Intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method and system

Abstract

The invention discloses an intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method and system. The method comprises the steps of multi-modal data acquisition, space-time alignment fusion, three-dimensional dynamic trajectory reconstruction, obstacle avoidance path planning and system verification. The invention belongs to the technical field of path planning, and particularly relates to an intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method and system.According to the scheme, an incremental dynamic updating mechanism is adopted, and the three-dimensional trajectory reconstruction precision is improved through a double-correction mechanism of hidden space normal estimation and global discretization constraint; bEV feature optimization fusion is adopted to generate a double-enhanced feature map; constructing a mixed bounding box for the trajectory prediction points, calculating a dynamic minimum distance, designing a dynamic safety threshold fusing the real-time speed of the unmanned aerial vehicle and the priority, constructing a space-time coupling risk index in combination with the potential collision time and the priority to complete refined risk grading, and constructing a reinforcement learning model of priority weighted rewards; the problem that the passing right distribution is not uniform in the low-altitude multi-priority unmanned aerial vehicle obstacle avoidance process is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of path planning, and particularly relates to an intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method and system. BACKGROUND

[0002] The low-altitude unmanned aerial vehicle obstacle avoidance path planning method is a method for planning a safe, efficient and collision-free flight path from a starting point to an ending point for an unmanned aerial vehicle in a low-altitude complex environment, avoiding various static and dynamic obstacles under the premise of meeting the unmanned aerial vehicle's own constraints such as power, endurance and flight attitude, and taking into account the path optimization.

[0003] However, the existing low-altitude unmanned aerial vehicle obstacle avoidance path planning method has the technical problems of low multi-source data fusion efficiency, poor three-dimensional trajectory reconstruction accuracy, inability to accurately reflect the change in the height of the unmanned aerial vehicle, resulting in distorted collision risk assessment in complex environments such as high-rise building dense areas and mountainous areas, insufficient real-time conflict warning, and the existing method only triggers warning through a distance threshold without considering dynamic parameters such as the speed, heading and acceleration of the unmanned aerial vehicle, resulting in a high false alarm rate, uneven distribution of low-altitude multi-level unmanned aerial vehicle obstacle avoidance traffic rights, and low reliability of obstacle avoidance decisions in multi-target and multi-constrained low-altitude obstacle avoidance scenarios. SUMMARY

[0004] In view of the above problems, the present application provides an intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method and system to overcome the defects of the prior art, and to solve the technical problems of low multi-source data fusion efficiency, poor three-dimensional trajectory reconstruction accuracy, and inability to accurately reflect the change in the height of the unmanned aerial vehicle, resulting in distorted collision risk assessment in complex environments such as high-rise building dense areas and mountainous areas, by using an incremental dynamic updating mechanism to improve the multi-source data fusion efficiency, using a dual correction mechanism of hidden space normal estimation and global discretization constraint to improve the three-dimensional trajectory reconstruction accuracy, and using BEV feature optimization fusion to generate a double-enhanced feature map to adapt to the trajectory reconstruction in complex environments and solve the problem of distorted collision risk assessment.

[0005] The technical solution adopted by this invention is as follows: This invention provides an intelligent obstacle avoidance path planning method for low-altitude unmanned aerial vehicles (UAVs), which includes the following steps:

[0006] Step S1: Multimodal data acquisition;

[0007] Step S2: Spatiotemporal alignment and fusion;

[0008] Step S3: 3D dynamic trajectory reconstruction;

[0009] Step S4: Obstacle avoidance path planning;

[0010] Step S5: System verification.

[0011] Further, in step S1, the multimodal data acquisition includes the following steps:

[0012] Step S11: Equipment deployment. Deploy sensing devices in the monitoring area according to the preset topology. All devices achieve nanosecond-level synchronization through a dedicated network.

[0013] Step S12: Collect multimodal data, including radar data, visual data, ADS-B data and RID data, and preprocess the data respectively to obtain preprocessed multimodal data.

[0014] Further, in step S2, the spatiotemporal alignment fusion specifically involves performing spatiotemporal alignment fusion on the preprocessed multimodal data to obtain standard multimodal data, which includes radar point cloud data, visual features, and RID location data.

[0015] Further, in step S3, the three-dimensional dynamic trajectory reconstruction involves inputting the standard multimodal data, after spatiotemporal alignment and fusion, into an improved NeRF network to obtain a high-precision three-dimensional dynamic trajectory, including the following steps:

[0016] Step S31: Initial digital twin scene construction. Input standard multimodal data to initialize the scene. Generate a fused feature map through feature stitching. Using GPS-calibrated ground control points as a reference, remove scene modeling anomalies through the RANSAC algorithm to ensure that the absolute position error of the initial digital twin scene is ≤3cm.

[0017] Step S32: Incremental dynamic update. In the initial digital twin scenario, to address the poor real-time performance caused by traditional NeRF full-scenario retraining, a strategy of trajectory change region determination, block parameter storage, and momentum update is adopted. Only the neural radiation field parameters in the trajectory change region are locally optimized, and the static region parameters are frozen to achieve efficient dynamic iteration of the digital twin scenario.

[0018] Step S33: Point cloud attitude correction. Input the dynamically updated digital twin scene and NeRF parameters. During the iteration process, the cumulative error of pose estimation of consecutive frames will be generated. A dual correction mechanism of latent space normal estimation and global discretization constraint is adopted to control the attitude error of adjacent frames of the trajectory to be within 0.1°, and the corrected point cloud attitude matrix is ​​obtained.

