Multi-rotor unmanned aerial vehicle adaptive obstacle avoidance strategy generation system and method

Abstract

The invention relates to the technical field of artificial intelligence and unmanned aerial vehicle path planning, in particular to a multi-rotor unmanned aerial vehicle adaptive obstacle avoidance strategy generation system and method, and the system comprises a multi-source environment sensing module which is used for collecting environment data; the dynamic obstacle modeling module is used for constructing a dynamic obstacle motion state model; the threat degree evaluation and analysis module is used for constructing a threat degree evaluation model and then outputting a threat evaluation result; the path generation optimization module is used for inputting the threat evaluation result into a path generation network model and generating N groups of candidate paths meeting constraints; performing simulation deduction on the candidate paths, and calculating a comprehensive score of expected energy consumption, a risk value and a task completion degree of each path based on a threat assessment result; and selecting an optimal path according to the comprehensive score, and generating a final obstacle avoidance strategy. According to the method, the defects of response lag and scene adaptability caused by dynamic obstacle prediction misalignment, layered decision delay and fixed parameter stiffness of a traditional obstacle avoidance system are overcome.

Description

Technical Field

[0001] The present invention relates to the technical field of artificial intelligence and unmanned aerial vehicle (UAV) path planning, and in particular to a system and method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV. Background Art

[0002] In the field of drone design and manufacturing, improving the autonomy and safety of aircraft in complex environments is one of the core challenges. Traditional obstacle avoidance system architectures have significant limitations, which restrict the improvement of their environmental adaptability. These limitations are mainly reflected in three aspects:

[0003] First, obstacle avoidance perception relies on single-sensor tracking methods that generally lack an effective mechanism to quantify the reliability of perception data. In common scenarios such as strong light interference or multipath effects, trajectory prediction errors can significantly increase, causing drones to frequently trigger aggressive emergency maneuvers (such as losing control during sharp turns) or overly conservative stalling behaviors (such as hovering and delaying missions), compromising flight safety and efficiency. Second, regarding the coordination between obstacle avoidance and path planning, existing solutions often employ a serial processing architecture of "perception → planning → control," separating environmental perception, path generation, and obstacle avoidance decision-making. When the number of dynamic obstacles increases (e.g., more than five), system response latency often exceeds 200ms, making it difficult to meet the real-time requirements of sudden or highly dynamic obstacle avoidance scenarios. Finally, safety decision-making models typically employ fixed threshold parameters, which are unable to flexibly adapt to the diverse obstacle avoidance strategy requirements of different flight scenarios (e.g., dense urban environments versus open terrain). This results in significant variations in obstacle avoidance success rates (up to 40%) across different scenarios.

[0004] These systematic deficiencies collectively limit the potential for widespread application of UAVs in complex and dynamic scenarios requiring highly reliable autonomous flight. Summary of the Invention

[0005] One of the objectives of the present invention is to provide an adaptive obstacle avoidance strategy generation system for multi-rotor UAVs, which overcomes the response lag and scene adaptability defects of traditional obstacle avoidance systems caused by inaccurate dynamic obstacle prediction, hierarchical decision delay and rigid fixed parameters, and realizes safe and efficient autonomous obstacle avoidance of UAVs in complex dynamic environments.

[0006] In order to achieve the above objectives, a multi-rotor UAV adaptive obstacle avoidance strategy generation system is provided, which includes the following modules:

[0007] Multi-source environmental perception module: used to collect environmental data through the millimeter-wave radar and binocular vision sensors on the multi-rotor drone, and output a three-dimensional grid map and confidence matrix with a time stamp after spatiotemporal alignment;

[0008] Dynamic obstacle modeling module: used to obtain the current position data of the UAV; set a threshold based on the confidence matrix, filter the grid cell data with confidence below the threshold in the 3D grid map, and obtain valid environmental data; extract independent obstacle clusters from the valid environmental data based on cluster analysis, and after removing the ground point cloud, output the spatial distribution matrix of the independent obstacle clusters; fit the motion trajectory equation of each obstacle through the spatial distribution matrix, and use the confidence matrix as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model;

[0009] Threat assessment and analysis module: used to output a motion state parameter set through a dynamic obstacle motion state model, the motion state parameter set including a velocity vector and an acceleration covariance matrix; after constructing a threat assessment model based on the current posture data of the UAV and the motion state parameter set, it outputs a threat assessment result, which includes the dynamic collision probability and collision time (TTC) value of each obstacle;

[0010] Path generation optimization module: used to input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the UAV; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path according to the comprehensive score and generate the final obstacle avoidance strategy, which includes a control instruction set for the heading angle correction, speed adjustment coefficient and emergency braking threshold.

