Unmanned aerial vehicle intelligent obstacle avoidance method based on forest scene

Abstract

The present application relates to a kind of forest scene-based unmanned aerial vehicle intelligent obstacle avoidance method, belong to unmanned aerial vehicle technical field, including the dynamic obstacle in forest is respectively established different motion model, to structured disturbed obstacle and carries out state prediction, for structured disturbed obstacle construction and high maneuvering unstructured obstacle construction risk field, and the above-mentioned risk field and static obstacle risk field are fused into global composite risk field, construct the unconstrained optimization objective function including smoothness cost, dynamics feasibility cost and continuous space-time risk cost, the optimal trajectory is generated by calculating the gradient of objective function to B spline control point.This method realizes the obstacle avoidance of dynamic obstacle by modeling and predicting dynamic obstacle in forest;Risk weight is introduced to build environment field, the collision risk of obstacle is estimated, and finally the optimal trajectory balanced energy consumption and risk is generated, to ensure that unmanned aerial vehicle energy consumption is least under the premise of safety.

Description

Technical Field

[0001] This invention belongs to the field of unmanned aerial vehicle (UAV) technology, specifically relating to an intelligent obstacle avoidance method for UAVs based on forest scenarios. Background Technology

[0002] In recent years, drones have been increasingly used in complex scenarios such as forest patrols and ecological monitoring. In these scenarios, drones not only need to avoid dense static obstacles, but also need to deal with complex and ever-changing dynamic obstacles.

[0003] In the field of dynamic obstacle avoidance for UAVs, prediction based on a single kinematic model or trajectory optimization using an isotropic risk field is commonly used. This method first extracts moving point clouds using vision or LiDAR, then uses standard Kalman filtering or extended Kalman filtering to predict the future center position of the dynamic obstacle based on a single model, and expands the predicted obstacle center outward into a physical protection sphere or constructs an isotropic Gaussian potential field. Finally, the expanded sphere is used as a hard constraint or the Gaussian potential field is used as a penalty term and added to the cost function, and the gradient descent method is used to solve it.

[0004] This method cannot handle highly maneuverable dynamic obstacles with multimodal motion characteristics, such as birds (e.g., constant speed cruise and high-maneuver startle), and it cannot describe anchoring constraints, which can easily lead the planner to make completely wrong obstacle avoidance decisions. During high-speed flight, existing technologies are prone to overprotection or ineffective obstacle avoidance when facing dynamic obstacles. In addition, under high-speed flight conditions, the risk field lacks analytical gradients, resulting in low efficiency in real-time optimization solutions. Summary of the Invention

[0005] The purpose of this invention is to overcome the problems of modeling distortion, over-conservatism, and low solution efficiency of existing dynamic obstacle avoidance planning algorithms for UAVs flying in complex forest scenarios, and to provide an intelligent obstacle avoidance method for UAVs based on forest scenarios.

[0006] The technical solution of the present invention is as follows:

[0007] A method for intelligent obstacle avoidance by drones in a forest setting includes the following steps:

[0008] S1: Classify dynamic obstacles in the forest into structured disturbed obstacles with anchor points and highly maneuverable unstructured obstacles with active avoidance behavior, and establish different motion models for each.

[0009] S2: Based on the kinematic model, the structured disturbed obstacle is predicted using a Gaussian distribution, and the highly maneuverable unstructured obstacle is predicted using a Gaussian mixture model for multimodal state prediction.

[0010] S3: Construct a risk field based on Mahalanobis distance for the structured disturbed obstacle, construct a risk field containing a relative velocity alignment factor for the highly maneuverable unstructured obstacle, and merge the above risk fields with the static obstacle risk field into a global composite risk field;

[0011] S4: Based on the B-spline parameterized UAV trajectory, construct an unconstrained optimization objective function that includes smoothness cost, dynamic feasibility cost, and continuous spatiotemporal risk cost of the global composite risk field. Generate the optimal trajectory by calculating the gradient of the objective function with respect to the B-spline control points.