[0019] Step S34: BEV feature optimization and fusion, using dual-channel attention;

[0020] BEV feature map generation, the BEV feature map includes BEV geometric feature map and BEV semantic feature map, the BEV geometric feature map is generated by the radar point cloud calibrated by the corrected point cloud attitude matrix through cylinder projection, and the BEV semantic feature map is generated by the visual semantic segmentation mask through bird's-eye view conversion.

[0021] Dual-head attention fusion uses attention weights to weight and fuse BEV geometric features and BEV semantic features to generate a BEV fusion feature map with both spatial and semantic enhancements.

[0022] Step S35: Generate a high-precision 3D dynamic trajectory. Input the incrementally updated radiation field parameters, the corrected point cloud attitude matrix, and the optimized BEV fusion features into the improved NeRF rendering module, and generate a 3D dynamic trajectory through dynamic sampling and radiation field rendering.

[0023] Further, in step S4, the obstacle avoidance path planning includes the following steps:

[0024] Step S41: Trajectory Prediction: Input the current 3D dynamic trajectory data into the LSTM-Transformer hybrid model to predict 100 trajectory prediction points within the next 5 seconds;

[0025] Step S42: Bounding box generation. Construct an AABB-OBB hybrid bounding box for each trajectory prediction point and calculate the minimum distance between the UAV bounding boxes.

[0026] Step S43: Conflict early warning, specifically, constructing a spatiotemporal coupled risk index to achieve quantitative classification of risk levels;

[0027] Step S44: Obstacle avoidance path planning. For multi-priority drone obstacle avoidance needs, a priority-weighted reward reinforcement learning model is constructed to prioritize the passage of high-priority drones. This includes the following steps:

[0028] Step S441: Dynamic state space construction, integrating self-state, environment state and priority information to construct a high-dimensional dynamic state vector;

[0029] Step S442: Priority-weighted reward function, design a composite reward mechanism of basic reward, priority reward and constraint penalty, and give priority to guaranteeing the right of way of high priority drones;

[0030] Step S443: Adopt the improved PPO algorithm, introduce the priority-weighted advantage function to ensure that the policy update is stable and biased towards the high-priority optimization objective, and output the optimized policy network parameters;

[0031] The real-time state of the current airspace is input into the optimized policy network. Based on the optimized policy network parameters, the policy network outputs the optimal action for that state. The optimal action is then mapped to the three-dimensional coordinate system of the airspace, and the corresponding discrete trajectory point coordinates are generated by combining the time step. ; Traverse the time steps to obtain the complete trajectory sequence ;

[0032] Step S444: Fit the complete trajectory sequence output by the strategy into a smooth curve to generate an obstacle avoidance path.

[0033] Furthermore, in step S5, the system verification specifically involves verifying whether the generated obstacle avoidance path is entirely within the legal airspace and without any risk of crossing the boundary. If there are issues such as path boundary crossing or high-priority UAV trajectory conflicts, the system feeds back to the reinforcement learning model. Based on the verification error, the reward function coefficients and strategy parameters are adjusted, and the path is regenerated. If the verification passes, the path and the corresponding warning level information are encapsulated and sent to the UAV flight control system and the airspace management center to ensure the executability of the obstacle avoidance path and airspace safety.

[0034] This invention provides an intelligent low-altitude unmanned aerial vehicle (UAV) obstacle avoidance path planning system, comprising a multimodal data acquisition module, a spatiotemporal alignment and fusion module, a three-dimensional dynamic trajectory reconstruction module, an obstacle avoidance path planning module, and a system verification module;

[0035] The multimodal data acquisition module specifically acquires multimodal data and preprocesses it to obtain preprocessed multimodal data; then sends the data to the spatiotemporal alignment and fusion module.

[0036] The spatiotemporal alignment and fusion module specifically performs spatiotemporal alignment and fusion on the preprocessed multimodal data to obtain standard multimodal data; and sends the data to the three-dimensional dynamic trajectory reconstruction module.

[0037] The 3D dynamic trajectory reconstruction module specifically involves inputting standard multimodal data, constructing an initial digital twin scene, performing incremental dynamic updates, using a dual correction mechanism of latent space normal estimation and global discretization constraints to achieve point cloud attitude correction, using dual-head channel attention to optimize and fuse BEV features, and finally generating a 3D dynamic trajectory; the data is then sent to the obstacle avoidance path planning module.

[0038] The obstacle avoidance path planning module specifically constructs a hybrid bounding box based on trajectory prediction points, calculates the minimum distance between UAV bounding boxes, constructs a spatiotemporal coupling risk index to classify conflict levels and trigger corresponding warnings, constructs a priority-weighted reward reinforcement learning model, and uses an improved PPO algorithm to generate a smooth obstacle avoidance path by fitting discrete trajectory points based on policy output with B-splines; the data is then sent to the system verification module.

[0039] The system verification module specifically verifies the legality and conflict-free nature of the obstacle avoidance path. If it fails, it provides feedback to adjust the model and regenerate it. If it passes, it encapsulates the warning information and sends it to the control center.