[0011] Furthermore, the multi-source environment perception module specifically includes:

[0012] Data acquisition submodule: used to obtain obstacle distance information through the millimeter-wave radar on the multi-rotor drone;

[0013] Texture features are collected through the binocular vision sensor on the multi-rotor drone;

[0014] Data processing submodule: used to align the two types of data in time and space, and output a three-dimensional grid map with a timestamp and a confidence matrix. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

[0015] Furthermore, in the threat assessment analysis module, the threat assessment model adopts a temporal neural network architecture, including:

[0016] The first sub-network processes the temporal characteristics of the obstacle motion state parameter set;

[0017] The second sub-network integrates the UAV pose data;

[0018] The output layer generates dynamic collision probability and TTC value, where the dynamic collision probability is normalized to the [0,1] range through the Sigmoid function.

[0019] Furthermore, the path generation optimization module also includes the following submodules:

[0020] Path update submodule: used to initiate local path replanning when a low-confidence grid cell is suddenly added to the confidence matrix: based on the current optimal path, a local optimization area is delineated in the three-dimensional grid map; the path topology continuity is maintained within the optimization area, and waypoints are updated based on the gradient descent method; the path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the UAV.

[0021] A second object of the present invention is to provide a method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV, comprising the following steps:

[0022] Multi-source environmental perception steps: Collect environmental data through millimeter-wave radar and binocular vision sensors, and output a three-dimensional grid map with a time stamp and a confidence matrix after spatiotemporal alignment;

[0023] Dynamic obstacle modeling steps: Obtain the current position data of the UAV; Set a threshold based on the confidence matrix, filter the grid cell data with confidence below the threshold in the 3D grid map to obtain valid environmental data; Extract independent obstacle clusters from the valid environmental data based on cluster analysis, and after removing the ground point cloud, output the spatial distribution matrix of the independent obstacle clusters; Fit the motion trajectory equation of each obstacle using the spatial distribution matrix, and use the confidence matrix as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model;

[0024] Threat assessment and analysis steps: outputting a motion state parameter set through a dynamic obstacle motion state model, wherein the motion state parameter set includes a velocity vector and an acceleration covariance matrix; constructing a threat assessment model based on the current posture data of the UAV and the motion state parameter set, and then outputting a threat assessment result, wherein the threat assessment result includes a dynamic collision probability and a collision time (TTC) value of each obstacle;

[0025] Path generation optimization step: input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the UAV; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path according to the comprehensive score, and generate a final obstacle avoidance strategy, which includes a control instruction set for a heading angle correction, a speed adjustment coefficient and an emergency braking threshold.

[0026] Furthermore, the multi-source environment perception step includes:

[0027] Obtain obstacle distance information through millimeter-wave radar;

[0028] Collect texture features through binocular vision sensors;

[0029] The two types of data are aligned in time and space, and a three-dimensional grid map with a timestamp and a confidence matrix are output. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

[0030] Furthermore, in the threat level assessment and analysis step, the threat level assessment model adopts a temporal neural network architecture, including:

[0031] The first sub-network processes the temporal characteristics of the obstacle motion state parameter set;

[0032] The second sub-network integrates the UAV pose data;

[0033] The output layer generates dynamic collision probability and TTC value, where the dynamic collision probability is normalized to the [0,1] range through the Sigmoid function.

[0034] Furthermore, the path generation optimization step also includes:

[0035] When a low-confidence grid cell is suddenly added to the confidence matrix, local path replanning is initiated:

[0036] Based on the current optimal path, a local optimization area is defined in the three-dimensional grid map;

[0037] Maintain path topology continuity within the optimization area and update waypoints based on gradient descent method;

[0038] The path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the drone.

