A laser radar-based unmanned aerial vehicle autonomous obstacle avoidance optimization method

Abstract

This invention relates to the field of autonomous obstacle avoidance technology for unmanned aerial vehicles (UAVs), and discloses an optimization method for autonomous obstacle avoidance of UAVs based on lidar. This method acquires real-time point cloud data streams of the UAV's flight environment through a lidar scanning module, which includes the three-dimensional position information and dynamic motion trajectories of obstacles. An obstacle spatial distribution model is constructed based on the real-time point cloud data stream to characterize the geometry and motion state of the obstacles. A reinforcement learning algorithm is used to calculate a collision risk value based on the position information and motion trajectory in the model, and an obstacle avoidance path optimization command is generated based on this risk value to adjust the UAV's flight direction and speed. Flight trajectory feedback data is recorded in real time during the execution of the command, and the parameters of the obstacle spatial distribution model are updated based on the feedback data. This method improves the adaptability and reliability of autonomous obstacle avoidance for UAVs in complex environments, ensuring flight safety.

Description

Technical Field

[0001] This invention relates to the field of autonomous obstacle avoidance technology for unmanned aerial vehicles (UAVs), specifically to an optimization method for autonomous obstacle avoidance of UAVs based on lidar. Background Technology

[0002] With the rapid development of drone technology, its applications in aerial photography, logistics transportation, power line inspection, and agricultural plant protection are becoming increasingly widespread. In these application scenarios, drones often need to fly autonomously in complex and ever-changing environments, and autonomous obstacle avoidance capability is one of the core elements to ensure drone flight safety. Currently, drone autonomous obstacle avoidance technology mainly relies on various sensors for environmental perception and algorithms to achieve obstacle recognition and path planning.

[0003] In existing technologies, some drones use visual sensors for environmental perception, acquiring and analyzing image data through cameras to identify obstacles. However, visual sensors are susceptible to environmental factors such as lighting conditions and weather. In scenarios with strong light, backlight, overcast skies, or fog, the clarity of image data decreases significantly, leading to reduced obstacle recognition accuracy and even missed or false detections. Other drones use ultrasonic or infrared sensors. While these sensors can detect obstacles at short distances, their detection range is limited, and they struggle to accurately acquire the three-dimensional position information of obstacles, failing to meet the obstacle avoidance requirements of drones flying at medium to long distances.

[0004] In obstacle modeling and risk assessment, traditional methods primarily focus on building models for static obstacles, mainly concentrating on their geometry and location, while neglecting the motion state of dynamic obstacles. In real-world flight environments, numerous dynamic obstacles exist, such as birds, other drones, and moving ground vehicles. The trajectories of these obstacles are uncertain. Relying solely on static models for collision risk assessment can easily lead to significant discrepancies between the assessment results and the actual situation, increasing the risk of collisions. Furthermore, existing obstacle avoidance path planning methods are mostly based on pre-defined algorithm models, and the generation of path optimization commands lacks dynamic response to real-time environmental changes. When the number of obstacles in the environment increases or their motion state suddenly changes, problems such as unreasonable obstacle avoidance paths and delayed adjustments can easily occur.

[0005] Existing systems often maintain fixed model parameters after initial setup, lacking an effective real-time feedback mechanism. During flight, the environment is constantly changing; new obstacles may appear, and the motion state of existing obstacles may also change. Fixed model parameters are unable to adapt to these dynamic changes, leading to a gradual decline in the system's obstacle avoidance performance and an inability to ensure the long-term stable flight safety of the UAV. Summary of the Invention

[0006] The purpose of this invention is to provide an optimization method for autonomous obstacle avoidance of unmanned aerial vehicles (UAVs) based on lidar, so as to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, this invention provides an optimization method for autonomous obstacle avoidance of unmanned aerial vehicles (UAVs) based on lidar, the method comprising:

[0008] The real-time point cloud data stream in the UAV flight environment is obtained by the lidar scanning module. The real-time point cloud data stream contains the three-dimensional position information and dynamic motion trajectory of obstacles.

[0009] An obstacle spatial distribution model is constructed based on the real-time point cloud data stream. The obstacle spatial distribution model is used to characterize the geometry and motion state of the obstacles.

[0010] A reinforcement learning algorithm is used to predict the collision risk value between the drone and the obstacle. The collision risk value is calculated based on the location information and motion trajectory in the spatial distribution model of the obstacle.

[0011] Based on the collision risk value, an obstacle avoidance path optimization instruction is generated, which is used to adjust the flight direction and speed of the UAV.

[0012] The system executes the obstacle avoidance path optimization command to control the drone's drive system and records flight trajectory feedback data in real time.

[0013] The parameters of the obstacle spatial distribution model are updated based on the flight trajectory feedback data.

[0014] Preferably, the step of acquiring real-time point cloud data streams in the UAV flight environment through the lidar scanning module includes:

[0015] Configure the scanning frequency and resolution parameters of the lidar sensor to ensure that the point cloud acquisition covers the preset monitoring area;

[0016] Noise filtering is performed on the raw point cloud data to remove environmental interference signals;

[0017] The filtered point cloud data is converted into a standardized three-dimensional coordinate sequence.

[0018] Preferably, the construction of the obstacle spatial distribution model includes:

[0019] The three-dimensional coordinate sequence is divided into multiple spatial grid units;

[0020] Perform point cloud clustering analysis within each spatial grid cell to identify obstacle boundary features;

[0021] Generate obstacle geometry description vectors and motion velocity estimates based on clustering results;

[0022] The outputs of all spatial grid cells are integrated to form the complete structure of the obstacle spatial distribution model.

[0023] Preferably, the step of using a reinforcement learning algorithm to predict the collision risk value between the drone and obstacles includes:

[0024] The obstacle geometry description vector and the estimated motion velocity are input into the state space of the reinforcement learning algorithm;

[0025] Design a reward function mechanism, which is calculated based on the relative distance between the current position of the UAV and the obstacle;

[0026] The reinforcement learning policy network is updated through iterative training, and the probability distribution of the collision risk value is output.

[0027] Preferably, the obstacle avoidance path optimization command includes:

[0028] Analyze the probability distribution of the collision risk values ​​to identify high-risk obstacle areas;

[0029] A dynamic path planning algorithm is used to calculate candidate schemes for safe flight paths, and the candidate schemes contain multiple path nodes;

[0030] Evaluate the feasibility score of each path node and select the path scheme with the highest feasibility score as the obstacle avoidance path optimization instruction;

[0031] The obstacle avoidance path optimization command includes specific directional angle adjustment values ​​and velocity changes.

[0032] Preferably, the drive system for controlling the drone includes:

[0033] The direction angle adjustment value and speed change are converted into motor control signals;

[0034] The motor control signal is sent to the UAV thruster via the drive interface module;

[0035] Simultaneously acquire thruster response data and real-time position data;

[0036] The thruster response data and real-time position data constitute a part of the flight trajectory feedback data.

