a landing path from the drone 12 to the landing platform 1 and controls the drone 12 to land on the landing platform 1.

[0047] This invention solves the positioning problem when there is no GPS signal or the sensor is blocked by visual positioning and dynamic guidance of the controllable light-emitting system 9, and improves the system's adaptability in complex environments, especially suitable for cargo-carrying scenarios of logistics drones 12.

[0048] Based on the above embodiments, in another embodiment of the present invention, such as... Figure 3 The positioning and guidance device 2 includes a columnar body, the independent controllable light-emitting system 9 includes multiple light-emitting units, the light-emitting units are arranged at equal heights on each side of the columnar body, the orientation sensing module 10 is a radar array, used to acquire the altitude and orientation information of the UAV 12 and transmit it to the controller, the controller controls and adjusts the brightness or on / off state of the light-emitting units.

[0049] In the specific implementation process, such as Figure 3 The positioning and guidance device 2 has a columnar main body made of aluminum alloy, which is a four-sided prism. It stands on one side of the landing platform 1, with a height of 3-5 meters and a width of 0.5-1 meter, and there is a safe distance between it and the landing platform 1. The orientation sensing module 10 is a millimeter radar array for detecting the UAV 12. It is fixed at the four corners of the top surface of the columnar main body. The sensing direction is installed at 180° with the diagonal of the top surface. It includes four radars, with a sensing distance of 10m, an accuracy of 0.1m, a sensing angle range of >90°, and an angle accuracy of 0.5°. The controllable light-emitting system 9 is a 10-watt LED positioning light, which is arranged in two rows at equal intervals on the four sides of the columnar main body. The controller is set inside the columnar main body. It receives the UAV 12 position information transmitted by the radar and controls the corresponding positioning lights to turn on and off. Specifically, the row of positioning lights facing the UAV 12 and the closest to the UAV 12 is kept constant, while the other positioning lights are turned off to avoid interference and save energy.

[0050] Based on the above embodiments, in another embodiment of the present invention, such as... Figure 3 The light-emitting units on each side of the columnar main body emit different colors. Based on the light-emitting colors, the space around the positioning guidance device 2 is divided into multiple spatial regions. The UAV terminal module pre-stores the relative positional relationship between each spatial region and the landing platform 1. The UAV terminal module includes a deep learning neural network for recognizing the color and coordinate features of the light-emitting units, uses a positioning algorithm to calculate the spatial coordinates of the positioning guidance device 2 relative to the UAV 12, and calculates the coordinates of the landing platform 1 in the coordinate system of the UAV 12 based on the determined spatial region.

[0051] In practical application, the positioning lights on the same side of the columnar body are the same color, with a total of four positioning light colors on all four sides. The diagonal of the positioning column divides the space into four regions A, B, C, and D, with the corresponding positioning light colors of the columnar body 11 being A, B, C, and D. Taking the columnar body 11 as the origin and using a north-south coordinate system, region A is located in the west, which is the location of the positioning column; region B is located in the north; region C is located in the east; and region D is located in the south. The UAV 12 can determine its own region by recognizing the positioning light colors, and then calculate the orientation of the landing platform 1 based on the pre-stored relative positional relationships between each spatial region and the landing platform 1.

[0052] This invention uses a multi-color positioning light linked with radar to enable the UAV 12 to determine its position without direct observation of the platform through a color area positioning method. This solves the problem of sensor field of view obstruction when carrying cargo and reduces dependence on external positioning signals.

[0053] Based on the above embodiments, in another embodiment of the present invention, such as... Figure 2 The landing platform 1 is equipped with a reset actuator, which includes a slide rail 8, a sliding assembly, a drive component 6, a control unit, and a displacement monitoring sensor 7. The sliding assembly includes a vertically arranged first sliding member and a second sliding member. The drive component 6 is used to drive the first sliding member and the second sliding member to move along the slide rail 8. The displacement monitoring sensor 7 is used to detect the displacement information of the sliding assembly and provide real-time feedback on the positional deviation between the sliding assembly and the center of the landing platform 1 to the control unit. The control unit controls the drive component 6 to drive the sliding assembly, thereby pushing the UAV 12 and / or logistics components to the center of the landing platform 1.

