Water rescue unmanned aerial vehicle accurate throwing method and system based on environment perception, electronic equipment and medium

Abstract

The invention relates to a water rescue unmanned aerial vehicle accurate throwing method and system based on environmental perception, electronic equipment and a medium, and provides an automatic solution that life-saving equipment can be accurately and reliably delivered to a drowning person by establishing a throwing kinetic model. The invention aims to solve the problems of inaccurate throwing, neglect of environmental influence and dependence on manual operation in the prior art, and realizes unmanned and intelligent accurate water rescue.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle emergency rescue, and in particular to an environment perception-based precise throwing method and system for water rescue unmanned aerial vehicles, an electronic device, and a medium. BACKGROUND

[0002] Water rescue, especially the rapid response to fallen persons, is the key to saving lives. Traditional rescue methods mainly rely on lifesavers to rescue or rescue boats to approach, which has inherent shortcomings such as slow response speed, great influence of wind and waves, and high risk to rescue personnel themselves.

[0003] In recent years, unmanned aerial vehicle technology has been gradually applied to the field of water rescue due to its high mobility and rapid response capability. Existing rescue unmanned aerial vehicles mostly use a simple throwing method, that is, they fly over the fallen person and manually or through video-assisted throwing of lifesaving equipment. This method has obvious defects: first, it relies heavily on the experience and reaction of the operator, making it difficult to accurately judge in complex sea conditions; second, the unmanned aerial vehicle will drift in position when hovering due to wind and rotor airflow, and simple vertical throwing may cause the lifesaving equipment to miss the target; finally, the dynamic influence of water flow and waves on the drifting trajectory of the fallen person and lifesaving equipment is not considered, and even if the lifesaving equipment falls near the fallen person, it may quickly drift away due to the high speed of the water flow, making it difficult for the fallen person to reach it, thus missing the golden rescue time.

[0004] Therefore, the existing technology lacks an automated solution that can intelligently perceive the environment and dynamically calculate the optimal throwing strategy to ensure that the lifesaving equipment can be accurately and reliably delivered to the fallen person. SUMMARY

[0005] Therefore, the present application proposes an environment perception-based precise throwing method and system for water rescue unmanned aerial vehicles to solve the problems of inaccurate throwing, neglect of environmental influence, and dependence on manual operation in the prior art, and to realize unmanned and intelligent precise water rescue.

[0006] The technical solution adopted by the present application is as follows:

[0007] An environment perception-based precise throwing method for water rescue unmanned aerial vehicles, the method comprising the following steps:

[0008] S1: Environment data acquisition and target identification: after the water rescue unmanned aerial vehicle flies to the target area, the on-board visible light camera, thermal imaging camera, and millimeter wave radar are used to cooperatively collect environmental data, and multi-sensor fusion is performed to form a continuous video stream. Through a trained multi-task AI visual recognition model, the video stream is analyzed in real time to detect and lock the fallen person target, calculate the real-time position of the fallen person, and identify the posture of the fallen person;

[0009] S2: Dynamic environment parameter estimation; based on the continuous video stream obtained in step S1, the moving vector of the water surface texture is analyzed by the optical flow method and the feature point tracking technology, and the real-time flow velocity and direction of the water surface are estimated;

[0010] S3: Optimal throwing point calculation; the real-time position Ptarget of the falling person, the attitude, the flow velocity V_water and direction obtained in step S2, the current height H of the unmanned aerial vehicle, the aerodynamic parameters of the lifesaving device, the local gravity acceleration g and the wind speed W are comprehensively considered, and a throwing dynamics model is established. A plurality of parabolic trajectories are simulated and calculated through the model, and the optimal air release point of the lifesaving device, i.e. the optimal throwing point Pdrop, is inversely solved, so that the lifesaving device falls into the falling person's reach after drifting to the falling person's reach;

[0011] S4: Precise maneuvering and executing throwing; the unmanned aerial vehicle autonomously flies to the optimal throwing point Pdrop calculated, adjusts its attitude during the flight to compensate for wind deflection. At the moment of reaching the optimal throwing point, the throwing mechanism is controlled to launch the life buoy at a predetermined initial speed V0 to ensure that it moves according to the pre-calculated parabolic trajectory;

[0012] S5: Throwing effect verification and closed-loop feedback; after the lifesaving ring falls into the water, the unmanned aerial vehicle continues to track the lifesaving ring and the falling person using the on-board visible light camera, thermal imaging camera and millimeter wave radar. If it is detected that the lifesaving ring and the falling person successfully meet (such as through the automatic locking of the towing joint on the lifesaving ring), the task is completed; if it is found to deviate from the expected, the re-throw logic is triggered immediately, and the step S1 is returned to perform a new round of calculation and throwing.