[0012] Furthermore, in step S1:

[0013] For structured disturbance obstacles, the position is modeled using a second-order damped harmonic oscillator stochastic differential equation driven by random forces. The modeling equation includes mass terms, damping terms, elastic recovery terms, and stochastic external disturbance force terms. State variables including position and velocity are defined to establish a state-space equation, in which a standard Wiener process is introduced to simulate gust disturbances.

[0014] For highly maneuverable unstructured obstacles, an interactive multi-model framework is adopted, which combines a three-dimensional uniform velocity model and a Singer maneuvering model for modeling. A three-dimensional state vector containing position, velocity, and acceleration is defined, and two state transition matrices are constructed for the three-dimensional uniform velocity model and the Singer maneuvering model.

[0015] Furthermore, in step S2: for structured disturbed obstacles, since their motion model is a linear Gaussian system, given an initial state Under the premise that the mean and covariance of the Gaussian distribution of the future state can be obtained by solving the problem.

[0016] Furthermore, in step S2, for highly mobile unstructured obstacles, the following sub-steps are performed sequentially:

[0017] The mixed probability is calculated based on the probability of the three-dimensional uniform velocity model and the Singh maneuver model from the previous moment;

[0018] Initialize the mixed input of state and covariance based on mixed probabilities;

[0019] Based on the initialized state, forward prediction is performed under both the three-dimensional uniform velocity model and the Singh maneuver model.

[0020] Calculate the measurement residuals and residual covariance, and update the state, covariance, and model probability of the three-dimensional uniform velocity model and the Singh maneuver model;

[0021] By fusing the outputs of the three-dimensional uniform velocity model and the Singh maneuver model, the best estimated state and covariance at the current moment are obtained.

[0022] By predicting the distribution at future moments, we obtain the spatial location distribution in the form of a Gaussian mixture model.

[0023] Furthermore, in step 3, the risk field based on Mahalanobis distance constructed for the structured disturbed obstacles has a risk value that is negatively exponentially related to the square of the Mahalanobis distance between the UAV's planned position and the expected position of the obstacle, and is multiplied by the entity confidence weight; wherein the Mahalanobis distance is calculated by the inverse matrix of the obstacle position covariance matrix.

[0024] Furthermore, in step 3, the formula for calculating the relative velocity alignment factor in the risk field constructed for highly maneuverable unstructured obstacles is as follows:

[0025]

[0026] in, This is the penalty gain coefficient; The vector transpose of relative velocities. Location of the drone; Let be the spatial position vector of the highly maneuverable unstructured obstacle at a future moment.

[0027] Combining the relative velocity alignment factor, the spatiotemporal risk field of highly maneuverable unstructured obstacles is defined as follows:

[0028] ;

[0029] For highly maneuverable unstructured obstacles in the future Spatial location The scalar risk value generated at that location; The predicted probability weights for highly maneuverable unstructured obstacles to maintain the first flight mode at future moments; This is the inverse of the future position covariance matrix; For highly maneuverable unstructured obstacles in maintaining the first In various flight modes, future moments The desired spatial location.

[0030] Furthermore, the global composite risk field in step 3 is represented as follows:

[0031] ;

[0032] in, Indicating drones in the future Planning location The global composite risk value at the location; This is a static obstacle risk field. and These represent the total number of detected structured dynamic obstacles and highly maneuverable unstructured obstacles, respectively. and They represent the first The structured dynamic obstacle and the first A highly mobile unstructured obstacle in the future Planning location The independent risk value generated at that location.

[0033] Furthermore, in step 4, the UAV trajectory is parameterized using uniform B-spline; the unconstrained optimization objective function is a weighted sum of smoothness cost, dynamic feasibility cost, and continuous spatiotemporal risk cost multiplied by their respective penalty weight coefficients.