[0039] Principles and advantages:

[0040] 1. Enhanced reliability of dynamic obstacle perception and modeling: A confidence matrix is ​​used to quantitatively evaluate multi-source sensor data, effectively filtering out low-reliability environmental information. A spatiotemporal alignment mechanism is combined to improve the accuracy of three-dimensional environmental representation. A confidence weighting coefficient is introduced into obstacle motion modeling to significantly improve the accuracy of dynamic obstacle trajectory prediction, avoiding the risk of misjudgment caused by sensor noise in traditional methods and ensuring drone flight stability.

[0041] 2. Breakthrough in real-time obstacle avoidance decision-making: This system utilizes an end-to-end decision-making architecture to deeply integrate threat assessment with path generation. This system reduces repeated computations through a shared feature extraction layer and directly outputs the optimal obstacle avoidance path and control instructions. This eliminates the time-consuming simulation bottleneck of traditional hierarchical decision chains and enables instantaneous response in complex dynamic scenarios.

[0042] 3. AI-based scenario-adaptive optimization: Embedded AI models perceive and understand flight scenario characteristics in real time, intelligently matching optimal parameter configurations to the threat assessment and path generation networks. This prioritizes obstacle avoidance agility in dense environments, while optimizing energy efficiency to extend flight range in open terrain, achieving intelligent dynamic optimization of flight performance in all scenarios.

[0043] 4. Comprehensive improvement in system robustness: From data reliability management at the perception layer, parallel processing at the decision layer, to the scenario-adaptive mechanism at the execution layer, a closed-loop obstacle avoidance logic is formed, significantly enhancing the drone's anti-interference ability under extreme conditions such as sudden changes in lighting, sensor interference, and sudden changes in obstacle motion patterns. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a logic block diagram of a multi-rotor drone adaptive obstacle avoidance strategy generation system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0045] The following is further described in detail through specific implementation methods:

[0046] Example

[0047] A multi-rotor UAV adaptive obstacle avoidance strategy generation system, basically as Figure 1 As shown, it includes a drone body and a controller, and the controller includes the following modules:

[0048] Multi-source environmental perception module: used to collect environmental data through the millimeter-wave radar and binocular vision sensor on the multi-rotor UAV, and output a three-dimensional grid map and confidence matrix with a time stamp after spatiotemporal alignment; the multi-source environmental perception module specifically includes:

[0049] Data acquisition submodule: used to obtain obstacle distance information through the millimeter-wave radar on the multi-rotor drone;

[0050] Texture features are collected through the binocular vision sensor on the multi-rotor drone;

[0051] Data processing submodule: used to align the two types of data in time and space, and output a three-dimensional grid map with a timestamp and a confidence matrix. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

[0052] Spatiotemporal alignment technology: After millimeter-wave radar data is converted from spherical coordinates to Cartesian coordinates, it is unified to the drone's center of mass coordinate system through the hand-eye calibration matrix; binocular vision uses SIFT feature matching to calculate disparity depth, and time synchronization is achieved through the PTP protocol to achieve microsecond-level alignment. Confidence matrix generation method: Using the Bayesian probability fusion model:

[0053] Conf ij =ω r P(radar accuracy) + ω v P (visual matching)

[0054] Among them, ω r =0.6,ω v =0.4 is the sensor weight coefficient, the radar accuracy is calculated using the multipath effect attenuation model, and the visual matching degree is evaluated based on the feature point overlap rate.

[0055] Dynamic obstacle modeling module: used to obtain the current posture data of the UAV; set a threshold based on the confidence matrix, filter the grid unit data with confidence below the threshold in the three-dimensional grid map, and obtain valid environmental data; extract independent obstacle clusters from the valid environmental data based on cluster analysis, and output the spatial distribution matrix of the independent obstacle clusters after removing the ground point cloud; fit the motion trajectory equation of each obstacle through the spatial distribution matrix, and use the confidence matrix as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model; remove the ground point cloud: use the random sampling consensus (RANSAC) algorithm to fit the ground plane. The weighted method calculation formula of the motion trajectory equation is as follows:

[0056]

[0057] The coefficient a m 、b m 、c m Fitted by RANSAC algorithm, is the corresponding grid confidence weighted value (high confidence grid dominates the fitting: The weight of data points is increased by 30%; low confidence grids suppress noise: The data points do not participate in the iteration), Represents the spatial coordinates of the obstacle at time t. a m represents the acceleration-related coefficient, which characterizes the acceleration change rate of the obstacle motion; b m represents the speed-related coefficient, which characterizes the instantaneous speed of the obstacle; c m Represents the position offset coefficient, which characterizes the position offset of the obstacle at the initial moment.