[0013] The application provides a water rescue unmanned aerial vehicle precise throwing system based on environmental perception, which comprises:

[0014] The environment data acquisition and target recognition module cooperates with the on-board visible light camera, thermal imaging camera and millimeter wave radar to acquire environmental data, and performs multi-sensor fusion. Through the trained multi-task AI visual recognition model, the falling person target is detected and locked in real time, and the attitude is recognized;

[0015] The dynamic environment parameter estimation module estimates the real-time flow velocity and direction of the water surface based on the continuous video stream obtained by the environment data acquisition and target recognition module, analyzes the moving vector of the water surface texture by the optical flow method and the feature point tracking technology, and estimates the real-time flow velocity and direction of the water surface;

[0016] An optimal throwing point calculation module: the optimal throwing point calculation module comprehensively obtains the real-time position Ptarget of the fallen person, the attitude, the water flow speed Vwater and the direction obtained in step S2, the current height H of the unmanned aerial vehicle, the aerodynamic parameters of the lifesaving device, the local gravity acceleration g and the wind speed W, and establishes a throwing dynamics model. Through the model simulation calculation of multiple parabolic trajectories, the optimal air release point of the lifesaving device, i.e. the optimal throwing point Pdrop, is inversely solved, so that the lifesaving device falls into the reachable position of the fallen person after drifting.

[0017] Precise maneuvering and executing throwing module: the precise maneuvering and executing throwing module controls the unmanned aerial vehicle to autonomously fly to the calculated optimal throwing point Pdrop, adjusts the attitude during the flight to compensate for the wind deflection. At the moment when the unmanned aerial vehicle reaches the predetermined point of the optimal throwing point, the throwing mechanism is controlled to launch the life buoy at a predetermined initial speed V0, so as to ensure that the life buoy moves according to the pre-calculated parabolic trajectory.

[0018] Throwing effect verification and closed-loop feedback: the throwing effect verification and closed-loop feedback controls the unmanned aerial vehicle to continue tracking the life buoy and the fallen person by using the on-board visible light camera, thermal imaging camera and millimeter wave radar. If it is detected that the life buoy and the fallen person successfully meet (such as through the automatic locking of the pulling joint on the life buoy), the task is completed; if it is found that the life buoy deviates from the expectation, the re-throwing logic is triggered immediately, and step S1 is returned to perform a new round of calculation and throwing.

[0019] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the above-mentioned precise throwing method of an unmanned aerial vehicle for water rescue based on environment perception when executing the computer program.

[0020] A computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the above-mentioned precise throwing method of an unmanned aerial vehicle for water rescue based on environment perception.

[0021] Compared with the prior art, the beneficial effects of the present application are:

[0022] 1. Precise: through the combination of AI vision and dynamics model, the throwing decision is improved from flying over the head to calculating the future position, which significantly improves the hit rate of the first throwing of the lifesaving device.

[0023] 2. Intelligent: full-process automation reduces the dependence on professional pilot operation, responds faster, and is especially suitable for large-scale disaster rescue scenarios.

[0024] 3. Environmental adaptability: for the first time, the water flow speed is taken as a core decision variable, so that the throwing strategy is more in line with the dynamic characteristics of the real water area, and the pain point of accurate throwing but far drifting is solved.

[0025] 4. High reliability: The design of the verification after throwing and the re-throw mechanism forms a decision closed loop, ensuring the final success of the rescue mission. BRIEF DESCRIPTION OF DRAWINGS

[0026] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of these drawings.

[0027] Figure 1 is the flowchart in the present application. DETAILED DESCRIPTION

[0028] The embodiments of the present application will be described in detail below with reference to the drawings.

[0029] The embodiments of the present application will be described in detail below with reference to the drawings.