[0034] Among them, the smoothness cost is based on the square integral of the third derivative of the B-spline trajectory; the dynamic feasibility cost is achieved by penalizing control points that exceed the preset maximum speed and maximum acceleration; and the continuous spatiotemporal risk cost is the integral of the global composite risk field along the trajectory in the prediction time domain.

[0035] Furthermore, in step 4, when calculating the gradient of the objective function with respect to the B-spline control points, the gradient of the continuous spatiotemporal risk cost is calculated using the chain rule; wherein, for the risk term caused by highly maneuverable unstructured obstacles, its analytical gradient includes: an exponential term, a relative spatial distance vector, and an orthogonal projection term of the relative velocity in a direction perpendicular to the relative distance vector, the orthogonal projection term being used to generate the lateral repulsion gradient.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1. This invention achieves obstacle avoidance by modeling and predicting dynamic obstacles in forests; it introduces risk weights to construct an environmental field, estimates the collision risk of obstacles, and finally generates an optimal trajectory that balances energy consumption and risk, ensuring that the UAV consumes the least amount of energy while maintaining safety.

[0038] 2. This invention proposes an intelligent obstacle avoidance method based on different types of obstacles commonly found in forests. Different models are established for different obstacles, solving the prediction delay problem caused by traditional constant-velocity models during target maneuvers. Simultaneously, this invention introduces a relative velocity alignment factor, calculating lateral repulsion gradients only for targets with a collision threat, effectively avoiding overly conservative behavior during obstacle avoidance and improving trajectory smoothness. Furthermore, this invention compresses the trajectory solution time, reducing the computational and hardware requirements of the UAV and effectively saving costs. Attached Figure Description

[0039] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0040] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0041] like Figure 1 As shown, a method for intelligent obstacle avoidance of drones based on forest scenarios is described, and the specific implementation steps are as follows:

[0042] The first step is dynamic obstacle modeling, which is divided into structured disturbances with anchor points and objects with subjective intent.

[0043] For structured disturbances with anchor points, such as tree branches or power lines affected by wind, the position of the second-order damped harmonic oscillator driven by random forces is determined using stochastic differential equations. Modeling is performed.

[0044] The equation of motion is defined as:

[0045] ;

[0046] in, For random external disturbances, The equivalent mass of the obstacle. Let the obstacle be the three-dimensional spatial position vector in the global coordinate system. The three-dimensional velocity of the obstacle. The three-dimensional motion acceleration of the obstacle. The damping coefficient is... The elastic recovery coefficient, This is the three-dimensional static equilibrium point of the obstacle when it is not subject to any wind force.

[0047] Let the state variable be The state-space equation is then:

[0048] ;

[0049] in, Represents the differential. The derivative representing time, The derivative of the state variable. The differential of the standard Wiener process; This is a standard Wiener process used to simulate gust disturbances; This is the system state transition matrix, used to describe the coupling relationship between position and velocity within the system; The input control matrix is ​​used to describe the constant stretching effect of the fixed anchor point position on the system state. This is the noise-driven matrix, used to map one-dimensional or low-dimensional random wind forces to the specific state dimension of the system. Where:

[0050] , ;

[0051] in, It is the identity matrix, which is a 3×3 diagonal matrix.

[0052] For dynamic obstacles with high maneuverability and active avoidance capabilities, such as birds, an interactive multi-model framework is adopted, combining a three-dimensional uniform velocity model and a Singh maneuvering model for modeling.

[0053] Define the three-dimensional state vector of a highly maneuverable unstructured obstacle as follows: ;

[0054] Represents the three-dimensional position components of a dynamic obstacle in the global coordinate system; Represents the three-dimensional velocity components of a dynamic obstacle; This represents the three-dimensional acceleration components of a dynamic obstacle.