[0058] Threat Assessment and Analysis Module: This module is used to output a set of motion state parameters using a dynamic obstacle motion state model. The motion state parameter set includes a velocity vector and an acceleration covariance matrix. Based on the current UAV posture data and the motion state parameter set, a threat assessment model is constructed, and then a threat assessment result is output. The threat assessment result includes the dynamic collision probability and collision time (TTC) value of each obstacle. The threat assessment model adopts a temporal neural network architecture and includes:

[0059] The first sub-network (bidirectional LSTM layer (hidden unit 32)) processes the temporal features of the obstacle motion state parameter set; the input is the temporal data of the obstacle velocity vector Vec{v} and the acceleration covariance matrix Sigma_a;

[0060] The second sub-network (fully connected layer) integrates the drone pose data;

[0061] The output layer generates dynamic collision probability and collision time TTC value, where the dynamic collision probability is normalized to the range [0,1] using the Sigmoid function. The calculation formula of collision probability is as follows:

[0062]

[0063] in, is the relative velocity vector (the velocity vector of the obstacle relative to the UAV), and k is the risk sensitivity coefficient.

[0064] Path generation optimization module: used to input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the drone; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path according to the comprehensive score, and generate the final obstacle avoidance strategy, which includes the control instruction set of heading angle correction, speed adjustment coefficient and emergency braking threshold. Path generation network model: uses a 3D convolutional neural network, with an input of 64×64×5 tensors (channels 1-3: threat assessment results; channel 4: drone posture; channel 5: grid map slice), and the output layer generates a waypoint sequence through full connection. Comprehensive scoring function for comprehensive scoring:

[0065] S=ω e ·E norm +ω r ·R norm +ω t T norm

[0066] Among them, S is the comprehensive score of the path, and the larger the value, the better the path.

[0067] Enorm is the expected energy consumption, which indicates the total energy consumed by the drone along the path. The calculation formula is as follows:

[0068] a is acceleration, V is velocity. k1·||a(t)|| 2 Characterizes the energy consumption of the motor to overcome inertia (the proportional coefficient k1 is determined by the mass of the drone), k2·V(t) 3 Characterizes the aerodynamic drag energy consumption (the proportional coefficient k2 is calibrated by wind tunnel test).

[0069]

[0070] R is the risk value, which indicates the comprehensive risk level of obstacles encountered on the path. The calculation formula is as follows:

[0071] R=max(P C )+λ·exp(-TTC),λ: risk sensitivity coefficient

[0072]

[0073] T is the mission completion degree, which indicates the efficiency of the path in achieving the scheduled flight mission. The calculation formula is as follows:

[0074] Among them, A is the covered path point, B is the total path point, is the time delay, and β is the time delay penalty term.

[0075] The path generation optimization module also includes the following submodules:

[0076] The path update submodule is used to initiate local path replanning when a low-confidence grid cell is suddenly added to the confidence matrix. Based on the current optimal path, a local optimization region is defined in the three-dimensional grid map. The path topology continuity is maintained within the optimization region, and waypoints are updated using a gradient descent method. The path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the UAV. The beneficial effects of local path replanning are as follows:

[0077] 1. Instantaneous response to sudden threats: Real-time detection of unknown obstacles (such as targets in visual blind spots) based on confidence matrix mutations, triggering millisecond-level emergency obstacle avoidance, and eliminating the risk of collision caused by response delays. 2. Efficient and lightweight path optimization: Delineate local optimization areas in the three-dimensional grid map, and only update the affected waypoints, significantly reducing computing overhead and ensuring real-time re-planning. 3. Enhanced flight stability: Maintain smooth heading transition through path topology continuity constraints, and synchronously embed roll angle and centripetal acceleration limits to prevent sharp turns without control or structural overload. 4. Resource-optimal obstacle avoidance path: Use the gradient descent method to search for the local optimal solution under physical constraints, achieve safe avoidance with minimal path deviation, and maximize the efficiency of the original mission.

