Low-altitude air route obstacle avoidance and safety window generation method and system, and storage medium

Abstract

The application relates to the technical field of unmanned aerial vehicle autonomous navigation and dynamic route planning, in particular to a low-altitude route obstacle avoidance and safety window generation method and system and a storage medium, which comprises the following steps: a dynamic obstacle avoidance decision model comprising a space-time attention network, a model predictive control optimizer and a safety corridor generator is constructed; standardized multi-modal data is input into the dynamic obstacle avoidance decision model, space-time perception features are extracted and fused through the space-time attention network, a space-time safety corridor is calculated through the safety corridor generator, and an optimal obstacle avoidance trajectory and a control instruction are solved under the safety corridor constraint through the model predictive control optimizer; a continuous space-time safety window sequence is generated based on the optimal obstacle avoidance trajectory; the optimal control instruction is executed, and dynamic route monitoring and re-planning are performed according to the safety window sequence. Through deep fusion and feature extraction of multi-modal information, the accuracy of environment perception and the context understanding ability are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of autonomous navigation and dynamic route planning technology for unmanned aerial vehicles (UAVs), specifically to a method, system, and storage medium for low-altitude route obstacle avoidance and safety window generation. Background Technology

[0002] Currently, in low-altitude flight missions such as UAV logistics, urban air traffic (UAM), and emergency rescue, autonomous obstacle avoidance and route planning in dynamic environments are core technologies for ensuring flight safety and mission reliability. The results directly impact the operational efficiency, safety, and robustness of the UAV. Traditional obstacle avoidance methods primarily rely on perception information from a single sensor (such as lidar or vision) and employ rule-based reactive obstacle avoidance (such as artificial potential field methods and dynamic window methods) or sampling-based global planning algorithms (such as RRT and A). While reactive methods offer high real-time performance, they lack consideration for the global mission and the future state of dynamic obstacles, making them prone to getting trapped in local optima or experiencing jitter. Sampling-based global planners, on the other hand, involve large computational demands in dynamic high-dimensional spaces, making it difficult to meet the millisecond-level real-time decision-making requirements.

[0003] With the development of sensor technology, multimodal perception fusion has become crucial for improving environmental perception capabilities. However, effectively aligning and fusing heterogeneous, asynchronous, and multi-precision data from different sources (such as vehicle status, multi-obstacle trajectories, wind field environment, and mission objectives) in time and space, and performing interpretable and verifiable safety planning based on the fused information, remains a challenge in the industry. Existing methods often face the problem of "information silos": the perception, prediction, and planning modules are relatively independent, lacking end-to-end collaborative optimization, resulting in insufficient safety or overly conservative planned trajectories in dynamic environments.

[0004] Model predictive control (MPC) demonstrates advantages in trajectory planning due to its ability to explicitly handle system dynamic constraints and multi-objective optimization. However, its direct application to environments with dense dynamic obstacles presents two core challenges: first, how to efficiently integrate complex, time-varying environmental safety constraints into the MPC optimization framework; and second, how to generate a continuous and connected safety space description (i.e., safety corridors or safety windows) over a future period while ensuring real-time performance, in order to enable forward monitoring and emergency replanning. Existing methods often simplify obstacles into points or static expanding regions, failing to fully consider their motion uncertainties and spatial occupancy evolving over time, resulting in inaccurate safety margin estimates.

[0005] As can be seen from the above, how to achieve deep fusion of multimodal perception information, generate spatiotemporal safety corridors in real time, and perform efficient, smooth, and safe trajectory optimization and monitoring under their constraints in complex and dynamic low-altitude environments is a problem that urgently needs to be solved in low-altitude autonomous flight technology. Summary of the Invention

[0006] The purpose of this invention is to provide a method, system, and storage medium for low-altitude route obstacle avoidance and safety window generation, in order to solve the technical problems of existing obstacle avoidance methods in dense dynamic obstacle scenarios, which are due to isolated perception information, disconnect between planning and safety constraints, lack of spatiotemporal foresight, resulting in insufficient safety of obstacle avoidance decisions, unstable trajectories, poor real-time performance, and inability to provide a continuous and monitorable description of the safety space.

[0007] To solve the above-mentioned technical problems, the present invention specifically provides the following technical solution:

[0008] A method for low-altitude route obstacle avoidance and safety window generation includes the following steps:

[0009] Step S1: Collect multimodal perception data of the low-altitude dynamic environment, including vehicle status modal data V, obstacle dynamic modal data O, environmental modal data E, and route target modal data G;

[0010] Step S2: Standardize and spatiotemporally align the collected modal data to form a spatiotemporally aligned standardized data tensor. , where m represents the modal index;

[0011] Step S3: Construct a dynamic obstacle avoidance decision model that includes a spatiotemporal attention network, a model predictive control optimizer, and a safe corridor generator;

[0012] Step S4: Input the standardized multimodal data into the dynamic obstacle avoidance decision model, extract and fuse spatiotemporal perception features through the spatiotemporal attention network, calculate the spatiotemporal safe corridor through the safe corridor generator, and solve the optimal obstacle avoidance trajectory and control command under the safe corridor constraint through the model prediction control optimizer.