[0055] Constructing state equations ,in This indicates two different flight modes; The three-dimensional state vector of a highly maneuverable unstructured obstacle at any given time; This represents the discrete time step number. Represents the next discrete moment; Indicates the dynamic obstacle in the first... Discrete state transition matrix under each mode; Indicates the dynamic obstacle in the first... Process noise gain matrix under each mode; Indicates the first The system process noise matrix under this mode is usually assumed to be Gaussian white noise with zero mean.

[0056] Assuming a highly maneuverable unstructured obstacle maintains a constant velocity when unflapped or under patrol, with zero acceleration or only minimal noise, this cruise mode can be modeled as follows:

[0057] ;

[0058] Here, T is the system's discretization sampling time. The basic sampling time is set to T=0.1, and it is not used as a matrix transpose when it is used as a superscript.

[0059] The acceleration of highly maneuverable unstructured obstacles can be calculated using a first-order Markov process. Modeling is performed, among which The maneuver frequency is used to describe the intensity of maneuvering of highly maneuverable unstructured obstacles. The target's current acceleration; The degree of urgency; Zero-mean Gaussian white noise; It is the ratio of the sensor sampling period to the time constant of the obstacle's change in acceleration; The time autocorrelation coefficient is calculated based on a first-order Markov process.

[0060] In the specific implementation of this plan, when the unstructured obstacle is a common bird in the forest (such as sparrows or flocks of pigeons), its actual maximum maneuvering speed is set. 4m / s 2 ~6m / s 2 maneuver frequency Set to 0.1s -1 In practice, measurement data is obtained through sensors (such as lidar), and the actual physical parameters of birds in flight are used as input to fit the real-world movements of birds suddenly startled, swooping down, or changing direction. The discretized state transition matrix is ​​as follows:

[0061] ;

[0062] The switching between two modes for highly maneuverable unstructured obstacles follows a Markov process, denoted by the state transition probability matrix. ,in Indicates from pattern Switch to mode The probability of.

[0063] The second step is to predict dynamic obstacles.

[0064] For structured disturbances with anchor points, such as tree branches or power lines disturbed by wind, since it is a linear Gaussian system, given an initial state... Under the premise of future time The state distribution within the range follows a Gaussian distribution. The mean and covariance can be expressed as:

[0065] ;

[0066] ;

[0067] in, This is the mean vector of the obstacle states; The rate of change of the mean vector over time; Let be the covariance matrix of the obstacle state; Let A be the expansion rate of the covariance matrix over time; let A be the system state transition matrix. A represents the matrix transpose, used to describe the coupling relationship between position and velocity within the system; B is the input control matrix, used to describe the constant stretching effect of the fixed anchor point position on the system state. The three-dimensional static equilibrium point of the obstacle when it is not subject to any wind force; Let be the noise covariance matrix for a continuous-time process.

[0068] For highly maneuverable dynamic obstacles such as birds that can actively avoid obstacles, at time k, the posterior probabilities of the two models based on the previous time k-1 are... Calculate the mixture probability of model i at the previous time step, given that the motion model is j at the current time step. :

[0069] ;

[0070] ;

[0071] in, For discrete time steps; for At any given moment, a highly maneuverable unstructured obstacle serves as the model. The posterior probability; For the Markov transition probability of mode switching; It is the prior predicted probability of the j-th motion mode at the current time, and plays a normalization role when calculating the mixed probability; For mixed probabilities.

[0072] Perform mixed input initialization for state and covariance:

[0073] ;

[0074] ;

[0075] in, For the first The model utilizes the cutoff point All data at any given time, for Optimal estimation of the state at time step; and Each is a single model The filtered state and covariance output at the previous time step; and The models at the current time are respectively The initial mixing state and mixing covariance used.

[0076] Based on the mixed state, forward predictions are performed under both the uniform velocity state and the maneuvering model:

[0077] ;

[0078] ;

[0079] in, For the first The state transition matrix of each model; The predicted position and velocity of the highly maneuverable unstructured obstacle at the current moment; The noise-driven equation; This is the process noise covariance matrix of the corresponding model; For posterior estimation, it means using the cutoff to Observational data at time Estimate the state at any given time; For prior prediction, it means using the cutoff date... Observational data at time The state at any given moment can be further predicted.