[0040] The beneficial results achieved by the present invention using the above solution are as follows:

[0041] (1) To address the technical problems of low efficiency of multi-source data fusion, poor accuracy of 3D trajectory reconstruction, inability to accurately reflect changes in UAV altitude, and distortion of collision risk assessment in complex environments such as densely populated high-rise buildings and mountainous areas, an incremental dynamic update mechanism is adopted to improve the efficiency of multi-source data fusion. The accuracy of 3D trajectory reconstruction is improved through a dual correction mechanism of latent space normal estimation and global discretization constraint. BEV feature optimization fusion is adopted to generate dual-enhanced feature maps to adapt to trajectory reconstruction in complex environments and solve the problem of distortion in collision risk assessment.

[0042] (2) To address the technical problems of insufficient real-time conflict warning, the existing methods only trigger warnings through distance thresholds without considering dynamic parameters such as speed, heading, and acceleration of UAVs, resulting in high false alarm rates, uneven distribution of right-of-way for low-altitude multi-level UAVs, and low reliability of obstacle avoidance decisions in low-altitude obstacle avoidance scenarios with multiple targets and constraints, trajectory prediction is adopted to construct a hybrid bounding box for prediction points, calculate the dynamic minimum distance, and design a dynamic safety threshold that integrates the real-time speed and priority of UAVs. A spatiotemporal coupling risk index is constructed in combination with potential collision time and priority to complete the refined risk classification, thereby solving the problem of high false alarm rates caused by relying solely on distance thresholds to trigger warnings. A priority-weighted reward reinforcement learning model is constructed to solve the problem of uneven distribution of right-of-way in low-altitude multi-priority UAV obstacle avoidance. An improved PPO algorithm is adopted to avoid sudden changes in low-altitude obstacle avoidance strategies and ensure the stability of obstacle avoidance decisions. Attached Figure Description

[0043] Figure 1 A flowchart illustrating an intelligent low-altitude unmanned aerial vehicle (UAV) obstacle avoidance path planning method provided by the present invention;

[0044] Figure 2 This is a schematic diagram of an intelligent low-altitude unmanned aerial vehicle (UAV) obstacle avoidance path planning system provided by the present invention.

[0045] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0047] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0048] Example 1, see Figure 1 The present invention provides an intelligent obstacle avoidance path planning method for low-altitude unmanned aerial vehicles (UAVs), which includes the following steps:

[0049] Step S1: Multimodal data acquisition, specifically, acquiring multimodal data and preprocessing it to obtain preprocessed multimodal data;

[0050] Step S2: Spatiotemporal alignment and fusion, specifically, performing spatiotemporal alignment and fusion on the preprocessed multimodal data to obtain standard multimodal data;

[0051] Step S3: 3D dynamic trajectory reconstruction, specifically, inputting standard multimodal data into the improved NeRF network, performing incremental dynamic updates for initial digital twin scene construction and local parameter optimization, using a dual correction mechanism of latent space normal estimation and global discretization constraints to achieve point cloud attitude correction, and using dual-head channel attention to optimize and fuse BEV features, finally generating a 3D dynamic trajectory;

[0052] Step S4: Obstacle avoidance path planning, specifically predicting future trajectory points, constructing a hybrid bounding box to calculate the minimum distance between drones, constructing a spatiotemporal coupling risk index to classify conflict levels and trigger corresponding warnings, constructing a dynamic state space that integrates multiple information, designing a composite reward mechanism, improving the PPO algorithm to construct a reinforcement learning model, and using B-spline fitting to generate a smooth obstacle avoidance path that prioritizes the passage of high-priority drones.

[0053] Step S5: System verification, specifically verifying the legality and conflict-free nature of the obstacle avoidance path. If it fails, feedback is provided to adjust the model and regenerate it. If it passes, the warning information is packaged and sent to the control center.

[0054] Example 2, see Figure 1 This embodiment is based on the above embodiment. In step S1, the multimodal data acquisition includes the following steps:

[0055] Step S11: Equipment Deployment. Deploy sensing equipment in the monitoring area according to the preset topology. Specifically, this includes 4 sets of 4D millimeter-wave radar (coverage radius of 500 meters), 1 ADS-B receiver (coverage radius of more than 10 kilometers), 1 optoelectronic device (4K resolution, 30fps frame rate), 4 RID receivers (coverage radius of 2 kilometers), and 1 edge computing node (NVIDIA Jetson AGXOrin). All devices are synchronized at the nanosecond level through a dedicated network.

[0056] Step S12: Collect multimodal data, including radar data, visual data, ADS-B data, and RID data, and perform data preprocessing on each; obtain preprocessed multimodal data;

[0057] Radar data preprocessing involves using statistical filtering algorithms to remove outliers and extracting initial features through the STEM module.

[0058] Visual data preprocessing involves using a ResNet-152 network for multi-scale feature extraction to generate a semantic segmentation mask, thus obtaining visual feature data.

[0059] ADS-B data preprocessing involves parsing fields such as ICAO address, latitude and longitude, altitude, and speed, and verifying data integrity using the SHA-256 algorithm.

[0060] RID data preprocessing involves parsing fields such as UAV ID, location, speed, altitude, and timestamp, calculating RID signal strength (RSSI value), and assessing data reliability.