[0030] It should be noted that the various aspects of the embodiments described below are within the scope of the appended claims. It should be apparent to those of ordinary skill in the art that the aspects described herein can be embodied in a wide variety of forms, and that any specific structure and / or function described herein is merely illustrative. Based on the teachings herein one skilled in the art should appreciate that an aspect described herein can be implemented independently of any other aspects and that an aspect described herein can be implemented both as any number of software and / or hardware configurations and that the examples described herein are not limited to any particular software and / or hardware configuration. It should also be understood that any logical and / or textual combinations of two or more aspects described herein can fall within the scope of the present application. For example, any number and combination of aspects set forth herein can be used to implement a device and / or practice a method. Additionally, the device and / or method can be implemented using other structures and / or functionality in addition to or other than those set forth herein.

[0031] In addition, in the following description, specific details are provided to facilitate thorough understanding of examples. However, one skilled in the art will understand that the examples can be practiced without these specific details.

[0032] The embodiment of the specification proposes an environment perception-based precise throwing method of a water rescue unmanned aerial vehicle: the method comprises the following steps:

[0033] S1: environment data acquisition and target identification: after the water rescue unmanned aerial vehicle flies to the target area, the on-board visible light camera, thermal imaging camera and millimeter wave radar are used to cooperatively collect environment data, and multi-sensor fusion is performed to form a continuous video stream, and through a trained multi-task AI visual recognition model, the video stream is analyzed in real time to detect and lock the target of a fallen person, and the posture of the fallen person is identified;

[0034] S1-1: multi-sensor data synchronous acquisition:

[0035] The visible light camera, thermal imaging camera and millimeter wave radar on board the unmanned aerial vehicle in the application need to be hardware time-synchronized. Through a unified trigger signal, each frame of data collected by the three is strictly aligned in the time stamp, laying a foundation for subsequent data fusion.

[0036] S1-2: training and deployment of multi-task AI visual recognition model based on deep learning:

[0037] This step is the key to accurately identifying the fallen person and the posture of the fallen person, and specifically as follows:

[0038] Data set construction: a large-scale, multi-scene water rescue image data set is collected and constructed. The data set should include: public data sets (such as COCO, VisDrone), fallen rescue video screenshots crawled from the network, images and videos simulated by shooting in different water environments such as swimming pools, lakes and sea areas.

[0039] For each training image in the water rescue image data set, not only the "person" bounding box is labeled, but also pixel-level segmentation (Semantic Segmentation) is performed to distinguish the human body, water surface, ripples, etc., and key point labeling (such as joint positions of head, hand, foot, etc.) of the posture of the person is performed, and posture attribute labels (such as "waving hands", "static floating", "struggling") are labeled.

[0040] Network model structure: an encoder-decoder structure convolutional neural network is preferably selected. The encoder (such as ResNet, EfficientNet) is used to extract multi-scale features. The decoder part includes three branches:

[0041] Target detection branch: based on the head of YOLO or Faster R-CNN, the bounding box and confidence of the fallen person are output.

[0042] Instance segmentation branch: based on the head of Mask R-CNN, the accurate pixel-level mask of the outline of the fallen person is output.

[0043] Pose estimation branch: head based on OpenPose, output human key point heat map.

[0044] Introduce attention mechanism after the encoder, such as SE module or CBAM, so that the network pays more attention to the significant targets on the water surface and suppresses the interference of wave reflection.

[0045] Model training: use the constructed dataset to train the multi-task model in an end-to-end manner. The loss function is the weighted sum:

[0046] Loss = α * Ldetection + β * Lsegmentation + γ * Lpose;

[0047] Where L_detection is the loss of target detection, L_segmentation is the segmentation loss, Lpose is the pose estimation loss, and α, β, γ are balance weight hyperparameters. The training uses Adam optimizer and adopts learning rate decay strategy.

[0048] S1-3: Real-time recognition: deploy the trained model on the UAV on-board computing unit after lightening. After the real-time video stream is input into the model, the real-time position, contour and pose classification result of the fallen person can be obtained.