[0080] The high-frequency cluster centroid provided by the LiDAR and vision fusion algorithm is used as the measurement at the current moment. Calculate the residual and residual covariance :

[0081] ;

[0082] ;

[0083] in, For residuals; This is a measurement vector, representing the actual three-dimensional position of the dynamic obstacle in the global coordinate system; For the first step calculated Predicting the prior state of each model; This is the observation matrix. The residual covariance matrix; For the first The prior prediction covariance matrix of each model; This is the measurement noise covariance matrix.

[0084] Using Kalman gain Update status: ;

[0085] Update covariance: ;

[0086] in, Here is the Kalman gain matrix. It is the inverse of the residual covariance matrix. To obtain the first [data] by combining the latest radar data. Optimal posterior state estimation for a given motion pattern Here, the posterior covariance matrix is ​​formed by incorporating the measurement data. It is a 6×6 identity matrix.

[0087] Calculate the likelihood function Update the probability model:

[0088] ;

[0089] Among them, No. The likelihood of each model; Let be a multivariate Gaussian probability density function, representing the variable With a mean of 0 and a covariance of The probability density of values ​​taken under a normal distribution; To make use of the current moment After observing the data, the dynamic obstacle was indeed in the first position. The posterior true probability of a pattern; It represents the prior prediction probability of the j-th motion mode at the current time and serves as a normalization factor when calculating the mixed probability.

[0090] At this moment, the dynamic obstacle is constantly The best estimated state and covariance are obtained through multimodal fusion:

[0091] ;

[0092] ;

[0093] in, This represents the final optimal state for the fusion of dynamic obstacles; This represents the final total covariance of the dynamic obstacles.

[0094] Predicting the distribution of future moments ,in, This is the current physical time; For prediction in the time domain, the projection matrix of the location dimension is extracted as follows: ,in It is a 3-order identity matrix. It is a 3rd order zero matrix.

[0095] For structured obstacles such as tree branches and power lines, the future positions follow a Gaussian distribution. ,in, and The mean vector and position covariance matrix of the future positions of the structured dynamic obstacles are respectively derived from the differential equation in the first step. Integral result:

[0096] ;

[0097] in, This represents the full state vector of the structured obstacle. Here is the system's state transition matrix; Input matrix for static equilibrium position; Let V be the static equilibrium position vector of the obstacle. For constant inputs caused by the static equilibrium position, This is the noise driving matrix; For the system's future prediction time, This is the initial moment of integration; This represents the full state vector of the structured obstacle at the initial moment; and This represents the state transition matrix of the linear continuous system; This represents the stochastic differential increment of the standard Wiener process over a time period.

[0098] Initial state mean , Let be the initial full-state vector of the structured obstacle. This is the mathematical expectation operator, used to calculate the statistical mean of a random variable or stochastic process. Since the incremental expectation of a standard Wiener process is zero, the expectation of the random integral term is also zero. Therefore:

[0099] ;

[0100] in, Integral infinitesimal element representing time. Let be the integral variable within the integration interval; is the base of the natural logarithm.

[0101] because It is a constant vector that does not change with time, and it is invertible with respect to the system's state transition matrix A, so it can be further written as:

[0102] ;

[0103] in, For state At any moment The statistical mean. State at the initial moment The statistical mean.

[0104] The covariance matrix is ​​defined as ;

[0105] Among them, deviation ;

[0106] Since the uncertainty of the initial state is unrelated to wind noise in the future period, the expectation of the cross term is zero, therefore:

[0107] ;

[0108] in, Let be the covariance matrix of the state variables; The covariance is known at the current initial time. Represents the standard Wiener process. This represents the random increment of the Wiener process.