[0061] Example 3, see Figure 1 This embodiment is based on the above embodiment. In step S2, the spatiotemporal alignment fusion specifically involves performing spatiotemporal alignment fusion on the preprocessed multimodal data to obtain standard multimodal data. The standard multimodal data includes radar point cloud data, visual features, and RID location data. The spatiotemporal alignment fusion includes the following:

[0062] Time synchronization is achieved by using GPS timing signals as a reference and calibrating the clocks of each device via the PTP protocol to ensure a time error of ≤50ns and eliminate timing deviations.

[0063] Spatial alignment is achieved by projecting the radar coordinate system point cloud onto the photoelectric coordinate system using orthogonal feature transformation, and point cloud registration is realized through ICP algorithm. The registration error is ≤2cm, and the target spatial positioning error is controlled at the centimeter level, meeting the requirements of low-altitude high-precision airspace perception.

[0064] RID data fusion combines RID location data with radar / visual trajectories using Kalman filtering to generate trajectory data with a unified spatiotemporal reference, integrating identity information with spatial trajectories to achieve unified identity-trajectory association.

[0065] Example 4, see Figure 1 This embodiment is based on the above embodiment. In step S3, the three-dimensional dynamic trajectory reconstruction involves inputting the standard multimodal data after spatiotemporal alignment and fusion into the improved NeRF network to obtain a high-precision three-dimensional dynamic trajectory, including the following steps:

[0066] Step S31: Initial digital twin scene construction. Input standard multimodal data to initialize the scene. Generate a fusion feature map through feature stitching. Using GPS-calibrated ground control points as a reference, remove scene modeling anomalies through the RANSAC algorithm to ensure that the absolute position error of the initial digital twin scene is ≤3cm. Build a high-precision static digital twin spatial reference to provide a reliable spatial reference for subsequent trajectory positioning.

[0067] Step S32: Incremental dynamic update. In the initial digital twin scenario, to address the poor real-time performance issue caused by traditional NeRF full-scenario retraining, a strategy of trajectory change region determination, block parameter storage, and momentum update is adopted. Only the neural radiation field parameters in the trajectory change region are locally optimized, and the static region parameters are frozen to achieve efficient dynamic iteration of the digital twin scenario. This includes the following:

[0068] Trajectory change region determination: Based on the trajectory prediction results output by Kalman filter, the trajectory overlap ratio (IOU) between the current frame and the previous frame is calculated. When the IOU < 0.7, it is determined to be a region with significant trajectory change. Combined with the motion vector of the radar point cloud, the spatial range of the dynamic target is located.

[0069] Block parameter storage: The block storage mechanism of fused feature map is adopted to divide the neural radiation field parameters according to the scene grid, activate only the parameter blocks corresponding to the trajectory change area, and freeze the static area parameters.

[0070] A momentum update strategy is introduced to fuse historical parameter information in the updated change region, avoiding scene feature drift caused by incremental updates. An adaptive learning rate is used to improve the convergence speed. The formula used is as follows:

[0071] ;

[0072] In the formula, This represents the dynamic parameters of the neural radiation field in the region where the frame t1+1 changes. This represents the dynamic parameters of the neural radiation field in the region where frame t1 changes. This represents the momentum coefficient, with a value of 0.9. This represents the parameter update amount in frame t1-1, i.e., the gradient direction of the historical updates. This represents the adaptive learning rate for frame t1. This indicates that the loss function has dynamic parameters. gradient, This represents the rendering loss of the neural radiation field. This represents the color / density value predicted by the model. Represents the actual observed color / density values. This represents a mask for the trajectory change area. The elements of the change area are 1, which participates in the parameter update. The elements of the static area are 0, which freezes the parameters and enables local incremental updates.

[0073] Step S33: Point cloud pose correction. Input the dynamically updated digital twin scene and NeRF parameters. During the iteration process, cumulative errors in pose estimation of consecutive frames will be generated. A dual correction mechanism of latent space normal estimation and global discretization constraint is adopted to control the pose error of adjacent frames of the trajectory to within 0.1°, and the corrected point cloud pose matrix is ​​obtained; including the following:

[0074] Latent space normal feature extraction involves inputting radar point cloud and visual features into an improved MLP module, mapping them to a 512-dimensional latent space, and calculating the latent space normal vector using the gradient properties of the latent space features; the formula used is as follows:

[0075] ;

[0076] ;

[0077] In the formula, Let denot p be the latent space feature vector of point p in frame t1, and let GELU() denote the Gaussian error linear unit activation function. , This represents the weight matrix of the improved MLP. The dimension is 64×128 (from input layer to hidden layer). The dimension is 128×512 (from hidden layer to hidden space output layer). , This represents the bias vector of the improved MLP, with dimensions respectively... , Output dimension matching, This represents the radar point cloud features of point p. The visual semantic features representing point p. This indicates a feature concatenation operation. This represents the latent space normal vector of point p in frame t1. This represents the gradient operator with respect to spatial coordinate p. This represents the hidden feature decoder. Represents the L2 norm;

[0078] By combining point coordinate matching and normal vector consistency to calculate the pose estimation dual-constraint formula, a pose solution optimization objective function is constructed. The pose matrix of adjacent frames is solved using the EPnP algorithm, and the formula used is as follows:

[0079] ;

[0080] ;