[0049] S2: Dynamic environment parameter estimation; based on the continuous video stream obtained in step S1, the moving vector of water surface texture is analyzed through optical flow method and feature point tracking technology, and the real-time water flow velocity and direction of the water surface are estimated; the real-time water flow velocity estimation includes:

[0050] S2-1: In the video stream, apply feature point detection algorithm to consecutive frames, and extract feature points in significant areas such as wave and floating object. Then, use optical flow method (such as Lucas-Kanade algorithm) to calculate the moving vector (u, v) of these feature points between two frames; where the consecutive frames are represented as t frame and t+1 frame;

[0051] S2-2: Convert the image pixel displacement of feature points to horizontal displacement in world coordinate system. Considering the motion of UAV itself, the body displacement calculated by IMU and GPS data of UAV is used to compensate the horizontal displacement of feature points. After compensation, the horizontal displacement vectors of all feature points are clustered and averaged, and the average water flow velocity Vwater = (Vx, Vy) of the current area of water surface can be estimated.

[0052] S3: Calculation of the optimal drop point; A drop dynamics model is established by combining the real-time position Ptarget of the person in the water, their attitude, the water flow velocity Vwater and direction obtained in step S2, the current altitude H of the UAV, and the aerodynamic parameters of the rescue equipment such as drag coefficient, local gravitational acceleration g, and wind speed W. Multiple parabolic trajectories are simulated using this model, with the optimization objective being the rescue equipment drifting to a position within reach of the person in the water after falling in, thus solving for the optimal aerial release point of the rescue equipment, i.e., the optimal drop point Pdrop.

[0053] S3-1: Throwing dynamics model establishment: The motion of the rescue equipment after leaving the drone is divided into two stages: the air flight stage and the water drifting stage.

[0054] The process of establishing the throwing dynamics model includes:

[0055] a) Dynamics model of the air flight phase

[0056] The trajectory of the rescue equipment is affected by gravity, air resistance, and wind. Its dynamic model during the flight phase can be described by the following set of differential equations:

[0057] ;

[0058] in:

[0059] m represents the mass of the lifesaving equipment;

[0060] (x, y, z) represents the position of the lifesaving equipment in the world coordinate system;

[0061] ρ is the air density, Cd is the drag coefficient, A is the frontal area, Vrel is the relative velocity, and |Vrel| is the magnitude of the relative velocity.

[0062] Wx represents the wind speed component in the x-direction, and Wy represents the wind speed component in the y-direction.

[0063] g is the acceleration due to gravity.

[0064] Initial conditions: At t=0, i.e. at the moment of throwing: the position of the rescue equipment is (X(0), Y(0), Z(0)), and the velocity of the rescue equipment is V(0) = Vdrone + V0, where Vdrone is the current velocity vector of the drone.

[0065] b) The model for the water drifting phase includes: After the life-saving equipment falls into the water, its motion is mainly governed by the water flow. Its motion in the water is simplified to the velocity of the equipment being equal to the velocity of the water flow at its location, described by the following set of differential equations:

[0066] ;

[0067] where (Xwater, Ywater) is the coordinate of the life-saving equipment after falling into the water, (Vwx, Vwy) is the velocity vector of the life-saving equipment after falling into the water, the falling point is denoted as Psplash, and the falling time is denoted as Tsplash.

[0068] S3-3: Numerical calculation process of the optimal throwing point Pdrop

[0069] The solution of the optimal throwing point Pdrop is a trajectory optimization problem, and the goal is to make the life-saving equipment meet the fallen person at a future time Tmeet after falling into the water. The calculation process is as follows:

[0070] S3-3-1: Motion prediction of the fallen person

[0071] According to the real-time position Ptarget of the fallen person, the water flow velocity Vwater, and the attitude of the fallen person, the position of the fallen person at any future time t is predicted as follows:

[0072] Ptargetpredicted(t) = Ptarget + (α*Vwater+β*R) * t;

[0073] where α and β are control parameters, and R is a normalized parameter of the attitude of the fallen person;

[0074] S3-3-2: Definition of objective function

[0075] The goal is to minimize the final distance between the life-saving equipment and the fallen person. The objective function J is defined as the distance between the life-saving equipment after falling into the water and drifting for a short time Δt and the predicted position of the fallen person:

[0076] J(Pdrop, V0) = | Psplash+ (1-α)(Vwater * Δt)- (Ptarget +Vwater *Tsplash+(β*R)(Tsplash + Δt))|;

[0077] The essence of optimization is to make the falling point Psplash of the life-saving equipment accurately fall on the predicted position of the fallen person at the falling time Tsplash after considering the water flow drift.