[0109] According to Itō's isomorphism theorem, the variance of the random integral is transformed into a deterministic integral. Let the covariance matrix of the continuous-time white noise be... ,have:

[0110] ;

[0111] For highly maneuverable dynamic obstacles such as birds that can actively avoid obstacles, their future positions follow a Gaussian mixture model, with spatial expectation and covariance as follows:

[0112] ;

[0113] ;

[0114] in, This represents the step extension term of the state transition matrix over time. The exponent is the current discrete time. To the future prediction time The equivalent prediction steps experienced are used to achieve continuous state deduction across time steps. To maintain the first for highly maneuverable unstructured obstacles In various flight modes, future moments Spatial location; The first result obtained after updating using the aforementioned Kalman filter The optimal state estimate of the model is used here as a forward-looking time. The initial state reference for spatial location simulation; This is the basic step size for the discrete system; To maintain the first for highly maneuverable unstructured obstacles In this mode, the future moment The spatial location covariance matrix; Let be the covariance matrix of a highly maneuverable unstructured obstacle in its full state at future time steps. The linear projection of the covariance matrix. for The transpose of a matrix.

[0115] The third step is to construct a spatiotemporal risk field. This involves assuming the drone is in a future time... The planned location within is .

[0116] Since its wind disturbance mainly manifests as random oscillations near the anchor point, a risk function based on Mahalanobis distance is adopted:

[0117] ;

[0118] in As the confidence weight of the entity; For drones in the future The three-dimensional planning location; To plan the three-dimensional location of structured obstacles for the UAV and time The scalar risk assessment value generated at the location; The desired location of the structured obstacle; It is the inverse of the future position covariance matrix.

[0119] For highly maneuverable unstructured obstacles, due to their initiative and high mobility, the collision risk is strongly correlated with the direction of relative velocity. The relative velocity is defined as... .in, It is a relative velocity vector; Let be the expected velocity vector of the highly maneuverable obstacle at a future moment; This is the velocity vector of the drone itself.

[0120] Introducing a relative velocity alignment factor :

[0121] ;

[0122] in, This is the penalty gain coefficient; Let be the spatial position vector of the highly maneuverable unstructured obstacle at a future moment.

[0123] Based on the predictions made, the spatiotemporal risk field of highly maneuverable unstructured obstacles is defined as follows:

[0124] ;

[0125] in, For highly maneuverable unstructured obstacles in the future 3D planning location The scalar risk value generated at that location; The predicted probability weights for highly maneuverable unstructured obstacles to maintain the first flight mode at future moments; This is the inverse of the future position covariance matrix; For highly maneuverable unstructured obstacles in maintaining the first In various flight modes, future moments The desired spatial location.

[0126] Static obstacle risk field The fusion of structured dynamic obstacles and highly maneuverable obstacles represents the global composite risk field as follows:

[0127] ;

[0128] in, Indicating drones in the future Planning location The global composite risk value at the location; This is a static obstacle risk field. and These represent the total number of detected structured dynamic obstacles and highly maneuverable unstructured obstacles, respectively. and They represent the first The structured dynamic obstacle and the first A highly mobile unstructured obstacle in the future Planning location The independent risk value generated at that location.

[0129] The fourth step is to generate a trajectory based on the risk gradient. To meet the flight speed requirements of the drone in the forest scenario and ensure real-time performance, B-splines are used to generate the drone trajectory. Perform parameterization. Let the control points of the B-spline be... .

[0130] Construct an unconstrained optimization objective function :

[0131] ;

[0132] in, For the sake of smoothness, ; Let be the third derivative of the drone trajectory with respect to time.

[0133] As a cost to dynamic feasibility, penalties are imposed for exceeding the maximum speed. and maximum acceleration Control points.

[0134] To mitigate continuous spatiotemporal risk, the integral of the trajectory over the entire prediction time domain must be minimized within the spatial risk field. . The global composite risk field constructed for the third step.