[0081] In the formula, , This represents the three-dimensional coordinates of point p in frame t1 and frame t1-1. This represents a 3×3 rotation matrix used to indicate the attitude rotation relationship between adjacent frames. This represents a 3×1 translation matrix, used to indicate the translation relationship between adjacent frames. , Let these represent the latent space normal vectors of point p in frame t1 and frame t1-1, respectively. Indicates coordinate observation noise. Indicates the observation noise of the normal vector. This represents the pose estimation loss function. , These represent the error weighting coefficients for coordinate matching and normal vector consistency, respectively. Represents a point cloud matching set;

[0082] Using the GPS reference point of the initial digital twin scene as the anchor point, the pose matrices of consecutive frames are combined to form a trajectory sequence. The trajectory sequence is divided into K discrete windows (K=10, 8 frames per window) according to time. The cumulative error of the pose matrix relative to the GPS anchor point in each window is calculated, and global discretization pose correction is performed to obtain the corrected point cloud pose matrix. The formula used is as follows:

[0083] ;

[0084] ;

[0085] In the formula, This represents the corrected point cloud pose matrix, which includes the rotation and translation matrices. This represents the original pose matrix obtained by the EPNP algorithm. This indicates the correction step size coefficient. This represents the gradient operator with respect to the cumulative error. Represents the three-dimensional coordinates of the i-th GPS reference anchor point. This represents the original observed coordinates of the i-th point in frame t1. This represents the temporal smoothing coefficient, with a value of 0.1, used to constrain the consistency of adjacent poses. The square of the Frobenius norm of the matrix is ​​used to quantify the overall difference in the pose matrix. This represents the average cumulative error of the k-th discrete time window. This represents the k-th discrete-time window. This represents the number of frames within the k-th window, with a value of 8.

[0086] Step S34: BEV feature optimization and fusion, using dual-channel attention;

[0087] BEV feature map generation, the BEV feature map includes BEV geometric feature map and BEV semantic feature map, the BEV geometric feature map is generated by the radar point cloud calibrated by the corrected point cloud attitude matrix through cylinder projection, and the BEV semantic feature map is generated by the visual semantic segmentation mask through bird's-eye view conversion.

[0088] Dual-head attention fusion is used to weight and fuse BEV geometric and semantic features through attention weights, generating a BEV fusion feature map with both spatial and semantic enhancements. The formula used is as follows:

[0089] ;

[0090] In the formula, This represents the final fusion weight matrix, enabling adaptive synergy of multimodal features. It is the Sigmoid activation function. , These represent the spatial attention weight matrix and the semantic attention weight matrix, respectively, and GAP represents the global average pooling operation. Represents the geometric feature map of BEV. This represents the semantic feature map of BEV, where b represents the bias term;

[0091] Step S35: Generate a high-precision 3D dynamic trajectory. Input the incrementally updated radiation field parameters, the corrected point cloud attitude matrix, and the optimized BEV fusion features into the improved NeRF rendering module, and generate a 3D dynamic trajectory through dynamic sampling and radiation field rendering.

[0092] The dynamic sampling is specifically based on the target region in the BEV fusion feature, and adopts an adaptive light sampling strategy to increase the sampling density in the area around the trajectory and decrease the sampling density in the background area, thereby balancing accuracy and efficiency.

[0093] The radiation field rendering specifically involves calculating the density and color value of each sampled ray through the neural radiation field, combining it with pose correction information to render a 3D scene containing the target trajectory, and extracting individual target trajectory point clouds through semantic mask segmentation.

[0094] The extracted target trajectory point cloud is smoothed by Gaussian filtering to remove rendering noise points, and finally outputs a three-dimensional dynamic trajectory. The three-dimensional dynamic trajectory includes the three-dimensional coordinates (x, y, z), attitude angles (roll, pitch, yaw), timestamps and target semantic tags for each frame, and the trajectory position accuracy is ≤5mm and the attitude accuracy is ≤0.1°.

[0095] By performing the above operations, an incremental dynamic update mechanism is adopted to improve the efficiency of multi-source data fusion. A dual correction mechanism of latent space normal estimation and global discretization constraint is used to improve the accuracy of 3D trajectory reconstruction. BEV feature optimization fusion is used to generate dual-enhanced feature maps to adapt to trajectory reconstruction in complex environments, thus solving the problem of collision risk assessment distortion. This solves the technical problems of low efficiency of multi-source data fusion, poor accuracy of 3D trajectory reconstruction, inability to accurately reflect changes in UAV altitude, and resulting in collision risk assessment distortion in complex environments such as densely populated high-rise buildings and mountainous areas.

[0096] Example 5, see Figure 1 This embodiment is based on the above embodiment. In step S4, the obstacle avoidance path planning includes the following steps:

[0097] Step S41: Trajectory Prediction: Input the current 3D dynamic trajectory data into the LSTM-Transformer hybrid model to predict 100 trajectory prediction points within the next 5 seconds;

[0098] Step S42: Bounding box generation. Construct an AABB-OBB hybrid bounding box for each trajectory prediction point and calculate the minimum distance between the UAV bounding boxes; the formula used is as follows:

[0099] ;

[0100] In the formula, This represents the minimum distance between the AABB-OBB hybrid bounding boxes of drone i and drone j. Let represent the minimum Euclidean distance between the axis-aligned bounding boxes (AABB) of drones i and j. This represents the minimum Euclidean distance between the orientation bounding boxes (OBB) of drones i and j. This represents the dynamic safety threshold of drone i. This represents the priority weight coefficient. This represents the RID priority coefficient of drone i. This indicates the real-time flight speed of drone i. Indicates the maximum permissible flight speed in the airspace. Indicates the basic safety distance threshold;

[0101] Step S43: Conflict early warning, specifically, constructing a spatiotemporal coupling risk index to quantitatively classify risk levels, using the following formula:

[0102] ;

[0103] In the formula, This indicates that drone i and drone j will be in the future. The total risk index of spatiotemporal coupling at any given moment. Indicates the future The minimum distance between the hybrid bounding boxes of drone i and drone j at any given time. Indicates the predicted time of potential conflict. Indicates the lead time for safety warnings. This represents the RID priority coefficient of drone j. This represents the dynamic safety threshold of drone j;

[0104] Risk level classification, when At that time, it was considered low risk, with no early warning, and continuous monitoring was required.