[0078] S3-3-3: Numerical solution and optimization algorithm

[0079] The specific steps are as follows:

[0080] Initialization: Set the size and direction of the initial throwing velocity V0. Set the initial guess value of the decision variable Pdrop as the current position of the drone Pdrone.

[0081] Numerical integration: For a given set of (Pdrop, V0), solve the differential equations of the air phase using numerical integration. The integration starts from t=0 until Z(t) = 0, obtaining the splash time Tsplash and the splash point Psplash.

[0082] Compute objective function: Substitute Psplash and Tsplash into the objective function J, and compute the error of the current release strategy.

[0083] Iterative optimization: Adjust the decision variable Pdrop using optimization algorithms to minimize the objective function J.

[0084] Gradient computation: To use gradient descent, we need to compute the gradient of the objective function J with respect to the decision variable Pdrop, ∇J. This can be done using the perturbation method: slightly change each component of Pdrop, re-do the numerical integration to get a new J, and thus approximate the partial derivative.

[0085] The gradient descent update formula is:

[0086] ;

[0087] where η is the learning rate and k is the iteration number.

[0088] When the objective function J is smaller than a pre-set threshold, or the iteration number reaches an upper limit, stop the optimization. The Pdrop obtained at this time is the optimal release point.

[0089] S4: Precise maneuvering and executing release: The UAV autonomously flies to the calculated optimal release point P_drop, adjusts its attitude to compensate for wind drift. At the moment of reaching the predetermined point, the release mechanism is controlled to launch the life buoy with a specific initial velocity V0, ensuring that it follows the pre-calculated parabolic trajectory;

[0090] S5: Release effect verification and closed-loop feedback: After the life buoy splashes, the UAV continues to track the life buoy and the swimmer using the vision system. If it detects that the life buoy and the swimmer have successfully met (such as through the automatic locking of the towing joint on the life buoy), the task is completed; if it finds that it deviates from the expectation, it immediately triggers the re-release logic and returns to step S1 for a new round of calculation and release.

[0091] The side of the life buoy is integrated with a permanent magnet ring and a mechanical self-locking hook. The end of the towing rope released by the UAV is a magnetic metal head and a buckle that matches the self-locking hook. When the life buoy is released near the swimmer, the UAV can slowly descend in height to approach the towing rope head to the life buoy. Through magnetic attraction, preliminary docking is achieved, and then under the pulling force or slight mechanical collision, the buckle and the self-locking hook complete the final locking, forming a reliable connection.

[0092] The UAV continuously tracks the life buoy and the person falling into the water after the throwing. The distance between the two is determined by a visual algorithm. If the distance is less than a threshold value within a set time and the tow rope is detected to be in a taut state, it is considered that the rescue is successful. Otherwise, the re-throwing program is immediately started, the steps S1-S4 are re-executed, and better parameters can be selected in the second throwing.

[0093] The present application converts the solution of the optimal throwing point into a numerical optimization problem by establishing an accurate throwing dynamics model considering air resistance and wind force. Through efficient iterative calculation by algorithms such as gradient descent, a release point is finally determined, which can make the lifesaving device accurately fall in front of the person falling into the water in a dynamic water flow environment, greatly improving the success rate and efficiency of rescue.

[0094] The present application provides a water rescue UAV accurate throwing system based on environmental perception, which comprises:

[0095] The environment data acquisition and target recognition module: the environment data acquisition and target recognition module uses the fusion of airborne visible light camera, thermal imaging camera and millimeter wave radar to collect environment data, detects and locks the target of the person falling into the water in real time through the trained multi-task AI visual recognition model, and identifies the attitude of the target;

[0096] The dynamic environment parameter estimation module: the dynamic environment parameter estimation module estimates the real-time water flow speed and direction of the water surface based on the continuous video stream obtained by the environment data acquisition and target recognition module through the optical flow method and feature point tracking technology;

[0097] The optimal throwing point calculation module: the optimal throwing point calculation module establishes a throwing dynamics model by comprehensively considering the real-time position Ptarget of the person falling into the water, the attitude, the water flow speed Vwater and direction obtained by the dynamic environment parameter estimation module, the current height H of the UAV, the aerodynamic parameters of the lifesaving device, the local gravitational acceleration g and the wind speed W. Through the model simulation calculation of multiple parabolic trajectories, the best air release point of the lifesaving device is inversely solved, that is, the optimal throwing point Pdrop, so that the lifesaving device falls into the water and drifts to the reachable position of the person falling into the water.