[0013] Step S5: Generate a continuous spatiotemporal safety window sequence based on the optimal obstacle avoidance trajectory;

[0014] Step S6: Execute the optimal control command and perform dynamic route monitoring and replanning according to the safety window sequence.

[0015] In a preferred embodiment of the present invention, in step S1, the autonomous vehicle state modal data V includes the location of the unmanned aerial vehicle. ,speed acceleration and attitude angle data, for The three-dimensional components, for The three-dimensional components;

[0016] The obstacle dynamic modal data O includes the obstacle position. ,speed acceleration ,category and motion uncertainty parameters ,in N is the number of obstacles. for The three-dimensional components;

[0017] The environmental modal data E includes wind speed vectors. air density and visibility Data, including for The three-dimensional components;

[0018] The route target modal data G includes the target point position. Expected arrival time and path reference point sequence ,in for The three-dimensional components.

[0019] As a preferred embodiment of the present invention, the spatiotemporal attention network in step S3 includes:

[0020] The vehicle state encoder employs a multilayer perceptron (MLP) to extract vehicle state features. ;

[0021] The obstacle dynamic encoder employs a combination of a graph attention network (GAT) and a long short-term memory network (LSTM) to extract dynamic spatiotemporal features of obstacles. ;

[0022] An environmental encoder, employing a convolutional neural network (CNN), is used to extract environmental features. ;

[0023] The target encoder, employing an MLP (Multi-Level Processing) algorithm, is used to extract features of targets along the route. ;

[0024] The cross-spatial attention fusion module is used for computation. and Cross-attention between them, and fusion and Information, outputting unified spatiotemporal perception features .

[0025] As a preferred embodiment of the present invention, the method for constructing the model predictive control optimizer in step S3 is as follows:

[0026] Define the prediction time domain With control time domain ;

[0027] Constructing state prediction equations based on UAV dynamics models: ,in For state, For control input;

[0028] Build including tracking cost Control costs Security costs Cost of comfort Optimization objective function ,in For the total cost, To predict the time domain, To control the time domain, In time step The system status, In time step The control input, To control the amount of change in the input;

[0029] Using the aforementioned spatiotemporal safety corridor as a hard constraint and the dynamics and actuator physical limits as boundary constraints, a constrained finite-time optimal control problem is constructed.

[0030] In a preferred embodiment of the present invention, the calculation process of the safety corridor generator in step S3 is as follows:

[0031] based on Obstacle prediction trajectory and uncertainty parameters Calculate the spatiotemporal occupancy region of obstacles in the prediction time domain. ,in To predict the location, For radius, For safety boundaries;

[0032] Calculate the free space at time t: ;

[0033] Will Perform convex decomposition and connectivity checks to form a series of convex polyhedra. ,in L represents the number of convex regions;

[0034] Connecting adjacent convex regions generates a continuous spatiotemporal safety corridor. ,in For the current moment, For time step.

[0035] As a preferred embodiment of the present invention, the method for solving the optimal obstacle avoidance trajectory and control command in step S4 includes:

[0036] The constrained finite-time optimal control problem is transformed into a sequential quadratic programming (SQP) problem.

[0037] In the current forecast state With safety corridor The problem is solved iteratively under constraints, with each iteration solving a quadratic programming (QP) subproblem and updating the control sequence.

[0038] After iterative convergence, the optimal control command sequence is output. and the corresponding predicted state sequence ,in ( () represents the optimal control input at the k-th time step. ( ) for the first Predicting the system state at each time step. ,in Let the optimal predicted position coordinates of the UAV at the k-th time step be... Let be the optimal predicted velocity of the UAV at the k-th time step.

[0039] As a preferred embodiment of the present invention, the method for generating a continuous spatiotemporal safety window sequence in step S5 is as follows:

[0040] Security Window Defined as within the time interval A safe and accessible area within the space;

[0041] From the predicted state sequence Extracting position sequences ,in ( () represents the optimal predicted position coordinates of the UAV at the k-th time step;

[0042] by Centered on, combined The geometry generates a safe window in the form of an ellipsoid or a convex polyhedron. ,satisfy ,and ;

[0043] Connect adjacent security windows to ensure This forms a continuous sequence of spatiotemporal safety windows. .

[0044] As a preferred embodiment of the present invention, step S6 further includes an emergency obstacle avoidance mode, triggered by the presence of an obstacle. This makes it possible to predict collision time. ,in Emergency threshold;

[0045] In emergency obstacle avoidance mode, obstacle avoidance commands are generated in real time using an analytical repulsive field method: ,in This is the original instruction. This is the repulsive force gain, used to control the intensity of the obstacle avoidance response.