[0037] Preferably, updating the parameters of the obstacle spatial distribution model based on the flight trajectory feedback data includes:

[0038] Compare the deviation values ​​between the real-time location data and the predicted path nodes;

[0039] If the deviation value exceeds the preset tolerance threshold, the division rule of the spatial grid unit is adjusted;

[0040] Re-execute the point cloud clustering analysis and update the obstacle geometry description vector;

[0041] The updated obstacle spatial distribution model is used for the next round of collision risk value prediction.

[0042] Preferably, the method further includes performing data preprocessing before constructing the obstacle spatial distribution model:

[0043] Extract timestamp information from the three-dimensional coordinate sequence to align the time dimension of the point cloud data;

[0044] Compensate for the scanning delay error of the LiDAR and generate a time-synchronized point cloud dataset.

[0045] Preferably, the method further includes optimizing the training process of the reinforcement learning algorithm, specifically:

[0046] Initialize the weight parameters of the policy network;

[0047] Simulate environmental interactions using historical flight datasets and calculate policy gradient update values;

[0048] The policy gradient update value is used to adjust the weights of the reward function mechanism.

[0049] Preferably, the method further includes an adaptive mechanism for handling dynamic obstacles, specifically:

[0050] The movement speed estimate is monitored for change trend. If the change trend exceeds the stable range, the path node is re-evaluated and a supplementary obstacle avoidance instruction is generated. The supplementary obstacle avoidance instruction is then merged into the obstacle avoidance path optimization instruction.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] This method acquires real-time point cloud data streams through a LiDAR scanning module. LiDAR is unaffected by lighting and weather conditions, and can stably output data containing the three-dimensional position information and dynamic motion trajectory of obstacles, enabling UAVs to perceive the flight environment more timely and comprehensively. Compared to methods relying on visual or ultrasonic sensors, this data acquisition method reduces the interference of environmental factors on the perception results, and can more accurately capture the positional changes and movement trends of obstacles, providing reliable basic data support for subsequent obstacle analysis and risk assessment.

[0053] When constructing the obstacle spatial distribution model, this method comprehensively considers the geometry and motion state of the obstacles, overcoming the limitations of traditional models that only focus on static features. Geometric information reflects the static attributes of obstacles such as size and outline, while motion state information reflects the dynamic characteristics of obstacles such as speed and direction. The combination of the two allows the model to more comprehensively depict the actual existence of obstacles in space. This comprehensive model description helps to more accurately judge the potential impact of obstacles on the flight path of UAVs, avoiding risk misjudgments caused by incomplete descriptions of obstacle states.

[0054] The algorithm employs reinforcement learning to predict collision risk values, fully leveraging its learning and adaptive capabilities in dynamic environments. Reinforcement learning continuously adjusts risk assessment strategies and parameters based on real-time data from an obstacle spatial distribution model, ensuring that collision risk value calculations dynamically adapt to environmental changes. When faced with complex multi-obstacle scenarios or sudden changes in obstacle motion, the algorithm can quickly respond and update risk assessment results, ensuring that the risk value accurately reflects the collision probability under the current flight environment and providing timely decision-making support for path optimization.

[0055] The obstacle avoidance path optimization commands generated based on collision risk values ​​can specifically adjust the drone's flight direction and speed. This adjustment is not based on a fixed obstacle avoidance pattern, but rather combines the real-time movement of obstacles and the changing trends of collision risk, making the drone's obstacle avoidance maneuvers more flexible and targeted. For example, when an obstacle approaches rapidly, the command can promptly adjust the flight direction to avoid a collision; when the obstacle moves slowly, the command can optimize the flight speed to reduce unnecessary detours, improving flight efficiency while ensuring safety.

[0056] After executing the obstacle avoidance path optimization command, the flight trajectory feedback data is recorded in real time and used to update the obstacle spatial distribution model parameters, forming a closed-loop dynamic optimization mechanism. The flight trajectory feedback data contains environmental interaction information during the actual obstacle avoidance process of the UAV, which can reflect the deviations and deficiencies of the model in practical applications. By integrating the feedback data into the model parameter update process, the obstacle spatial distribution model can continuously learn new environmental features and obstacle behavior patterns, constantly optimizing the accuracy of obstacle description and risk prediction capabilities. This continuous updating mechanism enables the system to adapt to different flight scenarios, maintaining good obstacle avoidance performance in urban building clusters, natural terrain, or areas with dense dynamic obstacles, thus enhancing the stability and safety of the UAV's autonomous flight in complex environments. Attached Figure Description

[0057] Figure 1 This is a schematic diagram illustrating the working principle of the autonomous obstacle avoidance optimization method for UAVs based on lidar as described in this invention.

[0058] Figure 2 Flowchart for constructing a spatial distribution model of obstacles;

[0059] Figure 3 Clustering analysis diagram of lidar point cloud;

[0060] Figure 4 A flowchart for generating obstacle avoidance path optimization instructions;

[0061] Figure 5 This is a diagram for path planning and collision risk assessment.

[0062] Figure 6 A flowchart for updating the parameters of the obstacle spatial distribution model. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] Please see Figure 1 This invention provides an optimization method for autonomous obstacle avoidance of unmanned aerial vehicles (UAVs) based on lidar, the method comprising:

[0065] A point cloud data stream, containing the 3D position information and dynamic trajectory of obstacles, is continuously acquired from the UAV's flight environment using a LiDAR scanning module. This data stream is used to construct an obstacle spatial distribution model, representing the geometry and motion of the obstacles. A reinforcement learning algorithm is employed to predict the collision risk value between the UAV and the obstacles, calculated from the model's coordinates and trajectory. Obstacle avoidance path optimization commands are generated based on the risk value, adjusting the UAV's flight direction and speed. These commands are executed to control the UAV's drive system, and flight trajectory feedback data is recorded in real time. The parameters of the obstacle spatial distribution model are updated based on the feedback data, achieving adaptive optimization.

[0066] Example 1: See Figure 2The configuration and data processing of the LiDAR scanning module are fundamental components of the UAV autonomous obstacle avoidance optimization method. This module uses high-precision sensors to collect point cloud data streams from the UAV's flight environment in real time, providing raw input for subsequent obstacle detection and path planning. The LiDAR sensor's scanning frequency is set to an adjustable range of 50Hz to 200Hz to adapt to the environmental perception requirements at different flight speeds. The scanning angle covers 360 degrees horizontally, and the vertical field of view is set to 30 degrees, ensuring all-round monitoring of the space around the UAV. The resolution parameters are adjusted according to the actual application scenario, with a typical configuration of 0.1-degree angular steps, achieving a lateral resolution of approximately 1.7 centimeters at a distance of 10 meters.

[0067] After the raw point cloud data is acquired, it needs to be preprocessed to eliminate environmental noise interference. A Gaussian filtering algorithm is used to smooth the point cloud data, with a filter window size of 3×3 pixels and a standard deviation parameter set to 0.5. This processing can effectively suppress outliers caused by atmospheric scattering, sensor noise, and other factors, while preserving the geometric features of real obstacles. The filtered point cloud data is then unified into a coordinate system centered on the UAV body through a coordinate transformation module. The coordinate system is defined with the X-axis representing the UAV's forward direction, the Y-axis representing the rightward direction, and the Z-axis pointing vertically downwards. The three-dimensional coordinate values ​​of each point are stored in floating-point format with millimeter-level precision.