[0054] In specific applications, this mainly targets logistics drones 12 carrying logistics items. The landing platform 1 can be square, made of aluminum alloy, fixed to the ground, and equipped with landing markers 3. The slide rail 8 is made of non-magnetic aluminum alloy and is symmetrically fixed to the four sides of the landing platform 1 using screws. The drive component 6 is a servo motor that can slide within the slide rail 8 and drive the first and second sliding components to move along the slide rail 8; the sliding components are made of lightweight aluminum alloy. The displacement monitoring sensor 7 can be a laser rangefinder sensor, fixed in the middle of the slide rail 8, used to measure the distance moved by the first and second sliding components, with an accuracy within 0.1mm. The control unit can be an embedded PID controller.

[0055] The first and second sliding members of the sliding assembly form a grid-like sliding reset mechanism for precise reset. When the drone 12 lands above the landing platform 1, the logistics landing platform 1 uses a reset actuator to push the logistics component or drone 12 under the drone 12, returning the logistics component and drone 12 to the center of the logistics landing platform 1, and then the sliding bar resets.

[0056] This invention also provides a visual positioning-based method for precise landing of a UAV 12, as illustrated below. The method is executed using any of the visual positioning-based UAV 12 precise landing systems described above. Figure 4 It includes the following steps: S1, Dynamic visual guidance signal generation: The orientation perception module 10 identifies the position information of the UAV 12 and transmits it to the controller. The controller controls the controllable light-emitting unit to generate a dynamic visual guidance signal according to the position information of the UAV 12. Specifically: The UAV 12 flies to a distance of 10m from the landing platform 1, slowly descends to a position below 5m, hovers, rotates in place to find the positioning guidance device 2, and flies in the direction of the positioning guidance device 2. After the UAV 12 enters the range of the orientation perception module 10 (radar), the orientation perception module 10 identifies the position of the UAV 12, and the controller controls the controllable light-emitting unit to generate a dynamic visual guidance signal. The dynamic visual guidance signal is based on the altitude and orientation of the UAV 12, so that the group of light-emitting units closest to the UAV 12 remains constantly lit, while the other light-emitting units are turned off or have their brightness reduced to avoid interference and save energy.

[0057] As mentioned above, when the orientation sensing module 10 transmits the perceived location information of the drone 12 to the controller, the row of positioning lights that are oriented towards the drone 12 and are closest to the drone 12 at the same height will remain on, while the remaining positioning lights will be turned off or have their brightness reduced.

[0058] S2, Spatial coordinate calculation: The UAV module determines the orientation and coordinates of the logistics landing platform 1 relative to the positioning guidance device 2 based on the dynamic visual guidance signal, and calculates the spatial coordinates of the landing platform 1 by combining the regional positioning algorithm. Specifically, after the UAV 12 identifies the color and coordinate characteristics of the positioning light, it determines the position of the landing platform 1 based on the pre-stored position mapping relationship between each area and the landing platform 1, calculates the spatial coordinates of the positioning guidance device 2 relative to the UAV 12 using a positioning algorithm, calculates the coordinates of the landing platform 1 in the coordinate system of the UAV 12 through a coordinate transformation matrix, and generates control commands to guide the landing.

[0059] like Figure 5 In the area positioning algorithm, the space is divided into different areas based on the color of the light-emitting unit. The UAV terminal module pre-stores the position mapping relationship between each area and the landing platform 1. The UAV terminal module identifies the color and coordinate features of the light-emitting unit, uses the positioning algorithm to calculate the spatial coordinates of the positioning guidance device 2 relative to the UAV 12, and determines the spatial area based on the color of the light-emitting unit. It calls the pre-stored position mapping relationship between each area and the landing platform 1, and calculates the coordinates of the landing platform 1 in the coordinate system of the UAV 12 through the coordinate transformation matrix.

[0060] The drone 12 uses a deep learning convolutional neural network to extract visual features and, through a YOLO pre-trained model, identifies the position of the positioning guidance device 2 using localization and other methods. Deep learning algorithms improve visual recognition accuracy, and the color region localization method allows the drone 12 to determine its position without directly observing the platform, solving the problem of sensor field-of-view obstruction during cargo loading. The pre-stored mapping relationship between each region and the landing platform 1 is as follows: Figure 5 and Figure 6 The space is divided into four areas, A, B, C, and D, along the diagonal of the positioning post. The corresponding positioning lights are the colors A, B, C, and D. The center of the column body 11 is the origin, and a north-south coordinate system is used. Area A is located in the west, which is the orientation of the positioning post; area B is located in the north; area C is located in the east; and area D is located in the south.

[0061] When drone 12 is located in area A, such as Figure 6 As shown in -a, the logistics landing platform 1 is located on the line connecting the drone 12 and the positioning post, with a distance dis from the positioning post.