[0098] The precise maneuvering and executing throwing module: the precise maneuvering and executing throwing module controls the UAV to fly to the optimal throwing point Pdrop calculated, adjusts its attitude to compensate for the wind deflection. At the moment of reaching the predetermined point, the throwing mechanism is controlled to launch the life buoy at a certain initial speed V0, ensuring that it moves according to the pre-calculated parabolic trajectory.

[0099] The throwing effect verification and closed-loop feedback: the throwing effect verification and closed-loop feedback control unmanned aerial vehicle continues to track the life buoy and the swimmer by using the vision system. If it is detected that the life buoy and the swimmer successfully meet (such as through the automatic locking of the towing joint on the life buoy), the task is completed; if it is found that it deviates from the expectation, the re-throw logic is triggered immediately, and the environment data acquisition and target identification module is returned to a new round of calculation and throwing.

[0100] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, and the processor implements the above-mentioned precise throwing method of an environment-aware water rescue unmanned aerial vehicle.

[0101] A computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement the above-mentioned precise throwing method of an environment-aware water rescue unmanned aerial vehicle.

[0102] Those of ordinary skill in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the above-mentioned embodiments of each method. Any reference to memory, storage, databases, or other media used in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0103] In this specification, the same or similar parts between various embodiments are referred to each other, and each embodiment focuses on the difference from other embodiments. Especially, for the embodiments described later, the description is relatively simple, and the relevant part can refer to the part of the description of the foregoing embodiments.

[0104] The above merely provides the specific implementation of the present application, but the protection scope of the present application is not limited to this. Any changes or replacements within the technical scope disclosed by the present application can be easily conceived by the person skilled in the art, and should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. An environment-sensing-based precise throwing method for water rescue drones, characterized in that, The method comprises the following steps: S1: environmental data acquisition and target recognition: after the water rescue unmanned aerial vehicle flies to the target area, the environmental data is collected by using the airborne visible light camera, thermal imaging camera and millimeter wave radar, and multi-sensor fusion is performed to form a continuous video stream, a trained multi-task AI visual recognition model is used to analyze the video stream in real time to detect and lock the target of the fallen person, calculate the real-time position of the fallen person, and identify the posture of the fallen person; S2: dynamic environment parameter estimation; based on the continuous video stream obtained in step S1, the water surface texture movement vector is analyzed by using the optical flow method and feature point tracking technology to estimate the real-time water flow velocity and direction of the water surface; S3: optimal throwing point calculation; the real-time position Ptarget of the fallen person, the posture R, the water flow velocity and direction obtained in step S2, the current height H of the unmanned aerial vehicle, the aerodynamic parameters of the lifesaving device, the local gravity acceleration g and the wind speed W are comprehensively considered to establish a throwing dynamics model; through the model simulation calculation of multiple parabolic trajectories, the optimal air release point of the lifesaving device, that is, the optimal throwing point Pdrop, is inversely solved, so that the lifesaving device falls into the reachable position of the fallen person after drifting in the water; S4: precise maneuvering and executing throwing: the unmanned aerial vehicle autonomously flies to the optimal throwing point Pdrop calculated, adjusts its own posture to compensate for the wind deflection during the flight; at the moment of reaching the optimal throwing point, the throwing mechanism is controlled to launch the life buoy at a predetermined initial speed V0 to ensure that it moves along the pre-calculated parabolic trajectory; S5: throwing effect verification and closed-loop feedback: after the life buoy falls into the water, the unmanned aerial vehicle continues to track the life buoy and the fallen person by using the airborne visible light camera, thermal imaging camera and millimeter wave radar; if it is detected that the life buoy and the fallen person successfully meet, the task is completed; If it is found that it deviates from the expectation, the re-throwing logic is triggered immediately, and the step S1 is returned to perform a new round of calculation and throwing.