[0078] A method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV comprises the following steps:

[0079] Multi-source environmental perception step: collecting environmental data through millimeter-wave radar and binocular vision sensors, and outputting a three-dimensional grid map and confidence matrix with a time stamp after spatiotemporal alignment; the multi-source environmental perception step includes:

[0080] Obtain obstacle distance information through millimeter-wave radar;

[0081] Collect texture features through binocular vision sensors;

[0082] The two types of data are aligned in time and space, and a three-dimensional grid map with a timestamp and a confidence matrix are output. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

[0083] Dynamic obstacle modeling steps: Obtain the current position data of the UAV; Set a threshold based on the confidence matrix, filter the grid cell data with confidence below the threshold in the 3D grid map to obtain valid environmental data; Extract independent obstacle clusters from the valid environmental data based on cluster analysis, and after removing the ground point cloud, output the spatial distribution matrix of the independent obstacle clusters; Fit the motion trajectory equation of each obstacle using the spatial distribution matrix, and use the confidence matrix as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model;

[0084] Threat assessment and analysis steps: Output a motion state parameter set through a dynamic obstacle motion state model, wherein the motion state parameter set includes a velocity vector and an acceleration covariance matrix; construct a threat assessment model based on the current posture data of the UAV and the motion state parameter set, and then output a threat assessment result, wherein the threat assessment result includes the dynamic collision probability and collision time (TTC) value of each obstacle; the threat assessment model adopts a temporal neural network architecture, including:

[0085] The first sub-network processes the temporal characteristics of the obstacle motion state parameter set;

[0086] The second sub-network integrates the UAV pose data;

[0087] The output layer generates dynamic collision probability and TTC value, where the dynamic collision probability is normalized to the [0,1] range through the Sigmoid function.

[0088] Path generation optimization step: Input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the UAV; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path based on the comprehensive score and generate the final obstacle avoidance strategy, which includes the control instruction set of the heading angle correction, speed adjustment coefficient and emergency braking threshold. The path generation optimization step also includes:

[0089] When a low-confidence grid cell is suddenly added to the confidence matrix, local path replanning is initiated:

[0090] Based on the current optimal path, a local optimization area is defined in the three-dimensional grid map;

[0091] Maintain path topology continuity within the optimization area and update waypoints based on gradient descent method;

[0092] The path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the drone.

[0093] The above is only an embodiment of the present invention. Common knowledge such as the known specific structures and characteristics in the scheme will not be described in detail here. Those of ordinary skill in the art are aware of all common technical knowledge in the technical field to which the invention belongs before the application date or priority date, can obtain all existing technologies in the field, and have the ability to apply conventional experimental means before that date. Those of ordinary skill in the art can improve and implement this scheme in combination with their own abilities under the enlightenment given by this application. Some typical known structures or known methods should not become obstacles for those of ordinary skill in the art to implement this application. It should be pointed out that for those skilled in the art, several variations and improvements can be made without departing from the structure of the present invention. These should also be regarded as the scope of protection of the present invention. These will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A multi-rotor UAV adaptive obstacle avoidance strategy generation system, characterized by: Includes the following modules: Multi-source environmental perception module: used to collect environmental data through the millimeter-wave radar and binocular vision sensors on the multi-rotor drone, and output a three-dimensional grid map and confidence matrix with a time stamp after spatiotemporal alignment; Dynamic obstacle modeling module: used to obtain the current posture data of the drone; A threshold is set based on the confidence matrix, and grid cell data with confidence lower than the threshold in the three-dimensional grid map is filtered to obtain valid environmental data; Based on cluster analysis, independent obstacle clusters are extracted from the valid environmental data. After removing the ground point cloud, the spatial distribution matrix of the independent obstacle clusters is output. The motion trajectory equation of each obstacle is fitted using the spatial distribution matrix, and the confidence matrix is ​​used as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model. Threat assessment and analysis module: used to output a motion state parameter set through a dynamic obstacle motion state model, the motion state parameter set including a velocity vector and an acceleration covariance matrix; after constructing a threat assessment model based on the current posture data of the UAV and the motion state parameter set, it outputs a threat assessment result, which includes the dynamic collision probability and collision time (TTC) value of each obstacle; Path generation optimization module: used to input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the UAV; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path according to the comprehensive score and generate the final obstacle avoidance strategy, which includes a control instruction set for the heading angle correction, speed adjustment coefficient and emergency braking threshold.