[0062] When drone 12 is located in area B, such as Figure 6 As shown in -b, the logistics landing platform 1 is located on the perpendicular line connecting the UAV 12 and the positioning post, with the perpendicular line falling in area A and the distance from the positioning post being a distance dis.

[0063] When drone 12 is located in area C, such as Figure 6 As shown in -c, the logistics landing platform 1 is located on the extension line of the line connecting the drone 12 and the positioning post. The direction of the extension line falls in area A, and the distance from the positioning post is a distance dis.

[0064] When drone 12 is located in area D, such as Figure 6 As shown in -d, the logistics landing platform 1 is located on the perpendicular line connecting the UAV 12 and the positioning post, with the perpendicular line falling in area A and the distance from the positioning post being a distance dis.

[0065] Where dis is the distance between the geometric center of the positioning column and the geometric center of the logistics landing platform 1 in the top view.

[0066] Using this method, the drone 12 does not need to see the logistics landing platform 1; it can determine the approximate location of the logistics landing platform 1 simply by the position of the positioning post and the color of the positioning light on the positioning post.

[0067] The UAV module estimates its own pose using a triangulation algorithm. It calculates the spatial coordinates of the positioning guidance device 2 relative to the UAV 12, and then uses a coordinate transformation matrix to calculate the coordinates of the landing platform 1 in the UAV 12 coordinate system. Control commands are then generated to guide the landing. The specific calculation principle is as follows: Figure 7The drone-side module includes a binocular stereo vision model, as shown in the figure. The schematic diagram of the binocular stereo vision model has a focal length of f, and the distance between the optical centers of the two camera lenses is the baseline distance B. These two cameras are located on the same plane. The same feature point in the left space of the object was captured at the same time. Image points of point P were acquired from the left and right cameras respectively. And their projection centers have the same image coordinate Y value, that is These relationships are as follows: Figure 7 As shown.

[0068] From trigonometric relations, we can obtain:

[0069] Parallax D is The distance between two points, i.e. The three-dimensional coordinates of point P in the head-up model system are as follows:

[0070] In this way, the coordinates of the positioning and guidance device 2 can be calculated, and the coordinates of the landing platform 1 in the coordinate system of the UAV 12 can be calculated by using a coordinate transformation matrix based on the determined spatial area.

[0071] S3, Path planning and landing: Using information including the spatial coordinates of landing platform 1 as input, calculate the control quantity of UAV 12, plan the landing path, and control UAV 12 to land on landing platform 1.

[0072] Specifically, the spatial coordinates of the landing platform 1, the state information of the UAV 12, and the information of environmental obstacles are input into the reinforcement learning model, and the motion control vector and acceleration control vector of the UAV 12 are output and updated step by step to guide the UAV 12 to land on the landing platform 1.

[0073] After determining the coordinates of landing platform 1 in the coordinate system of UAV 12, the logistics UAV 12 will transfer the coordinates of landing platform 1 to the coordinate system. 12 drone positions UAV end module line-of-sight direction vector Environmental information The current velocity vector of UAV 12 The current acceleration vector of UAV 12 As input, the control variables of UAV 12 at the next moment, including the velocity vector, are obtained through a pre-trained actor-critic (T-Actor-Critic) reinforcement learning model. and acceleration vector In the next moment, the drone 12 updates the system coordinates of the logistics landing platform 1, obtains the motion plan again through the reinforcement learning algorithm, and controls the drone 12 to gradually fly above the logistics landing platform 1.

[0074] Specifically: The reinforcement learning model is a confidence-based contingency actor-critic model, and the confidence-based contingency actor-critic (T-Actor-Critic) reinforcement learning model is described below: Model building: The logistics drone's status is 12:

[0075] in, 12 drone locations: Velocity and acceleration; Unit direction of the UAV's line of sight (unit vector); Information about surrounding obstacles; The 12 actions of the logistics drone defined in the reinforcement learning model are:

[0076] in, Speed / acceleration commands; : View axis direction rotation command; construct the state transition model according to the discretization method:

[0077] in For time step, This represents vector normalization.

[0078] Thus, the T-Actor-Critic reinforcement learning model is used to generate speed and acceleration control variables, enabling fully autonomous landing. Even when communication is interrupted, the operation can still be completed through visual positioning, improving path planning efficiency and possessing strong generalization ability.