[0135] , , This is the penalty weighting coefficient.

[0136] During implementation, the maximum planning speed of the drone is set. Maximum dynamic acceleration Set the prediction time domain. To ensure obstacle avoidance safety while maintaining the smoothest possible trajectory, the objective function weight coefficients are set as smoothness penalty weights. Dynamic penalty weight Continuous spatiotemporal risk penalty weight .

[0137] Calculate the gradient with respect to the B-spline control points using the chain rule:

[0138] ;

[0139] symbol It represents the partial derivative in calculus.

[0140] For dynamic risks caused by highly maneuverable unstructured obstacles, the analytical gradient is: ;in, , indicating the first The exponential term of the risk function under various flight modes; This represents the relative spatial distance vector between the UAV and the desired location of the highly maneuverable obstacle; relative velocity Perpendicular to the aforementioned relative distance vector orthogonal projection vectors on.

[0141] Based on the above implementation, this invention introduces the real acceleration of a flying bird in real time. A parameter-corrected state transition model is constructed, and a velocity alignment factor is included. The complex spatiotemporal risk field allows drones to initiate avoidance responses in advance, improving the success rate of dynamic obstacle avoidance. It also maintains a sufficient safe distance from birds throughout the process, solving the problem that existing traditional trajectory planning technologies lack accurate prediction of highly maneuverable obstacles such as birds, are prone to lag when dealing with highly maneuverable obstacles, have a low obstacle avoidance success rate, and the extreme safe distance is usually less than the safety threshold.

[0142] When faced with sudden, highly maneuverable obstacles, the avoidance paths generated by existing methods typically exhibit drastic curvature changes, easily exceeding the dynamic feasible boundary of the UAV. This invention addresses this by parameterizing with B-splines and utilizing the analytical gradient obtained through differentiation to mitigate the smoothness cost. With risk and cost By comprehensively searching for the optimal trajectory, the obtained optimal path outperforms traditional methods in terms of maximum curvature and speed jump, effectively avoiding high-frequency oscillations in the underlying flight control system.

[0143] In summary, compared with existing technologies, this invention significantly optimizes the smoothness and dynamic feasibility of flight trajectories while ensuring obstacle avoidance safety in complex dynamic environments.

[0144] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent obstacle avoidance by unmanned aerial vehicles (UAVs) in a forest setting, characterized in that: Includes the following steps: S1: Classify dynamic obstacles in the forest into structured disturbed obstacles with anchor points and highly maneuverable unstructured obstacles with active avoidance behavior, and establish different motion models for each. S2: Based on the kinematic model, the structured disturbed obstacle is predicted using a Gaussian distribution, and the highly maneuverable unstructured obstacle is predicted using a Gaussian mixture model for multimodal state prediction. S3: Construct a risk field based on Mahalanobis distance for the structured disturbed obstacle, construct a risk field containing a relative velocity alignment factor for the highly maneuverable unstructured obstacle, and merge the above risk fields with the static obstacle risk field into a global composite risk field; S4: Based on the B-spline parameterized UAV trajectory, construct an unconstrained optimization objective function that includes smoothness cost, dynamic feasibility cost, and continuous spatiotemporal risk cost of the global composite risk field. Generate the optimal trajectory by calculating the gradient of the objective function with respect to the B-spline control points.

2. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that: In step S1: For structured disturbance obstacles, the position is modeled using a second-order damped harmonic oscillator stochastic differential equation driven by random forces. The modeling equation includes mass terms, damping terms, elastic recovery terms, and stochastic external disturbance force terms. State variables including position and velocity are defined to establish a state-space equation, in which a standard Wiener process is introduced to simulate gust disturbances. For highly maneuverable unstructured obstacles, an interactive multi-model framework is adopted, which combines a three-dimensional uniform velocity model and a Singer maneuvering model for modeling. A three-dimensional state vector containing position, velocity, and acceleration is defined, and two state transition matrices are constructed for the three-dimensional uniform velocity model and the Singer maneuvering model.

3. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 2, characterized in that, In step S2: For structured disturbed obstacles, since their motion model is a linear Gaussian system, given an initial state... Under the premise that the mean and covariance of the Gaussian distribution of the future state can be obtained by solving the problem.

4. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 2, characterized in that, In step S2, for highly mobile unstructured obstacles, the following sub-steps are performed sequentially: The mixed probability is calculated based on the probability of the three-dimensional uniform velocity model and the Singh maneuver model from the previous moment; Initialize the mixed input of state and covariance based on mixed probabilities; Based on the initialized state, forward prediction is performed under both the three-dimensional uniform velocity model and the Singh maneuver model. Calculate the measurement residuals and residual covariance, and update the state, covariance, and model probability of the three-dimensional uniform velocity model and the Singh maneuver model; By fusing the outputs of the three-dimensional uniform velocity model and the Singh maneuver model, the best estimated state and covariance at the current moment are obtained. By predicting the distribution at future moments, we obtain the spatial location distribution in the form of a Gaussian mixture model.

5. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that, In step 3, the risk field based on Mahalanobis distance is constructed for the structured disturbed obstacle. Its risk value has a negative exponential relationship with the square of the Mahalanobis distance between the UAV's planned position and the expected position of the obstacle, and is multiplied by the entity confidence weight. Mahalanobis distance is calculated using the inverse of the obstacle position covariance matrix.

6. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that, In step 3, the formula for calculating the relative velocity alignment factor in the risk field constructed for highly maneuverable unstructured obstacles is as follows: in, This is the penalty gain coefficient; The vector transpose of relative velocity. Location of the drone; Let be the spatial position vector of the highly maneuverable unstructured obstacle at a future moment. Combining the relative velocity alignment factor, the spatiotemporal risk field of highly maneuverable unstructured obstacles is defined as follows: ; For highly maneuverable unstructured obstacles in the future Spatial location The scalar risk value generated at that location; The predicted probability weights for highly maneuverable unstructured obstacles to maintain the first flight mode at future moments; This is the inverse of the future position covariance matrix; For highly maneuverable unstructured obstacles in maintaining the first In various flight modes, future moments The desired spatial location.

7. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that, The global composite risk field in step 3 is represented as follows: ; in, Indicating drones in the future Planning location The global composite risk value at the location; This is a static obstacle risk field. and These represent the total number of detected structured dynamic obstacles and highly maneuverable unstructured obstacles, respectively. and They represent the first The structured dynamic obstacle and the first A highly mobile unstructured obstacle in the future Planning location The independent risk value generated at that location.

8. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that, In step 4, the UAV trajectory is parameterized using uniform B-spline; the unconstrained optimization objective function is a weighted sum of smoothness cost, dynamic feasibility cost, and continuous spatiotemporal risk cost multiplied by their respective penalty weight coefficients. Among them, the smoothness cost is based on the square integral of the third derivative of the B-spline trajectory; the dynamic feasibility cost is achieved by penalizing control points that exceed the preset maximum speed and maximum acceleration; and the continuous spatiotemporal risk cost is the integral of the global composite risk field along the trajectory in the prediction time domain.

9. The intelligent obstacle avoidance method for unmanned aerial vehicles (UAVs) based on a forest scene according to claim 1, characterized in that, In step 4, when calculating the gradient of the objective function with respect to the B-spline control points, the gradient of the continuous spatiotemporal risk cost is calculated using the chain rule; wherein, for the risk term caused by highly maneuverable unstructured obstacles, its analytical gradient includes: an exponential term, a relative spatial distance vector, and an orthogonal projection term of the relative velocity in a direction perpendicular to the relative distance vector, the orthogonal projection term being used to generate the lateral repulsion gradient.

Images