[0105] when At that time, the risk level was classified as medium, triggering a general warning and pushing the risk information to the drone flight control system.

[0106] when When the situation is considered high-risk, an emergency warning is triggered, and the obstacle avoidance path planning preparatory process is initiated.

[0107] when When the situation is considered high-risk, a mandatory warning is triggered, the obstacle avoidance path is directly invoked and sent to low-priority drones, while the original flight path of high-priority drones is locked to ensure the execution of high-priority tasks.

[0108] Step S44: Obstacle avoidance path planning. For multi-priority drone obstacle avoidance needs, a priority-weighted reward reinforcement learning model is constructed to prioritize the passage of high-priority drones. This includes the following steps:

[0109] Step S441: Dynamic state space construction. By fusing the self-state, environment state, and priority information, a high-dimensional dynamic state vector is constructed. The formula used is as follows:

[0110] ;

[0111] In the formula, This represents the standardized dynamic state vector at time t. This represents the three-dimensional position coordinates of the UAV at time t. This represents the three-dimensional coordinates of the target point of the UAV. This represents the three-dimensional velocity vector of the UAV at time t. This represents the three-dimensional acceleration vector of the UAV at time t. This represents the maximum acceleration of the drone, and Norm() represents the L2 normalization function. This represents the distance vector between the UAV and the obstacle at time t. This represents the real-time priority coefficient of the UAV at time t. This represents the remaining time for the drone to reach the target point at time t. Indicates the maximum allowed time for the drone to reach the target point;

[0112] Step S442: Priority-weighted reward function. Design a composite reward mechanism consisting of basic reward, priority reward, and constraint penalty to prioritize the right-of-way for high-priority drones. The formula used is as follows:

[0113] ;

[0114] ;

[0115] In the formula, This represents the instantaneous composite reward value at time t. This represents the goal-oriented reward at time t. This represents the priority reward at time t. Indicates the collision penalty item. This represents the trajectory smoothing penalty term. Indicates the priority reward coefficient. This indicates the minimum distance between a high-priority drone and surrounding low-priority drones. This represents the safe distance threshold, and exp() represents the exponential function. This indicates the minimum distance between a low-priority drone and a nearby high-priority drone. This indicates an indicator function that takes the value 1 when the condition is met and 0 otherwise. The low-priority drone receives an avoidance bonus only when the distance between it and the high-priority drone is greater than the safety threshold.

[0116] Step S443: Adopt the improved PPO algorithm, introduce a priority-weighted advantage function to ensure stable policy updates that favor high-priority optimization objectives, and output the optimized policy network parameters. The formula used is as follows:

[0117] ;

[0118] ;

[0119] In the formula, This represents the priority-weighted advantage function at time t. This represents the discount factor, with a value of 0.98, used to reduce the weight of future rewards. This represents the total time step of a single trajectory. This represents the instantaneous composite reward value at time t+k. This represents the state value at time T, the endpoint of the trajectory. This represents the state value at time t. This means finding the optimal policy parameters that maximizes the minimum expected return. This indicates that the current policy network parameters are In state Select action The probability, This indicates the network parameters of the old policy in state. Select action The probability is used to limit the policy update magnitude. `clip()` represents the clipping function, which limits the policy probability ratio to a certain value. Within the interval, Indicates the cutting factor;

[0120] The real-time state of the current airspace is input into the optimized policy network. Based on the optimized policy network parameters, the policy network outputs the optimal action for that state. The optimal action is then mapped to the three-dimensional coordinate system of the airspace, and the corresponding discrete trajectory point coordinates are generated by combining the time step. By iterating through the time steps, a complete trajectory sequence consisting of multiple discrete trajectory points is obtained. ;

[0121] Step S444: Fit the complete trajectory sequence output by the strategy into a smooth curve to generate an obstacle avoidance path. The formula used is as follows:

[0122] ;

[0123] In the formula, This represents the final generated continuous obstacle avoidance path, which is a smooth curve fitted from discrete trajectory points. This represents the B-spline curve fitting function, which transforms discrete points into continuous paths that satisfy the dynamic constraints of the UAV.

[0124] By performing the above operations, trajectory prediction is used to construct a hybrid bounding box for the predicted points, the dynamic minimum distance is calculated, and a dynamic safety threshold that integrates the real-time speed and priority of the UAV is designed. A spatiotemporal coupling risk index is constructed by combining potential collision time and priority to complete the refined risk classification, which solves the problem of high false alarm rate caused by relying solely on distance threshold to trigger warnings. A priority-weighted reward reinforcement learning model is constructed to solve the problem of uneven right-of-way allocation in obstacle avoidance of low-altitude multi-priority UAVs. An improved PPO algorithm is adopted to avoid abrupt changes in low-altitude obstacle avoidance strategy and ensure the stability of obstacle avoidance decision.