[0079] The T-Actor-Critic reinforcement learning model includes a main decision channel and a confidence evaluation channel. The main decision channel is used to generate the motion control vector and acceleration control vector of the UAV 12. The confidence evaluation channel is equipped with a visual positioning quality evaluation module and is used to evaluate the visual positioning quality in real time. When the visual positioning confidence is greater than or equal to a predetermined threshold, the main decision channel action is directly output; when the visual positioning confidence is less than the predetermined threshold, the main decision channel action and the conservative safety action are fused according to the confidence weight.

[0080] To address the vision loss issues caused by cargo obstruction, sudden changes in lighting, or excessive deviation of the UAV 12's line of sight, and to prevent the UAV 12 from going out of control, this patent employs a dual-channel reinforcement learning algorithm, the specific design of which is shown in Table 1:

[0081] Confidence-driven action fusion mechanism

[0082] The above confidence level The quality assessment is calculated by the visual positioning module, which is an LSTM network model with the following inputs and outputs: Input: Features of 5 consecutive frames of images (YOLOv8 output), number of positioning lights recognized, image blur, IMU angular velocity.

[0083] Output: Confidence level The confidence level indicates the credibility of the drone's actions at this moment.

[0084] This invention introduces a dual-channel mechanism: a main decision-making channel and a confidence evaluation channel. The main decision-making channel, as described above, is the conventional Actor-Critic structure, responsible for generating flight commands under normal circumstances. The confidence evaluation channel is implemented using a Long Short-Term Memory (LSTM) network. The LSTM network can process time-series data, taking into account image features from multiple consecutive frames, the number of location lights identified, image blurriness, and the angular velocity of the UAV's inertial measurement unit (IMU). This information comprehensively reflects the quality of visual positioning and the stability of the UAV's motion state. After training, the LSTM network outputs a confidence score between 0 and 1, representing the reliability of the current visual positioning information. When the confidence score is higher than a set threshold, the model primarily relies on the actions generated by the main decision-making channel for flight control; when the confidence score is lower than the threshold, the model will fuse the main channel actions with conservative safety actions generated based on Model Predictive Control (MPC) according to certain weights, thereby ensuring that the UAV can still fly safely and attempt to find a landing path even when visual positioning is inaccurate or interfered with.

[0085] The main channel of this invention generates regular actions based on the Actor-Critic algorithm, while the confidence evaluation channel calculates a visual positioning quality score using an LSTM network. If the score is less than 0.7, the main channel actions are mixed with the safety actions generated by the MPC model according to the confidence weight. In this way, a dual-channel mechanism is used to address scenarios such as cargo obstruction and sudden changes in lighting, preventing the UAV from going out of control and improving the robustness and safety of path planning.

[0086] The confidence-based contingency actor-evaluation model includes a reward function. For UAV 12-path planning, the reward function is designed as a weighted sum of multiple reward items: The target approach reward, used to encourage drone 12 to approach landing platform 1, is calculated using the following formula:

[0087] in, This is the location of landing platform 1. For the 12 positions of the drone, This represents the unit direction vector of the UAV module relative to the target. Rewards for approaching the goal The coefficient is greater than 0; To ensure the smoothness of the trajectory planning for the logistics drone, a trajectory smoothness bonus is added, using the difference between the acceleration control vectors from the previous and current time steps as the cost. The trajectory smoothness bonus, used to penalize abrupt acceleration changes, is calculated using the following formula:

[0088] in, for The acceleration vector of the Time Drone 12 for The acceleration vector of the Time Drone 12 This represents the unit direction vector of the UAV module relative to the target. For trajectory smoothing reward, The coefficient is greater than 0; To avoid collisions with drone 12, an obstacle avoidance reward / penalty system is implemented. A collision penalty is applied when a collision occurs, while the reward increases the further drone 12 is from the obstacle when no collision occurs. The obstacle avoidance reward encourages drone 12 to maintain a safe distance, and its calculation formula is as follows:

[0089] in, For trajectory smoothing reward, l represents the collision penalty. All are coefficients, and all are greater than 0. For the 12 positions of the drone, This refers to the location information of the obstacle. To ensure the positioning post remains within the line of sight of UAV 12 as much as possible, a line-of-sight alignment bonus is added. This bonus is calculated by taking the inner product of the optical axis of UAV 12 and the line connecting UAV 12 and the positioning post. The line-of-sight alignment bonus guides the line of sight towards landing platform 1, and its calculation formula is as follows:

[0090] in, For trajectory smoothing reward, The coefficient is greater than 0; The unit direction of the UAV's line of sight; The formula for calculating the comprehensive reward is:

[0091] in, For comprehensive rewards, These are the weighting coefficients for each item, and they are greater than 0.