[0068] The standardized 3D coordinate sequence then proceeds to the spatial mesh generation stage. The monitoring space is divided into uniform cubic mesh units, with a unit size of 10 cm × 10 cm × 10 cm. This size selection comprehensively considers computational efficiency and environmental characterization accuracy, ensuring accurate description of the spatial distribution of obstacles while maintaining real-time performance. Each mesh unit maintains an independent point cloud data index, facilitating subsequent clustering analysis. After mesh generation, the system performs density clustering analysis on the point cloud within each unit. An improved DBSCAN algorithm is used for clustering, which can automatically identify spatially continuous point cloud clusters while filtering out discrete noise points. The algorithm parameters are set to a neighborhood search radius of 0.5 meters and a minimum threshold of 5 points to form a cluster.

[0069] Each point cloud cluster output by clustering analysis represents a potential obstacle entity. Geometric features are extracted from these point cloud clusters to generate an obstacle description vector. The description vector comprises three parts: size features, surface curvature features, and boundary features. Size features are obtained by calculating the maximum span of the point cloud cluster in the X, Y, and Z directions; surface curvature features are obtained by calculating the rate of change of the normal vector after fitting a local plane to the point cloud; and boundary features are described by extracting the convex hull vertices of the point cloud cluster. Simultaneously, the velocity of the same obstacle is calculated by comparing the positional changes of the same obstacle in two adjacent frames of point cloud data. Velocity estimation uses the least squares method to fit the displacement-time relationship, outputting a three-dimensional velocity vector.

[0070] The clustering results of all grid cells are integrated to form a complete obstacle spatial distribution model. The integration process includes establishing a global spatial index and topological relationships. Each obstacle entity is assigned a unique identifier, and its grid cell location is recorded. The model is organized using an octree data structure to achieve fast spatial query and update. For dynamic obstacles, the model additionally maintains their historical trajectory data to predict their future positions.

[0071] The obstacle spatial distribution model is updated in sync with LiDAR scanning. Each time new scan data arrives, the system first matches the obstacle positions predicted by the current model with the actual detection results. Successfully matched obstacles update their geometric features and motion states; unmatched predictions are considered as obstacles disappearing; newly detected point cloud clusters are added to the model as new obstacles. This dynamic update mechanism enables the model to accurately reflect real-time changes in the environment.

[0072] During model building, the system employs a multi-level verification mechanism to ensure data reliability. The initial verification checks the geometric rationality of the cloud clusters, filtering out detection results that do not conform to physical laws. The intermediate verification judges the authenticity of obstacles through the consistency of tracking across multiple consecutive frames. The advanced verification, combined with the UAV's motion state, evaluates the degree of agreement between the detection results and inertial measurement data. Only obstacles that pass all verification stages are included in the final spatial distribution model.

[0073] Spatiotemporal alignment of LiDAR data is crucial for model accuracy. The system employs a hardware timestamp synchronization mechanism to ensure a strict correspondence between point cloud data and UAV pose information. For point cloud distortion caused by the UAV's own motion during scanning, a motion compensation algorithm is used for correction. This algorithm utilizes high-frequency inertial measurement unit data to reconstruct the precise acquisition time and spatial position of each LiDAR scanning point.

[0074] The output of the obstacle spatial distribution model provides environmental perception data for the subsequent path planning module. The model not only includes the spatial location information of static obstacles but also accurately describes the motion state of dynamic obstacles. For each obstacle, the model provides complete information such as its geometric dimensions, surface shape, motion speed, and predicted trajectory. This data is organized in a structured form, supporting efficient spatial querying and collision detection calculations.

[0075] The point cloud data processing workflow is accelerated using a parallel computing architecture. The reception, filtering, and coordinate transformation of LiDAR data are executed on dedicated hardware processing units to ensure real-time performance. Computationally intensive tasks such as spatial mesh generation and cluster analysis are allocated to multi-core processors for parallel processing. The system employs a task pipeline design, with data transferred between different processing stages via a circular buffer to minimize processing latency.

[0076] The accuracy of environmental modeling directly impacts the quality of obstacle avoidance decisions. The system optimizes modeling performance through an adaptive parameter adjustment mechanism. When complex scenes are detected, the system automatically reduces the mesh cell size and cluster radius to improve modeling accuracy; in open environments, these parameters are appropriately increased to reduce computational load. This adaptive mechanism enables the system to maintain optimal modeling performance under various flight conditions.

[0077] Calibration and calibration of lidar sensors are prerequisites for ensuring data quality. During system startup, the system automatically executes the sensor calibration process, including internal parameter calibration and external installation location calibration. The calibration process utilizes a target object with known geometric features to optimize calibration parameters by minimizing reprojection error. During daily operation, the system continuously monitors the sensor status and prompts for maintenance when performance degradation is detected.

[0078] The obstacle spatial distribution model also possesses incremental update capabilities. For continuously tracked obstacles, the model employs a Kalman filter algorithm to smooth their trajectory, reducing the impact of measurement noise. The filter's state vector contains position, velocity, and acceleration information, and the system dynamically adjusts process noise parameters to adapt to different motion modes. This processing significantly improves the accuracy of dynamic obstacle state estimation.

[0079] The effective use of point cloud data is also reflected in the identification of special scenarios. By analyzing point cloud distribution patterns, the system can identify obstacle types that are difficult for traditional sensors to detect, such as glass curtain walls and barbed wire. For these special obstacles, the model will label their special attributes for the decision-making module to handle accordingly. At the same time, the system can also identify ground features and ceiling structures, providing three-dimensional spatial positioning references for drones.

[0080] The entire environmental modeling process forms a closed-loop quality control system. The system monitors the output indicators of each processing stage in real time, and automatically triggers data reprocessing or parameter adjustment when an anomaly is detected. Quality control indicators include dozens of dimensions such as point cloud density, clustering success rate, and trajectory continuity, ensuring the reliability of the modeling results. This closed-loop mechanism enables the system to adapt to operational requirements under various complex environmental conditions.

[0081] See Figure 3 It demonstrates the entire process of lidar point cloud processing, including four key analytical views:

[0082] Raw point cloud data distribution: Displays the raw point cloud acquired by LiDAR, including ground points, obstacle points, and noise points; Point cloud clustering and boundary recognition: Shows the results after processing by the DBSCAN clustering algorithm, with different gray levels representing different obstacles; 3D spatial point cloud distribution: Displays the spatial distribution of the point cloud from a 3D perspective, including height information; Obstacle spatial modeling and motion trajectory: Displays the obstacle modeling results after mesh generation and the motion trajectories of dynamic obstacles. This verifies the system's ability to identify static and dynamic obstacles in complex environments, providing an accurate environmental model for autonomous obstacle avoidance.