2. The method of claim 1, wherein the method further comprises: The step S1 specifically comprises: S1-1: multi-sensor data synchronous acquisition; S1-2: training and deployment of multi-task AI visual recognition model based on deep learning: S1-3: real-time recognition: the trained model is lightened and deployed on the airborne computing unit of the unmanned aerial vehicle; after the real-time video stream is input into the model, the accurate pixel position, contour and posture classification result of the fallen person can be obtained.

3. The method of claim 2, wherein the method further comprises: The step S1-2 of training and deploying the multi-task AI visual recognition model based on deep learning specifically comprises: Collect and construct a water rescue image dataset; For each training image in the water rescue image dataset, not only the boundary box of the person is labeled, but also pixel-level segmentation is performed to distinguish the human body, water surface, spray and floating object, and the key points of the posture of the person are labeled, and the posture attribute label is labeled; Network model structure: a convolutional neural network with an encoder-decoder structure is selected; the encoder is used to extract multi-scale features; the decoder part includes three branches: Target detection branch: YOLO or Faster R-CNN head based, outputting the boundary box and confidence of the fallen person; Instance segmentation branch: Mask R-CNN head based, outputting the accurate pixel-level mask of the contour of the fallen person; Pose estimation branch: head based on OpenPose, output human key point heat map; Introducing attention mechanism after the encoder, making the network pay more attention to the human, water surface, splash, and floating object on the water surface; Using the constructed water rescue image dataset to train the multi-task model in an end-to-end manner; the loss function is the weighted sum: Loss = α * Ldetection + β * Lsegmentation + γ * Lpose; Where L_detection is the loss of target detection, L_segmentation is the segmentation loss, Lpose is the pose estimation loss, and α, β, γ are balance weight hyperparameters; the Adam optimizer is used for training, and the learning rate decay strategy is adopted.

4. The method of claim 3, wherein the method further comprises: The step S2 specifically comprises: S2-1: In the video stream, a feature point detection algorithm is applied to consecutive frames to extract feature points in the splash and floating object area; then, the optical flow method is used to calculate the movement vector (u, v) of these feature points between two frames; where the consecutive frames are represented as t frame and t+1 frame; S2-2: Convert the image pixel displacement of the feature points to the horizontal displacement in the world coordinate system; calculate the displacement of the unmanned aerial vehicle body itself using the IMU and GPS data of the unmanned aerial vehicle, and then compensate the horizontal displacement of the feature points; after compensation, cluster and average the horizontal displacement vectors of all feature points, that is, the average water flow velocity Vwater = (Vx, Vy) of the current area water surface can be estimated.

5. The method of claim 1, wherein the method further comprises: The throwing dynamics model specifically comprises: an air flight stage dynamics model and a water drifting stage model, wherein the air flight stage dynamics model specifically comprises: The motion trajectory of the lifesaving device is affected by gravity, air resistance and wind force; the motion equation of the air flight stage dynamics model can be described by the following differential equation set: ； Where: m is the mass of the lifesaving device; (x, y, z) is the position of the lifesaving device in the world coordinate system; ρ is the air density, Cd is the resistance coefficient, A is the windward area, Vrel is the relative velocity, and |Vrel| is the modulus of the relative velocity; Wx is the component of wind speed in the x direction, and Wy is the component of wind speed in the y direction; g is the acceleration of gravity; The water drifting stage model specifically comprises: after the lifesaving device falls into the water, its motion is mainly dominated by the water flow, and its water motion is simplified as the velocity of the device being equal to the water flow velocity at its position, which is described by the following differential equation set: ； Where (Xwater, Ywater) is the coordinate of the lifesaving device after falling into the water, (Vwx, Vwy) is the velocity vector of the lifesaving device after falling into the water, the splash point is denoted as Psplash, and the splash time is denoted as Tsplash.