[0125] Example 6, see Figure 1 This embodiment is based on the above embodiment. In step S5, the system verification specifically verifies whether the generated obstacle avoidance path is completely within the legal airspace and has no risk of crossing the boundary. If there are problems such as path crossing the boundary or high-priority UAV trajectory conflict, the feedback is sent to the reinforcement learning model. Based on the verification error, the reward function coefficient and strategy parameters are adjusted, and the path is regenerated. If the verification is successful, the path and the corresponding warning level information are encapsulated and sent to the UAV flight control system and airspace control center to ensure the executability of the obstacle avoidance path and airspace safety.

[0126] Example 7, see Figure 2 Based on the above embodiments, this embodiment provides an intelligent low-altitude UAV obstacle avoidance path planning system, including a multimodal data acquisition module, a spatiotemporal alignment and fusion module, a three-dimensional dynamic trajectory reconstruction module, an obstacle avoidance path planning module, and a system verification module.

[0127] The multimodal data acquisition module specifically acquires multimodal data and preprocesses it to obtain preprocessed multimodal data; then sends the data to the spatiotemporal alignment and fusion module.

[0128] The spatiotemporal alignment and fusion module specifically performs spatiotemporal alignment and fusion on the preprocessed multimodal data to obtain standard multimodal data; and sends the data to the three-dimensional dynamic trajectory reconstruction module.

[0129] The 3D dynamic trajectory reconstruction module specifically involves inputting standard multimodal data, constructing an initial digital twin scene, performing incremental dynamic updates, using a dual correction mechanism of latent space normal estimation and global discretization constraints to achieve point cloud attitude correction, using dual-head channel attention to optimize and fuse BEV features, and finally generating a 3D dynamic trajectory; the data is then sent to the obstacle avoidance path planning module.

[0130] The obstacle avoidance path planning module specifically constructs a hybrid bounding box based on trajectory prediction points, calculates the minimum distance between UAV bounding boxes, constructs a spatiotemporal coupling risk index to classify conflict levels and trigger corresponding warnings, constructs a priority-weighted reward reinforcement learning model, and uses an improved PPO algorithm to generate a smooth obstacle avoidance path by fitting discrete trajectory points based on policy output with B-splines; the data is then sent to the system verification module.

[0131] The system verification module specifically verifies the legality and conflict-free nature of the obstacle avoidance path. If it fails, it provides feedback to adjust the model and regenerate it. If it passes, it encapsulates the warning information and sends it to the control center.

[0132] 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.

[0133] 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.