[0083] Example 2: See Figure 4 The collision risk prediction process of the reinforcement learning algorithm receives obstacle description vectors from the environment modeling module as its core input. These description vectors contain geometric features and motion state data. Geometric features include dimensions, surface curvature, and boundary vertex coordinates, while motion state data includes a three-dimensional velocity vector and estimated acceleration. The system normalizes these features into a numerical sequence with uniform dimensions, combining them to form a multi-dimensional state vector. The dimensions of the state vector are dynamically adjusted according to the scene complexity, with a typical configuration containing 15 to 30 feature dimensions.

[0084] A deep Q-network framework is used to construct a multi-layer neural network to process state inputs. The network structure adopts a fully connected design, with the number of nodes in the input layer consistent with the dimension of the state vector. Two processing layers are set in the intermediate hidden layers, each containing 64 neurons, and the ReLU nonlinear activation function is used. The number of nodes in the output layer corresponds to the discretized action space, including 8 basic flight directions and 4 speed levels, for a total of 32 action combinations. The neural network weights are initialized using a randomized, small-value strategy to avoid prediction bias in the initial state.

[0085] The reward function is designed based on the spatial relationship between the drone and obstacles. The system calculates the minimum Euclidean distance from each obstacle surface to the drone's fuselage in real time. A negative gradient mechanism is used for the reward value: a high penalty is applied when the minimum distance is below 0.5 meters; a linearly decreasing penalty value is applied in the range of 0.5 meters to 2 meters; and no reward or penalty is applied when the distance exceeds 5 meters. An additional time penalty factor is introduced, applying a small penalty if the target point is not reached after each decision cycle. The reward signal is transmitted to the network weight update process through a time difference algorithm.

[0086] The policy network training employs an experience replay mechanism. The system establishes a fixed-capacity memory to store state transition records; each record includes the current state, the action performed, the reward received, and the state after the transition. During each network update, a batch of historical records is randomly sampled to avoid the influence of temporal correlations between data. The network weights are adjusted based on the Bellman optimal equation to calculate the target Q-value, and the difference between the target value and the predicted value is used to calculate the loss function through mean squared error. The optimization algorithm uses an adaptive moment estimation method to adjust the learning rate and control the magnitude of parameter updates.

[0087] The collision risk value is output in the form of a probability distribution. After calculating the Q-value corresponding to 32 action selections, the system converts it into a probability distribution using the Softmax function. The probability values ​​are mapped to a three-dimensional physical space to form a risk heatmap, which is divided into spatial regions centered on the current location of the UAV. High-risk areas are defined as the physical space range where the probability peak exceeds a preset threshold, and the system marks the boundary coordinates and the direction of risk gradient change of this region.

[0088] The path planning module activates a dynamic obstacle avoidance algorithm based on a risk heatmap. First, a three-dimensional path search space is constructed, discretized into a cubic grid with 1-meter sides. The search starting point is set to the UAV's current position, and the ending point is the preset flight target point. The system evaluates the passage cost for each grid node; the cost is a composite of the basic path length, risk heatmap value, and motion energy consumption. The weighting coefficient of the risk heatmap value is dynamically adjusted based on the peak value of the heatmap; the higher the peak value, the greater the weighting ratio.

[0089] The search algorithm unfolds in a 3D mesh space. It maintains two sets of nodes: open and closed, and iteratively expands neighboring nodes starting from the starting point. The heuristic function value for each node uses Euclidean distance to guide the search towards the target. A node selection mechanism comprehensively evaluates the actual cost from the starting point to the node and the estimated cost to the destination, prioritizing the expansion of nodes with the lowest total cost. During the search, mesh cells marked as high-risk areas are skipped; when it is necessary to traverse a high-risk area, the path node density is automatically increased to improve accuracy.

[0090] After candidate paths are generated, the feasibility assessment phase begins. The system performs segmented verification on each candidate path: calculating whether the radius of curvature at each turning point of the path conforms to the UAV's maneuverability; verifying whether the path altitude exceeds flight airspace restrictions; and checking whether the total path length is within the range supported by remaining energy. Each path segment is scored out of 100, with path safety accounting for 40%, energy efficiency for 35%, and time cost for 25%. The scoring process considers the spatiotemporal intersection probability of the predicted trajectories of dynamic obstacles and the path's timeline.

[0091] Once the optimal path is determined, it is converted into specific control commands. The system calculates the direction vector of the initial segment of the path and converts it into a yaw angle relative to the drone's nose direction, with an angle output accuracy of 0.1 degrees. The speed control is automatically adjusted according to the path curvature: the maximum cruising speed is used for straight sections, and the speed limit is calculated based on the radius of curvature for turning sections. The final generated commands include instantaneous azimuth angle adjustments and acceleration commands. The azimuth angle adjustment range covers an omnidirectional space from -180 degrees to +180 degrees, and the speed adjustment range is continuously variable from 0 to 5 m / s².

[0092] The real-time control layer employs a sliding window mechanism to execute the path. The system continuously tracks the deviation between the UAV's actual position and the planned path, setting a 2-meter distance tolerance threshold. When the deviation exceeds the threshold or a sudden obstacle is detected, the path planning module interrupts the current command sequence and immediately restarts the risk assessment and path search process. Under normal conditions, the system refreshes the control commands for the next path segment every 200 milliseconds, achieving high-frequency dynamic response.

[0093] The reinforcement learning policy network features a periodic update mechanism to support online learning. The system collects state-action-reward data sequences from actual flight operations and trains by batch replaying them during nighttime maintenance periods. A conservative learning rate parameter is set for network weight updates, with each update not exceeding 1% of the original weights to ensure a smooth transition in policy evolution. New training data undergoes multiple validations: verifying the consistency between the state data and the environment model; confirming that action selection meets physical constraints; and confirming that reward calculations align with real-time monitoring results.

[0094] The accuracy monitoring of collision risk prediction employs 3D trajectory reconstruction technology. The system records the complete flight path after each obstacle avoidance decision. In the offline phase, the path is reprojected onto the obstacle spatial distribution model to verify whether the minimum distance between the actual flight path and the obstacle surface meets expectations. For cases with excessive prediction deviation, the system traces back the neural network's processing of obstacle feature vectors to pinpoint the root cause of perception misjudgment or decision-making defects.

[0095] The path optimization module's specialized scenario handling capabilities cover common flight conditions. For dense obstacle clusters, the system automatically switches to a random sampling method, uniformly generating path samples within the feasible space to replace a global search. For narrow passages, it increases position control precision requirements and reduces flight speed. For high-speed moving obstacles, it employs predictive position interpolation technology to construct a temporal-spatial composite risk model. The identification of each scenario is based on a specific combination of obstacle distribution characteristics and motion patterns, triggering corresponding parameter configuration schemes.

[0096] The output format of control commands is compatible with various UAV platforms. The heading angle adjustment value conforms to the standard definition of heading angle, maintaining consistency with the heading reference of mainstream flight control systems. Speed ​​commands offer two output modes: relative speed change and absolute speed value, adapting to the interface protocols of different manufacturers. The system is compatible with both pulse width modulation signals and serial data buses as physical transmission methods, with a data transmission cycle controlled within 50 milliseconds, meeting the real-time control requirements of highly dynamic environments.