2. The adaptive obstacle avoidance strategy generation system for a multi-rotor UAV according to claim 1, characterized in that: The multi-source environment perception module specifically includes: Data acquisition submodule: used to obtain obstacle distance information through the millimeter-wave radar on the multi-rotor drone; Texture features are collected through the binocular vision sensor on the multi-rotor drone; Data processing submodule: used to align the two types of data in time and space, and output a three-dimensional grid map with a timestamp and a confidence matrix. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

3. The adaptive obstacle avoidance strategy generation system for a multi-rotor UAV according to claim 2, characterized in that: In the threat assessment and analysis module, the threat assessment model adopts a temporal neural network architecture, including: The first sub-network processes the temporal characteristics of the obstacle motion state parameter set; The second sub-network integrates the UAV pose data; The output layer generates dynamic collision probability and TTC value, where the dynamic collision probability is normalized to the [0,1] range through the Sigmoid function.

4. The adaptive obstacle avoidance strategy generation system for a multi-rotor UAV according to claim 3, characterized in that: The path generation optimization module also includes the following submodules: Path update submodule: used to initiate local path replanning when a low-confidence grid cell is suddenly added to the confidence matrix: based on the current optimal path, a local optimization area is delineated in the three-dimensional grid map; the path topology continuity is maintained within the optimization area, and waypoints are updated based on the gradient descent method; the path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the UAV.

5. A method for generating an adaptive obstacle avoidance strategy for a multi-rotor drone, characterized by: The following steps are involved: Multi-source environmental perception steps: Collect environmental data through millimeter-wave radar and binocular vision sensors, and output a three-dimensional grid map with a time stamp and a confidence matrix after spatiotemporal alignment; Dynamic obstacle modeling steps: obtain the current posture data of the UAV; set a threshold based on the confidence matrix, filter the grid cell data with confidence lower than the threshold in the 3D grid map, and obtain valid environmental data; Based on cluster analysis, independent obstacle clusters are extracted from the valid environmental data. After removing the ground point cloud, the spatial distribution matrix of the independent obstacle clusters is output. The motion trajectory equation of each obstacle is fitted using the spatial distribution matrix, and the confidence matrix is ​​used as the weighting coefficient of the motion trajectory equation to construct a dynamic obstacle motion state model. Threat assessment and analysis steps: outputting a motion state parameter set through a dynamic obstacle motion state model, wherein the motion state parameter set includes a velocity vector and an acceleration covariance matrix; constructing a threat assessment model based on the current posture data of the UAV and the motion state parameter set, and then outputting a threat assessment result, wherein the threat assessment result includes a dynamic collision probability and a collision time (TTC) value of each obstacle; Path generation optimization step: input the threat assessment results into the path generation network model, and generate N groups of candidate paths that meet the constraints under the maximum roll angle constraint of the UAV; simulate and deduce the candidate paths, and calculate the expected energy consumption, risk value and comprehensive score of task completion of each path based on the threat assessment results; select the optimal path according to the comprehensive score, and generate a final obstacle avoidance strategy, which includes a control instruction set for a heading angle correction, a speed adjustment coefficient and an emergency braking threshold.

6. The method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV according to claim 5, characterized in that: The multi-source environment perception step includes: Obtain obstacle distance information through millimeter-wave radar; Collect texture features through binocular vision sensors; The two types of data are aligned in time and space, and a three-dimensional grid map with a timestamp and a confidence matrix are output. The confidence matrix is ​​used to quantify the reliability of the sensor data of each grid cell.

7. The method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV according to claim 6, wherein: In the threat assessment and analysis step, the threat assessment model adopts a temporal neural network architecture, including: The first sub-network processes the temporal characteristics of the obstacle motion state parameter set; The second sub-network integrates the UAV pose data; The output layer generates dynamic collision probability and TTC value, where the dynamic collision probability is normalized to the [0,1] range through the Sigmoid function.

8. The method for generating an adaptive obstacle avoidance strategy for a multi-rotor UAV according to claim 7, characterized in that: The path generation optimization step also includes: When a low-confidence grid cell is suddenly added to the confidence matrix, local path replanning is initiated: Based on the current optimal path, a local optimization area is defined in the three-dimensional grid map; Maintain path topology continuity within the optimization area and update waypoints based on gradient descent method; The path update process must simultaneously meet the maximum roll angle constraint and centripetal acceleration limit of the drone.

Images