[0092] The reward function, as the mathematical expression of the optimization objective, drives the Actor-Critic network to iteratively learn, enabling the UAV's 12-action policy to gradually converge from random exploration to the optimal path planning policy. In practical applications, the optimization objective is defined through the reward function construction step, and the reinforcement learning decision step iteratively updates the policy based on the reward function, making the UAV's 12 actions tend to maximize the cumulative reward. Furthermore, the weight coefficients can be adjusted... Balancing the priority of various rewards, such as increasing them in complex obstacle environments. (Obstacle avoidance weight) is increased in areas where visual positioning is weak. (Focus your gaze on the weights.)

[0093] S4, Sliding Reset: After the UAV 12 lands above the landing platform 1, the reset actuator pushes the UAV 12 and / or the logistics component to return the logistics component and / or the UAV 12 to the middle of the landing platform 1, and then performs the sliding component reset.

[0094] Once the UAV 12 lands above the landing platform 1, the system immediately triggers a reset command: the high-frequency sampling rate of the laser rangefinder sensor monitors the displacement data of the first slide bar 4 (lateral slide bar) and the second slide bar 5 (longitudinal slide bar) in real time, generating a dynamic position deviation signal relative to the geometric center of the landing platform 1; this signal is input to the embedded PID controller to calculate the servo motor control quantity in real time; the servo motor performs the reset in two stages—first, it drives the lateral slide bar along the X-axis to push the logistics item to a deviation ≤0.5mm, and after the electromagnetic lock of the lateral slide bar is pre-locked, it then starts the longitudinal slide bar along the Y-axis to complete the reset with the same precision; the entire process uses a closed-loop feedback mechanism to dynamically adjust the slide bar speed. After the reset is completed, the electromagnetic lock built into the slide bar assembly is automatically activated, applying a strong 200N adsorption force to fix the cargo and resist the takeoff airflow disturbance.

[0095] For those skilled in the art, various improvements and modifications can be made without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention.

Claims

1. A visual positioning-based precision landing system for unmanned aerial vehicles (UAVs), characterized in that, include: Landing platform: Used for landing drones; Positioning and guidance device: Located on one side of the landing platform, it includes multiple independent controllable light-emitting systems arranged in layers, an orientation perception module for sensing the position of the UAV, and a controller. The orientation perception module transmits the sensed UAV position information to the controller, and the controller is used to control the controllable light-emitting units to generate dynamic visual guidance signals. The UAV terminal module is used to identify the dynamic visual signals of the positioning and guidance device, calculate the spatial coordinates of the landing platform, and generate the landing path of the UAV to the landing platform based on the spatial coordinate information of the landing platform.

2. The visual positioning-based UAV precision landing system according to claim 1, characterized in that, The positioning and guidance device includes a columnar body, and the independent controllable light-emitting system includes multiple light-emitting units. The light-emitting units are arranged at equal heights on each side of the columnar body. The orientation sensing module is a radar array used to acquire the altitude and orientation information of the UAV and transmit it to the controller. The controller controls and adjusts the brightness or on / off state of the light-emitting units.

3. The visual positioning-based UAV precision landing system according to claim 2, characterized in that, The light-emitting units on each side of the columnar main body emit different colors. Based on the light-emitting colors, the space around the positioning and guidance device is divided into multiple spatial regions. The UAV terminal module pre-stores the relative positional relationship between each spatial region and the landing platform. The UAV terminal module includes a deep learning neural network for recognizing the color and coordinate features of the light-emitting units, uses a positioning algorithm to calculate the spatial coordinates of the positioning and guidance device relative to the UAV, and calculates the coordinates of the landing platform in the UAV coordinate system based on the determined spatial region.

4. The visual positioning-based UAV precision landing system according to any one of claims 1 to 3, characterized in that, The landing platform is equipped with a reset actuator, which includes a slide rail, a sliding assembly, a drive component, a control unit, and a displacement monitoring sensor. The sliding assembly includes a vertically arranged first sliding member and a second sliding member. The drive component drives the first and second sliding members to move along the slide rail. The displacement monitoring sensor detects the displacement information of the sliding assembly and feeds back the positional deviation of the sliding assembly from the center of the landing platform to the control unit in real time. The control unit controls the drive component to move the sliding assembly, thereby pushing the drone and / or logistics items to the center of the landing platform.