[0097] The decision-making system employs a multi-layered anomaly handling mechanism for state monitoring. During operation, it continuously monitors the validity of input data: confirming that the environment model is updated periodically; verifying that the number of obstacles is within a reasonable range; and monitoring whether the probability distribution of the neural network output satisfies normalization characteristics. When an abnormal interruption of input data to the decision-making module is detected, it immediately switches to a simplified obstacle avoidance strategy: stopping forward flight, performing a stationary rotation scan, and re-entering the main path planning process after the environmental data regains continuity.

[0098] See Figure 5 The results of collision risk assessment and path planning based on reinforcement learning are shown: Collision risk heatmap and path planning: showing the distribution of collision risks in the environment (the darker the color, the higher the risk) and the planned safe path; 3D path planning view: showing the relationship between the UAV's planned path and the risky terrain from a 3D perspective, verifying the effectiveness of reinforcement learning algorithm in real-time risk assessment and path planning in complex dynamic environments.

[0099] Example 3: See Figure 6 The execution of obstacle avoidance path optimization commands begins with the conversion of control signals. The steering angle adjustment value is input to a proportional-integral-derivative (PID) controller, which contains three independent channels to handle pitch, roll, and yaw angles respectively. The proportional element directly responds to angle deviations, the integral element accumulates historical deviations, and the derivative element predicts trend changes. The current adjustment value corresponds to the speed change input, which is mapped to the motor control quantity through a voltage-thrust curve. The controller outputs two signals: the steering channel outputs a pulse-width modulated waveform with a duty cycle proportional to the angle correction value; the speed channel outputs an analog voltage signal with an amplitude matched to the acceleration requirements.

[0100] The drive interface module uses a serial communication protocol to transmit control signals. The protocol frame structure includes a start bit, address code, data field, and parity bit. The data field is divided into 32 bytes to store 16-bit pulse width and 16-bit voltage values. The transmission baud rate is fixed at 115200bps, and a complete command frame is sent every 50 milliseconds. The interface hardware implements electrical isolation to prevent electromagnetic interference from the thruster from affecting the control core. The module has a built-in acknowledgment detection mechanism that triggers redundant retransmission if no confirmation is received from the device.

[0101] The UAV propulsion system receives commands and executes them accordingly. Four brushless motors receive their respective pulse-width modulation signals, which are converted into three-phase drive current by an electronic speed controller. The motor speed variation rate is limited to within 2000 revolutions per second to avoid mechanical shock. Propulsion response data is collected synchronously: each motor is equipped with a Hall sensor to measure the actual speed, a current sensor to monitor the phase load, and a temperature sensor to record the operating status. The response data is sampled with 12-bit precision at a sampling frequency of 1kHz and transmitted back in real time via a dedicated data bus.

[0102] The position tracking system operates independently in the feedback loop. The onboard high-precision global positioning system module, model MS-1521, continuously outputs three-dimensional coordinates with a positioning update frequency of 10Hz and a horizontal error radius of 0.1 meters. An inertial measurement unit (IMU) assists in motion compensation, including a three-axis microelectromechanical gyroscope and accelerometer, with a data fusion frequency of 200Hz. Real-time position data is fused from multiple sources using a Kalman filter to output the fuselage center point coordinates and attitude angles, achieving an attitude measurement accuracy of 0.1 degrees. Thruster response data is merged with real-time position data into a flight trajectory feedback data packet, with its transmission cycle synchronized with the control system clock.

[0103] The parameter updates of the obstacle spatial distribution model are triggered by trajectory deviation. The system compares the spatial differences between the actual flight position and the predicted path nodes, and calculates the deviation value using three-dimensional Euclidean distance.

[0104]

[0105] in, This indicates the positional deviation value. These are the actual coordinates. This predicts the node coordinates. This is based on three consecutive sampling periods. When the deviation reaches a certain value, the model reconstruction program is initiated. The deviation threshold setting has dynamic adjustment capabilities and can be reduced to 0.3 meters in complex environments.

[0106] The spatial grid cell division rules are adjusted using a progressive refinement strategy. The initial grid is divided from 10 cm cubes into 5 cm cells, with the division depth limited to three levels or less. Each subdivided grid inherits the point cloud index of its parent cell and concurrently launches four computation threads for parallel processing. An automatic grid density balancing algorithm allocates computational resources based on point cloud distribution characteristics: the smallest cell size is used in areas with dense obstacles, while coarse-grained division is retained in open spaces. The grid topology is updated in real time with an adjacency table to maintain spatial continuity.

[0107] The point cloud clustering analysis process has been re-executed, and the algorithm parameters have been upgraded. The neighborhood search radius ε has been reduced from the baseline of 0.5 meters to 0.3 meters, and the minimum number of cluster points has been increased from 5 to 10. A normal vector consistency check has been introduced, requiring that the angle between the normals of point clouds within the same cluster be less than 30 degrees. Boundary integrity checks have been added to the clustering results to exclude point cloud clusters containing internal holes. The new geometric shape description vector generation process enhances the surface continuity features, adding a surface concavity / convexity index based on the original size and curvature features. This index is achieved by detecting the rate of change of point cloud density using ray tracing.

[0108] An updated obstacle spatial distribution model establishes a version management mechanism. Each model snapshot records the generation timestamp and parameter configuration, and historical versions are stored at a depth of the last five iterations. Model switching employs a smooth transition algorithm to avoid abrupt changes in geometric description: the weights of the new model increase linearly from 0.2 to 1.0 with a transition period of 200 milliseconds. When reconstructing the motion trajectory of dynamic obstacles, velocity trend data from historical models are integrated, and acceleration abrupt changes are smoothed through the second derivative.

[0109] The model accuracy verification system operates in parallel. Virtual probe points are evenly distributed in the model space. For each point, the theoretical value of the nearest obstacle surface is calculated and compared with the actual value matched by the point cloud. The verification report automatically generates a defect heatmap, marking areas with spatial location deviations exceeding 10%. These abnormal areas are prioritized for analysis in the next round of data processing, with the local grid density increased to 1 cm accuracy for repeated clustering.

[0110] The control execution process employs a closed-loop monitoring architecture. The actuator status watchdog circuit operates independently, monitoring the corresponding delay between the motor control signal and the response data. When the time difference between the steering command and the change in body attitude exceeds 80 milliseconds, or when the difference between the acceleration command and the actual speed fails to converge for more than 0.3 seconds, dynamic gain adjustment of the control loop is triggered. The adjustment process automatically updates the integral time constant of the proportional-integral-derivative controller based on the real-time monitored load inertia.

[0111] The system employs a dual-timebase configuration for environment-adaptive update cycle management. The basic update cycle is locked to the LiDAR scanning interval, forcibly synchronized with data acquisition. The emergency update channel has preemptive capabilities; when continuous position deviation alarms or sudden obstacle warnings occur, the regular processing flow is interrupted to immediately initiate a priority update. All model update operations implement a resource pre-allocation strategy, distributing computational tasks across four processing cores to balance the load, compressing the time of a single update to within 8 milliseconds.