6. The method of claim 1, wherein the method further comprises: The numerical calculation process of the optimal throwing point Pdrop comprises: S3-3-1: Motion prediction of the person falling into the water According to the real-time position Ptarget of the person falling into the water, the water flow velocity Vwater and the attitude of the person falling into the water, the position of the person falling into the water at any time t in the future is predicted: Ptargetpredicted(t) = Ptarget + (α*Vwater+β*R) * t; where α, β are control parameters, and R is a normalized parameter of the fallen person's posture; S3-3-2: Define the objective function The goal is to minimize the final distance between the lifesaving device and the fallen person; define the objective function J as the distance between the lifesaving device after falling and drifting for a short time Δt and the predicted position of the fallen person: J(Pdrop, V0) = | Psplash+ (1-α)(Vwater * Δt)- (Ptarget +Vwater * Tsplash+(β*R)(Tsplash + Δt))|; where Psplash is the splash point, and Tsplash is the splash time; S3-3-3: Numerical solution and optimization algorithm The specific steps are as follows: Initialization: Set the size and direction of the initial throw speed V0; set the initial guess value of the decision variable Pdrop as the current position of the drone Pdrone; Numerical integration: for a given (Pdrop, V0), use numerical integration method to solve the differential equations of the air phase; the integration starts from t=0 until Z(t) = 0, and the splash time Tsplash and the splash point Psplash are obtained; Calculate the objective function: substitute Psplash and Tsplash into the objective function J to calculate the error of the current throw strategy; Iterative optimization: adjust the decision variable Pdrop using the optimization algorithm to minimize the objective function J; Gradient calculation: use the perturbation method to calculate the gradient ∇J of the objective function J with respect to the decision variable Pdrop, including: change each component of the decision variable Pdrop; based on the changed decision variable Pdrop, recalculate the new value of the new objective function J through numerical integration, and estimate the partial derivative of each component of Pdrop according to the change of the objective function J; The gradient descent update formula is: ； where η is the learning rate, and k is the iteration number; When the objective function J is less than a preset threshold, or the iteration number reaches the upper limit, stop optimization; at this time, the obtained Pdrop is the optimal throw point.

7. An environment perception-based precise throwing system for water rescue drones, comprising: An environment data acquisition and target recognition module: the environment data acquisition and target recognition module cooperates with the airborne visible light camera, thermal imaging camera and millimeter wave radar to acquire environment data, and performs multi-sensor fusion; through the trained multi-task AI visual recognition model, the fallen person target is detected and locked in real time, and the posture is recognized; A dynamic environment parameter estimation module: based on the continuous video stream obtained by the environment data acquisition and target recognition module, the real-time water flow velocity and direction of the water surface are estimated by analyzing the water surface texture movement vector through the optical flow method and feature point tracking technology; The optimal throwing point calculation module: the optimal throwing point calculation module comprehensively considers the real-time position Ptarget of the falling person, the water flow velocity Vwater and direction obtained by the dynamic environment parameter estimation module, the current height H of the unmanned aerial vehicle, the aerodynamic parameters of the lifesaving device, the local gravitational acceleration g, and the wind speed W, establishes a throwing dynamics model; through the model simulation calculation of multiple parabolic trajectories, the lifesaving device is taken as the optimization target after falling into the falling person's reach, and the best air release point of the lifesaving device, that is, the optimal throwing point Pdrop, is inversely solved; The precise maneuvering and executing throwing module: the precise maneuvering and executing throwing module controls the unmanned aerial vehicle to fly to the optimal throwing point Pdrop calculated, adjusts the attitude to compensate for the wind deflection; at the moment of reaching the predetermined point, the throwing mechanism is controlled to launch the life buoy at a specific initial speed V0, and the life buoy is ensured to move according to the pre-calculated parabolic trajectory; The throwing effect verification and closed-loop feedback: the throwing effect verification and closed-loop feedback control the unmanned aerial vehicle to continue to track the lifesaving ring and the falling person by using the visual system; if the lifesaving ring and the falling person are successfully detected to meet, the task is completed; If it is found that the deviation from the expectation, the re-throw logic is triggered immediately, and the environment data acquisition and target recognition module is returned to perform a new round of calculation and throwing.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements a precise throwing method of an environment-aware water rescue unmanned aerial vehicle according to any one of claims 1 to 6 when executing the computer program.

9. A computer readable storage medium, the computer readable storage medium stores a computer program, wherein the computer program is executed by a processor to implement a precise throwing method of an environment-aware water rescue unmanned aerial vehicle according to any one of claims 1 to 6.

Images