5. A method for precise landing of a drone based on visual positioning, characterized in that, The precise landing system for unmanned aerial vehicles (UAVs) based on visual positioning, as described in any one of claims 1 to 4, comprises the following steps: S1, Dynamic visual guidance signal generation: The orientation perception module identifies the UAV's position information and transmits it to the controller. The controller controls the controllable light-emitting unit to generate a dynamic visual guidance signal based on the UAV's position information. S2, Spatial coordinate calculation: The UAV module determines the orientation and coordinates of the logistics landing platform relative to the positioning guidance device based on the dynamic visual guidance signal, and calculates the spatial coordinates of the landing platform by combining the regional positioning algorithm. S3, Path Planning and Landing: Using information including the spatial coordinates of the landing platform as input, calculate the control variables of the UAV, plan the landing path, and control the UAV to land on the landing platform.

6. A method for precise landing of a UAV based on visual positioning, executed using the precise landing system for a UAV based on visual positioning as described in claim 4, characterized in that, The step S3 is followed by a sliding reset step: after the UAV lands above the landing platform, the reset actuator pushes the UAV and / or logistics component to return the logistics component and / or UAV to the middle of the landing platform, and then the sliding component reset is performed.

7. The method for precise landing of a UAV based on visual positioning according to claim 5 or 6, characterized in that, In step S1, the dynamic visual guidance signal is based on the drone's altitude and orientation, keeping the group of light-emitting units closest to the drone constantly lit, while turning off or reducing the brightness of the remaining light-emitting units; the area positioning algorithm divides the space into different areas based on the color of the light-emitting units; in step S2, the drone-side module pre-stores the position mapping relationship between each area and the landing platform; the drone-side module identifies the color and coordinate characteristics of the light-emitting units; uses the positioning algorithm to calculate the spatial coordinates of the positioning guidance device relative to the drone; determines the spatial area based on the color of the light-emitting units; calls the pre-stored position mapping relationship between each area and the landing platform; and calculates the coordinates of the landing platform in the drone's coordinate system through a coordinate transformation matrix.

8. The method for precise landing of a UAV based on visual positioning according to claim 7, characterized in that, In step S3, the spatial coordinates of the landing platform, the UAV status information, and the environmental obstacle information are input into the reinforcement learning model, and the motion control vector and acceleration control vector of the UAV are output and updated step by step to guide the UAV to land on the landing platform.

9. The method for precise landing of a UAV based on visual positioning according to claim 8, characterized in that, The reinforcement learning model includes a main decision channel and a confidence evaluation channel. The main decision channel is used to generate the motion control vector and acceleration control vector of the UAV. The confidence evaluation channel is equipped with a visual positioning quality evaluation module and is used to evaluate the visual positioning quality in real time. When the visual positioning confidence is greater than or equal to a predetermined threshold, the main decision channel action is directly output. When the visual positioning confidence is less than the predetermined threshold, the main decision channel action and the conservative safety action are fused according to the confidence weight.

10. The method for precise landing of a UAV based on visual positioning according to claim 9, characterized in that, The reinforcement learning model is a confidence-based contingency actor-evaluation model. This model includes a reward function, designed for UAV path planning, which is a weighted sum of multiple reward items. The target approach reward, used to encourage the drone to approach the landing platform, is calculated using the following formula: ; in, Location of the landing platform Location of the drone. This represents the unit direction vector of the UAV module relative to the target. Rewards for approaching the goal The coefficient is greater than 0; The trajectory smoothing reward, used to penalize sudden acceleration changes, is calculated using the following formula: ; in, for The acceleration vector of the drone at any time. for The acceleration vector of the drone at any time. This represents the unit direction vector of the UAV module relative to the target. The reward for trajectory smoothing is denoted as a coefficient, which is greater than 0. Obstacle avoidance rewards are used to encourage drones to maintain a safe distance, and the calculation formula is as follows: ; in, For trajectory smoothing reward, As a penalty for collision, All are coefficients, and all are greater than 0. Location of the drone. This refers to the location information of the obstacle. The line-of-sight reward is used to guide the visual axis towards the landing platform, and the calculation formula is as follows: ; in, For trajectory smoothing reward, The coefficient is greater than 0; The unit direction of the UAV's line of sight; The formula for calculating the comprehensive reward is as follows: ; in, For comprehensive rewards, These are the weighting coefficients for each item, and they are greater than 0.

Images