[0112] The spatiotemporal consistency of model data is maintained using a prediction correction mechanism. After each model update, the extrapolated trajectory of each obstacle is reconstructed synchronously. The trajectory prediction is based on the current velocity vector, superimposed with acceleration perturbation compensation from the UAV collected by the inertial measurement unit. The prediction results are stored in a spatial index tree for collision pre-detection before the next decision cycle. When a potential conflict is detected by the pre-detection, the path replanning stage is initiated in advance, forming a closed-loop response chain from model update to decision optimization.

[0113] The real-time performance monitoring of the point cloud processing system includes 22 metrics. Key metrics include grid processing latency, clustering completion rate, and feature extraction error count. These metrics record the most recent 100 periods of data through a circular buffer, automatically generating an operational baseline. When a single-period metric deviates from the baseline by 30%, a diagnostic analysis process is triggered: checking hardware interrupt response timeliness, verifying memory bandwidth usage, analyzing algorithm branch prediction failure rate, and generating optimization parameter adjustment suggestions.

[0114] The coordination mechanism between control commands and model updates sets priority levels. Model updates receive medium-priority computing resources during normal flight; control command execution occupies the highest priority during obstacle avoidance maneuvers; model reconstruction tasks are limited to processor idle periods. The resource allocation manager implements a hard real-time scheduling strategy to ensure that critical control cycles maintain a precise 50-millisecond time window.

[0115] Example 4: In the preprocessing stage of LiDAR point cloud data, time synchronization is a crucial step in ensuring the accuracy of environmental modeling. When the system acquires raw point cloud data from the LiDAR hardware interface, each point cloud frame is accompanied by a timestamp accurate to the microsecond level. The timestamp is generated based on the hardware clock synchronized by the Global Positioning System (GPS) and is kept synchronized with the UAV's inertial navigation system. The time alignment processing of the point cloud data adopts a three-level buffer mechanism: the raw data buffer queue stores unprocessed point cloud frames; the time calibration module marks the offset of each frame of data relative to the system's unified time reference; and the synchronization output queue reassembles the point cloud sequence in chronological order.

[0116] Compensating for scanning delay in lidar requires precise measurement of the time delay characteristics at each stage. The system maintains a delay parameter table, recording the end-to-end time delay components from laser emission to data output. This table includes key parameters such as sensor internal processing delay, data transmission delay, and coordinate transformation calculation delay. Each delay component is determined through specialized calibration experiments, and compensation curves are fitted at different operating temperatures. Typical delay parameters are shown in Table 1.

[0117] Table 1: Typical delay parameters.

[0118]

[0119] The time-synchronized point cloud dataset is then used in the reinforcement learning training process. The policy network weights are initialized using a hierarchical randomization strategy: input layer weights follow a uniform distribution, hidden layer weights are initialized using a truncated normal distribution, and output layer weights are set to tiny random values. The number of input layer nodes dynamically matches the dimension of the state vector, and the hidden layers use a residual connection structure to avoid gradient vanishing. The historical flight dataset contains millions of state-action-reward records, with each record storing a 128-dimensional state vector, a 32-dimensional action selection vector, and a scalar reward value.

[0120] The simulated environment interactive training employs an asynchronous parallel architecture. Eight computing nodes run the environment simulator simultaneously, with each simulator instance loading different scene configuration parameters. Scene parameters include variables such as obstacle density, movement speed range, and spatial distribution pattern. During training, the simulator generates state transition data, which is then passed to the learning algorithm via a shared memory queue. Policy gradient updates are calculated using a sliding window averaging method, with the gradient direction calculated by sampling from the most recent 1000 training batches each time.

[0121] The reward function mechanism adjusts weights based on policy performance evaluation. The system defines ten evaluation metrics to monitor policy evolution, including average obstacle avoidance distance, path efficiency index, and motion smoothness. After every 500 training iterations, the validation module runs the current policy on an independent test set and generates a policy performance report. The weight adjustment algorithm analyzes the contribution of each reward component to the final policy and dynamically adjusts the coefficient ratios of different reward items. An early stopping mechanism is implemented during training; the current training phase is automatically terminated if no significant improvement is observed after three consecutive validation cycles.

[0122] The network weight update process implements strict numerical stability control. Before each parameter update, the norm of the gradient values ​​is checked, and gradient pruning is automatically enabled if it exceeds a threshold. The learning rate employs a cosine annealing scheduling strategy, with an initial value set to a moderate size, gradually decreasing as training progresses. A rolling backup mechanism is used to save network parameters, retaining the ten most recent training checkpoints, allowing for rapid restoration to a previous stable version in the event of performance regression.

[0123] Quality monitoring of training data involves multiple verification steps. Raw state data must undergo a reasonableness check: verifying that the numerical range is within the sensor's measurement range; verifying the physical consistency between data across different dimensions; and confirming the continuity of the time series. Action selection records must be reviewed to ensure they comply with UAV dynamic constraints, including maximum turning angular velocity and acceleration limits. Reward value calculations must be re-verified against the original observation data to ensure strict correspondence with the environmental conditions at the time of real-time decision-making.

[0124] The simulation environment configuration emphasizes scene diversity and physical realism. The basic scene library contains twenty typical environment templates, covering different spatial structures such as open airspace, urban canyons, and indoor corridors. Each template allows adjustment of ten parameters to generate variant scenes, including obstacle material properties, lighting conditions, and atmospheric disturbance intensity. The physics engine accurately simulates the aerodynamic characteristics of the drone, taking into account parameters such as airframe size, weight distribution, and thruster thrust curves.

[0125] Resource management during training employs a dynamic allocation strategy. Computing nodes allocate GPU resources based on task priority, with core training tasks dedicating their own high-performance computing cards, while data preprocessing tasks share mid-level computing units. Memory usage is managed using paging; frequently accessed training data resides in memory, while historical checkpoints are stored on high-speed solid-state drives. Network communication utilizes a hybrid protocol; large-capacity data transmission uses remote direct memory access (RDBMI), and control commands are rapidly exchanged via User Datagram Protocol (UDP).

[0126] Pre-deployment validation testing includes comprehensive functional checks. Thousands of random scenario tests are run in a simulation environment, recording key metrics such as collision rate, path deviation, and control stability. Hardware-in-the-loop testing deploys the policy network onto actual flight control hardware, simulating sensor inputs through signal injection to verify real-time response capabilities. The final deployment package generates differentiated policy versions, automatically selecting appropriate network structure simplification schemes for hardware platforms with varying computing capabilities.

[0127] The online learning mechanism is designed with practical operational constraints in mind. Newly collected flight data undergoes rigorous screening before being added to the training pool, prioritizing samples containing rare scenarios. Incremental training limits computational resource usage to ensure no impact on real-time control tasks. Policy updates employ a canary release model, first testing new strategies on select drone nodes to verify stability before full rollout. A version rollback function remains readily available, allowing for rapid switching to the previous stable version when performance degradation is detected.