[0134] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A method for intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning, characterized in that: The method comprises the following steps: Step S1: multi-modal data acquisition, specifically, acquiring multi-modal data and preprocessing to obtain preprocessed multi-modal data; Step S2: spatio-temporal alignment and fusion, specifically, spatio-temporal alignment and fusion are performed on the preprocessed multi-modal data to obtain standard multi-modal data; Step S3: three-dimensional dynamic trajectory reconstruction, specifically, inputting the standard multi-modal data, performing initial digital twin scene construction, performing incremental dynamic updating, adopting a double correction mechanism of implicit space normal estimation and global discretization constraint to realize point cloud pose correction, adopting double-head channel attention for BEV feature optimization and fusion, and finally generating a three-dimensional dynamic trajectory; Step S4: obstacle avoidance path planning, specifically, constructing a hybrid bounding box based on trajectory prediction points, calculating the minimum distance between the bounding boxes of the unmanned aerial vehicles, constructing a time-space coupled risk index to divide the conflict level and trigger the corresponding warning, constructing a reinforcement learning model with priority weighted rewards, and adopting an improved PPO algorithm to generate a smooth obstacle avoidance path by B-spline fitting of discrete trajectory points based on the strategy output; Step S5: system verification, specifically, checking the legality and conflict-free of the obstacle avoidance path, if not qualified, feeding back to adjust the model to regenerate, if qualified, packaging the warning information and issuing it to the control center. 2.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method according to claim 1, characterized in that: In step S3, the three-dimensional dynamic trajectory reconstruction, the standard multi-modal data after spatio-temporal alignment and fusion is input into the improved NeRF network to obtain a high-precision three-dimensional dynamic trajectory, comprising the following steps: Step S31: initial digital twin scene construction, inputting the standard multi-modal data for scene initialization, generating a fusion feature map through feature splicing, taking the ground control points calibrated by GPS as the reference, and removing the scene modeling abnormal points through the RANSAC algorithm to ensure that the absolute position error of the initial digital twin scene is ≤3cm; Step S32: incremental dynamic updating, in the initial digital twin scene, to solve the problem of poor real-time performance caused by traditional NeRF full-scene retraining, trajectory change area judgment, block parameter storage, and momentum update strategy are adopted to only optimize the neural radiation field parameters in the trajectory change area, and the parameters in the static area are frozen, thereby realizing efficient dynamic iteration of the digital twin scene; Step S33: point cloud pose correction, inputting the dynamic updated digital twin scene and NeRF parameters, the cumulative error of continuous frame pose estimation will be generated in the iteration process, a double correction mechanism of implicit space normal estimation and global discretization constraint is adopted to control the attitude error of adjacent frames of the trajectory within 0.1°, and a corrected point cloud attitude matrix is obtained; Step S34: BEV feature optimization and fusion, adopting double-head channel attention; BEV feature map generation, the BEV feature map comprises a BEV geometric feature map and a BEV semantic feature map, the BEV geometric feature map is generated by projecting the radar point cloud calibrated by the corrected point cloud attitude matrix, and the BEV semantic feature map is generated by converting the visual semantic segmentation mask into an overhead view; Double-head attention fusion, the BEV geometric features and the BEV semantic features are weighted and fused through attention weights to generate a "space-semantic" double-enhanced BEV fusion feature map; Step S35: generating a high-precision three-dimensional dynamic trajectory, inputting the incrementally updated radiation field parameters, the corrected point cloud attitude matrix and the optimized BEV fusion features into a rendering module of the improved NeRF, and generating a three-dimensional dynamic trajectory through dynamic sampling and radiation field rendering. 3.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method of claim 1, wherein: In step S4, the obstacle avoidance path planning includes the following steps: Step S41: trajectory prediction: inputting the current three-dimensional dynamic trajectory data into an LSTM-Transformer hybrid model to predict 100 trajectory prediction points within 5 seconds in the future; Step S42: bounding box generation: constructing an AABB-OBB hybrid bounding box for each trajectory prediction point and calculating the minimum distance between the UAV bounding boxes; Step S43: conflict warning, specifically constructing a space-time coupled risk index to realize quantitative division of risk levels; Step S44: obstacle avoidance path planning, constructing a reinforcement learning model with priority weighted rewards for multi-priority UAV obstacle avoidance requirements, and prioritizing high-priority UAV passage, including the following steps: Step S441: dynamic state space construction, constructing a high-dimensional dynamic state vector by fusing self-state, environmental state and priority information; Step S442: priority weighted reward function, designing a composite reward mechanism of basic reward, priority reward and constraint penalty to prioritize high-priority UAV passage; Step S443: using an improved PPO algorithm, introducing a priority weighted advantage function to ensure stable policy update and bias towards high-priority optimization goals, and outputting optimized policy network parameters; Input the real-time state of the current airspace into the optimized strategy network, and the strategy network outputs the optimal action in the state according to the optimized strategy network parameters, and maps the optimal action to the three-dimensional coordinate system of the airspace to generate corresponding discrete trajectory point coordinates combined with the time step ; traverse the time step to obtain a complete trajectory sequence ; Step S444: fitting the complete trajectory sequence output by the policy into a smooth curve to generate an obstacle avoidance path. 4.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method of claim 1, wherein: In step S1, the multi-modal data acquisition includes the following steps: Step S11: device deployment, deploying sensing devices according to a preset topology in a monitoring area, and achieving nanosecond-level synchronization among all devices through a dedicated network; Step S12: collecting multi-modal data, including radar data, visual data, ADS-B data and RID data, and performing data preprocessing respectively; obtaining preprocessed multi-modal data. 5.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method of claim 1, wherein: In step S2, the space-time alignment fusion is specifically performed on the preprocessed multi-modal data to obtain standard multi-modal data, including radar point cloud data, visual features and RID position data. 6.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method of claim 1, wherein: In step S5, the system verification is specifically performed to check whether the generated obstacle avoidance path is completely within the legal airspace range and has no border crossing risk. If there are path border crossing and high-priority UAV trajectory conflict problems, feedback to the reinforcement learning model, adjust the reward function coefficients and policy parameters based on the verification error, regenerate the path, and if the verification is passed, encapsulate the path and corresponding warning level information and issue it to the UAV flight control system and airspace control center.

7. An intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning system for implementing an intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning method according to any one of claims 1-6. The multi-modal data acquisition module, specifically for collecting multi-modal data and preprocessing to obtain preprocessed multi-modal data; 8.The intelligent low-altitude unmanned aerial vehicle obstacle avoidance path planning system according to claim 7, characterized in that: ​ The multi-modal data acquisition module, specifically for collecting multi-modal data and preprocessing, obtaining preprocessed multi-modal data; send data to the space-time alignment fusion module; The space-time alignment fusion module, specifically for the preprocessed multi-modal data for space-time alignment fusion, get standard multi-modal data; send data to the three-dimensional dynamic trajectory reconstruction module; The three-dimensional dynamic trajectory reconstruction module, specifically for input standard multi-modal data, initial digital twin scene construction, incremental dynamic update, point cloud pose correction using hidden space normal estimation and global discretization constraint double correction mechanism, double head channel attention for BEV feature optimization fusion, finally generate three-dimensional dynamic trajectory; send data to the obstacle avoidance path planning module; The obstacle avoidance path planning module, specifically for constructing a hybrid bounding box based on trajectory prediction points, calculating the minimum distance between the bounding boxes of the unmanned aerial vehicles, constructing a space-time coupled risk index to divide the conflict level and triggering the corresponding warning, constructing a reinforcement learning model with priority weighted reward, using the improved PPO algorithm, generating a smooth obstacle avoidance path by B-spline fitting based on the discrete trajectory points output by the strategy; send data to the system verification module; The system verification module, specifically for checking the legality and conflict-free of the obstacle avoidance path, if unqualified, feedback adjustment model regeneration, if qualified, package warning information and issue to the control center.

Images