[0128] Continuous monitoring of time synchronization accuracy employs a hardware-assisted approach. A high-precision time interval measuring instrument is installed on the UAV body to directly measure the time deviation between lidar data acquisition and inertial measurement unit sampling. The measurement results are fed back to the time calibration module, dynamically adjusting compensation parameters. A temperature sensor network monitors the operating temperature of each hardware module in real time and predicts delay change trends based on temperature-delay characteristic curves.

[0129] Spatial calibration and temporal synchronization of point cloud data are performed collaboratively. Each point cloud frame includes aircraft attitude data, and the original point cloud is transformed to a unified world coordinate system using a coordinate transformation matrix. The calculation of the transformation matrix takes into account temporal interpolation to accurately match the UAV pose at the moment of point cloud acquisition. The calibrated point cloud data is stored in an octree structure, supporting efficient spatial range queries and nearest neighbor searches. Data storage uses a compressed format to reduce memory usage while maintaining accuracy.

[0130] The reproducibility of the reinforcement learning training environment is achieved through seed control. A fixed seed value for the random number generator is used in each training cycle to ensure consistent training trajectories from the same input. The simulator's runtime configuration is saved as a description file, recording all parameter settings affecting policy performance. The training log contains a complete snapshot of the environment state and decision records, supporting precise replay analysis of any training step. This design facilitates the identification of policy flaws and the optimization of the training process.

[0131] Example 5: The operational basis of the dynamic obstacle adaptive mechanism is the continuous monitoring of the changing characteristics of obstacle motion states. The system periodically extracts the three-dimensional velocity vector data of each identified obstacle from the obstacle spatial distribution model. The sampling interval is strictly synchronized with the LiDAR scanning cycle, typically 10 milliseconds. The velocity estimation sequence of each obstacle first undergoes outlier removal: the moving standard deviation of five consecutive sampling points is calculated, and instantaneous data exceeding three times the standard deviation are considered noise and discarded. Linear interpolation is used to fill in the missing values. The preprocessed velocity sequence is input into a moving average filter with a fixed filter window size of five sampling periods, and a smoothed velocity trend curve is output.

[0132] The quantitative analysis of velocity change trends employs a first-order difference algorithm. The system calculates the rate of change of the vector magnitude of two adjacent filtered velocity values ​​and takes the weighted average of three consecutive difference values ​​as the final trend indicator. When the absolute value of this indicator exceeds 2 m / s², it is determined to enter an unstable state, triggering a dynamic response process. The judgment logic is set with a two-stage confirmation mechanism: after the first exceedance, a 100-millisecond observation period is initiated, and the state switch is only confirmed after three consecutive exceedances during this period. Specially marked volatile obstacles (such as birds or swaying objects) use an independent judgment threshold, and their rate of change threshold is relaxed to 4 m / s².

[0133] The path node reassessment process immediately interrupts the current path tracking process. The system saves all node data of the current path plan to the buffer stack and freezes the drone in hover at the current position (altitude error controlled within ±0.3 meters). When the dynamic path planning algorithm is reactivated, it loads the updated spatial distribution model, focusing on expanding the search space around obstacles with sudden velocity changes. The weight adjustment of high-risk areas adopts an incremental overlay strategy: the weight coefficient is increased by 20% on the original weight, and the adjustment range generates concentric spherical layers centered on the obstacle. The area within 1 meter of the obstacle is the high-risk core area, and 1-3 meters is the buffer transition area. The search grid density is tripled in the core area, and the grid side length is reduced to 0.3 meters.

[0134] The supplementary obstacle avoidance command generation mechanism includes a three-level correction logic. The primary correction calculates the intersection angle between the extended line of the obstacle's motion trend and the current heading, outputting a fine-tuning value of the azimuth angle within a range of ±5 degrees. The intermediate correction predicts the time window for the UAV to pass through the danger zone, combining this with the sudden speed change to calculate the collision time margin; when the margin is less than 1 second, a speed correction command of ±0.5 m / s is triggered. The advanced correction, for scenarios with multiple obstacles, establishes a local cost map of the optimal avoidance path and solves for the optimal avoidance direction. The correction parameters undergo dynamic feasibility verification: the direction adjustment value is checked to ensure it does not exceed the maximum turning rate limit; the speed adjustment value is verified to ensure it is within the feasible range of motor thrust.

[0135] Command merging and execution employs a priority overriding strategy. The main obstacle avoidance path optimization command queue maintains its original structure, while supplementary commands are added before the first command via an insertion mechanism. Smooth transition sections are established between commands: direction adjustments use trapezoidal acceleration / deceleration curves, limiting angular acceleration to no more than 30 degrees / second²; velocity changes follow a parabolic gradual change model, with the rate of acceleration change controlled within 2 meters / second³. All merged commands are appended with a timestamp sequence when injected into the execution system to ensure the execution order conforms to the expected timeline. Execution status is monitored in real time: the accuracy standard for direction adjustment completion is within ±0.5 degrees; the convergence tolerance for velocity adjustment is ±0.1 meters / second.

[0136] The sudden obstacle response mechanism features cross-module collaboration. When the LiDAR detects a new obstacle during reassessment, the model update module immediately pauses the current clustering analysis and prioritizes feature extraction for the new obstacle. The generation time for the description vector of the new obstacle is compressed to within 8 milliseconds, and it is directly inserted into the existing spatial distribution model. After obtaining the updated model, the reinforcement learning decision module recalculates the collision probability distribution only for the local state space of the affected area. This partial update mechanism keeps the overall response latency within 50 milliseconds.

[0137] The dynamic scene adaptation mechanism includes scene type recognition. The system analyzes the spatial distribution patterns of obstacles with sudden velocity changes: it activates an emergency avoidance mode for single-change obstacles; it initiates cooperative path planning for obstacles involving multiple obstacles (such as flocks of birds); and it establishes an oscillation model for predicting periodically moving obstacles. Recognition is based on five characteristic parameters: obstacle density gradient change, velocity vector correlation, motion trajectory periodicity, spatial distribution symmetry, and historical behavior matching degree. Pre-set parameter templates are activated for each scene type, and optimal path search constraints are automatically configured.

[0138] A self-healing recovery mechanism ensures system stability. When the position deviation persists for more than 1 meter after a supplementary instruction is executed, or when the environmental condition triggers a reassessment twice consecutively, the system automatically performs a safety reset: clearing all planned path nodes, climbing to a safe altitude (set to 20 meters above the current position), and initiating a 360-degree omnidirectional environmental scan. After updating the model with scan data, the path planning is reinitialized, and previously failed areas are marked as areas of special concern. Subsequent path searches in these areas employ an ultra-conservative strategy—the risk assessment weight is increased by an additional 50%.

[0139] Dynamic quota management is implemented for resource allocation. In normal mode, dynamic obstacle handling tasks consume 15% of computing resources; when a response is triggered, the resource quota increases to 50%, achieved by pausing non-critical background tasks (such as historical data archiving and redundant verification calculations). The processing task is divided into four concurrent sub-threads: trend monitoring and threshold judgment, path space reparameterization, instruction correction and generation, and execution status tracking. Thread priorities are dynamically adjusted at the millisecond level, with latency-sensitive tasks scheduled first.

[0140] The historical behavior learning system optimizes response parameters. After each response event, an event report is automatically generated, recording the triggering conditions, response strategy, and actual execution effect. The event database uses a multi-dimensional index structure, storing data in clusters based on obstacle movement patterns, environmental complexity, and response effect. Once sufficient samples are accumulated, the offline analysis module mines the correlation between parameter settings and obstacle avoidance success rates, automatically optimizing the judgment threshold and correction magnitude. New parameter configurations are incrementally updated to the online system after being validated through shadow testing.

[0141] Hardware acceleration modules enhance response speed. A programmable gate array (PGA) is configured with dedicated circuitry for motion trend prediction: differential computation units support parallel processing of six sets of velocity data; threshold determination logic achieves nanosecond-level response. Path search space generation is accelerated using a graphics processor: mesh generation calculations employ a parallel rasterization architecture; cost map construction utilizes 256 parallel processing cores. These hardware modules compress environmental dynamic response time to the millisecond level.

[0142] The normal handover process employs a segmented handover strategy. After the obstacle speed trend recovers to a stable range for three consecutive sampling periods, the system assesses the risk value of the current route: if the risk value is below 0.05, it directly switches back to the main path; in a medium-risk state (0.05-0.2), two to three transition paths are generated, and after testing execution stability, it returns to the main path; in a high-risk state, a supplementary mechanism is retained until the danger zone is cleared. Throughout the handover process, trajectory smoothness is monitored, and the heading angle offset is controlled within 0.5 degrees per meter of path.

[0143] The mechanism's performance is monitored in a closed-loop manner throughout its entire lifecycle. The state tracker maintains 40 operational metrics: from false alarm rate of motion trends to command merging delay time, from path replanning frequency to obstacle avoidance action completion rate. Metric data is stream-processed to form an operational baseline, and anomaly warnings are implemented based on statistical control chart principles. The maintenance diagnostic interface provides deep debugging capabilities, supporting the reproduction of system response sequences within any time window for fault analysis.

[0144] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0145] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An optimization method for autonomous obstacle avoidance of unmanned aerial vehicles (UAVs) based on lidar, characterized in that, include: The real-time point cloud data stream in the UAV flight environment is obtained by the lidar scanning module. The real-time point cloud data stream contains the three-dimensional position information and dynamic motion trajectory of obstacles. An obstacle spatial distribution model is constructed based on the real-time point cloud data stream. The obstacle spatial distribution model is used to characterize the geometry and motion state of the obstacles. A reinforcement learning algorithm is used to predict the collision risk value between the drone and the obstacle. The collision risk value is calculated based on the location information and motion trajectory in the spatial distribution model of the obstacle. Based on the collision risk value, an obstacle avoidance path optimization instruction is generated, which is used to adjust the flight direction and speed of the UAV. The system executes the obstacle avoidance path optimization command to control the drone's drive system and records flight trajectory feedback data in real time. The parameters of the obstacle spatial distribution model are updated based on the flight trajectory feedback data; The construction of the obstacle spatial distribution model includes: The point cloud 3D coordinate sequence is divided into multiple spatial grid units; Perform point cloud clustering analysis within each spatial grid cell to identify obstacle boundary features; Generate obstacle geometry description vectors and motion velocity estimates based on clustering results; The outputs of all spatial grid cells are integrated to form the complete structure of the obstacle spatial distribution model; The obstacle avoidance path optimization instructions include: Analyze the probability distribution of the collision risk values ​​to identify high-risk obstacle areas; A dynamic path planning algorithm is used to calculate candidate schemes for safe flight paths, and the candidate schemes contain multiple path nodes; Evaluate the feasibility score of each path node and select the path scheme with the highest feasibility score as the obstacle avoidance path optimization instruction; The obstacle avoidance path optimization command includes specific directional angle adjustment values ​​and velocity changes; The parameters for updating the obstacle spatial distribution model based on the flight trajectory feedback data include: Compare the deviation values ​​between the real-time position data of the thrusters and the predicted path nodes; If the deviation value exceeds the preset tolerance threshold, the division rule of the spatial grid unit is adjusted; Re-execute the point cloud clustering analysis and update the obstacle geometry description vector; The updated obstacle spatial distribution model is used for the next round of collision risk value prediction.

2. The optimization method for autonomous obstacle avoidance of UAVs based on lidar according to claim 1, characterized in that, The process of acquiring real-time point cloud data streams in the UAV flight environment via a lidar scanning module includes: Configure the scanning frequency and resolution parameters of the lidar sensor to ensure that the point cloud acquisition covers the preset monitoring area; Noise filtering is performed on the raw point cloud data to remove environmental interference signals; The filtered point cloud data is converted into a standardized three-dimensional coordinate sequence.

3. The optimization method for autonomous obstacle avoidance of UAVs based on lidar according to claim 1, characterized in that, The method of using reinforcement learning algorithms to predict the collision risk value between the drone and obstacles includes: The obstacle geometry description vector and the estimated motion velocity are input into the state space of the reinforcement learning algorithm; Design a reward function mechanism, which is calculated based on the relative distance between the current position of the UAV and the obstacle; The reinforcement learning policy network is updated through iterative training, and the probability distribution of the collision risk value is output.

4. The optimization method for autonomous obstacle avoidance of UAVs based on lidar according to claim 1, characterized in that, The drive system for controlling the UAV includes: The direction angle adjustment value and speed change are converted into motor control signals; The motor control signal is sent to the UAV thruster via the drive interface module; Simultaneously acquire thruster response data and real-time position data; The thruster response data and real-time position data constitute a part of the flight trajectory feedback data.

5. The optimization method for autonomous obstacle avoidance of unmanned aerial vehicles based on lidar according to claim 1, characterized in that, The method also includes performing data preprocessing before constructing the obstacle spatial distribution model: Extract timestamp information from the three-dimensional coordinate sequence to align the time dimension of the point cloud data; Compensate for the scanning delay error of the LiDAR and generate a time-synchronized point cloud dataset.

6. The optimization method for autonomous obstacle avoidance of unmanned aerial vehicles based on lidar according to claim 5, characterized in that, The method also includes optimizing the training process of the reinforcement learning algorithm, specifically: Initialize the weight parameters of the policy network; Simulate environmental interactions using historical flight datasets and calculate policy gradient update values; The policy gradient update value is used to adjust the weights of the reward function mechanism.

7. The optimization method for autonomous obstacle avoidance of unmanned aerial vehicles based on lidar according to claim 6, characterized in that, The method also includes an adaptive mechanism for handling dynamic obstacles, specifically: The movement speed estimate is monitored for change trend. If the change trend exceeds the stable range, the path node is re-evaluated and a supplementary obstacle avoidance instruction is generated. The supplementary obstacle avoidance instruction is then merged into the obstacle avoidance path optimization instruction.

Images