[0046] As a preferred embodiment of the present invention, the present invention provides a low-altitude route obstacle avoidance and safety window generation system, which is applied to a method for low-altitude route obstacle avoidance and safety window generation. The system includes:

[0047] The multi-source sensing and data acquisition unit is used to collect multimodal data on vehicle status, obstacle dynamics, environment, and route targets.

[0048] The data preprocessing and alignment unit is used to standardize and spatiotemporally align the data of each modality.

[0049] The dynamic obstacle avoidance decision-making model unit includes a spatiotemporal attention network module, a safe corridor generation module, and a model prediction control optimization module;

[0050] The trajectory optimization and safety window generation unit is used to solve for the optimal trajectory and generate a safety window sequence.

[0051] The control execution and monitoring unit is used to execute control commands and monitor the security window.

[0052] As a preferred embodiment of the present invention, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a method for low-altitude route obstacle avoidance and safety window generation.

[0053] Compared with the prior art, the present invention has the following advantages:

[0054] This invention designs a spatiotemporal attention network that includes a graph attention network (GAT), a long short-term memory network (LSTM), a convolutional neural network (CNN), and a multilayer perceptron (MLP). This network enables deep fusion and feature extraction of multimodal information such as vehicle state, dynamic interaction with multiple obstacles, environmental field, and task objectives, significantly improving the accuracy of environmental perception and contextual understanding.

[0055] The proposed safety corridor generator can dynamically calculate a spatio-temporal safe corridor (ST-SC) over a future period based on obstacle prediction trajectories and their uncertainties. This corridor transforms complex environmental safety information into a series of continuous convex geometric space constraints, which are then directly embedded as hard constraints into the optimization problem of model predictive control (MPC), giving the planned trajectory an inherently verifiable safety guarantee.

[0056] This invention employs a model predictive control (MPC) framework to simultaneously optimize trajectory tracking accuracy, control efficiency, flight comfort, and active safety distance under the hard constraints of a spatiotemporal safety corridor. This integrated optimization method overcomes the problems of error accumulation and suboptimal decision-making in traditional perception-planning-control pipelines, generating globally superior and dynamically feasible obstacle avoidance trajectories. Attached Figure Description

[0057] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0058] Figure 1 A flowchart of a low-altitude route obstacle avoidance and safety window generation method provided in an embodiment of the present invention;

[0059] Figure 2 This is a structural diagram of the dynamic obstacle avoidance decision-making model provided in an embodiment of the present invention;

[0060] Figure 3 This is a block diagram of a low-altitude flight path obstacle avoidance and safety window generation system provided in an embodiment of the present invention. Detailed Implementation

[0061] 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.

[0062] like Figure 1 As shown in the figure, this invention discloses a method for low-altitude route obstacle avoidance and safety window generation, including:

[0063] Step S1: Collect multimodal perception data of the low-altitude dynamic environment, including vehicle status modal data V, obstacle dynamic modal data O, environmental modal data E, and route target modal data G.

[0064] In this embodiment, in response to the environmental perception requirements of UAVs performing missions at low altitudes, this application embodiment first uses airborne sensors (such as GPS / IMU, visual sensors, lidar, millimeter-wave radar, etc.) and external information links to simultaneously collect raw data of four modalities.

[0065] Specifically, the vehicle state modal data V includes the drone's location. ,speed acceleration Attitude angles (roll, pitch, yaw); obstacle dynamic modal data O including obstacle position. ,speed acceleration ,category (Such as static buildings, dynamic vehicles, other drones, etc.) and motion uncertainty parameters ,in N represents the number of obstacles; environmental modal data E includes wind speed vectors. air density and visibility Data; Route target modal data G includes target point positions. Expected arrival time and path reference point sequence .

[0066] Specifically, all data acquisition and processing procedures involved in this invention strictly comply with relevant national laws, regulations, and industry standards. The sole purpose of the environmental and obstacle data collected by the system is to ensure the navigation safety of the aircraft itself and public airspace, and to achieve obstacle avoidance and path planning functions. Especially when processing perception information that may involve non-cooperative targets, the system design follows the principle of 'minimum necessity'. All algorithmic decisions (including emergency obstacle avoidance modes) prioritize the safety of life and property, and do not contain any discriminatory logic based on irrelevant characteristics such as race, ethnicity, religious belief, or personal identity. Furthermore, the algorithm's output will not be used for any other purpose unrelated to flight safety. The implementation of this invention aims to improve the level of automated safety in low-altitude flight, which aligns with social ethics and public interest.

[0067] Step S2: Standardize and spatiotemporally align the collected modal data to form a spatiotemporally aligned standardized data tensor. , where m represents the modal index.

[0068] In this embodiment, due to the different sources, dimensions, sampling frequencies, and timestamps of the data from each modality, preprocessing is required. First, Z-score standardization or min-max normalization is performed on the continuous variables in each modality's data to map the values ​​to similar scales. Second, based on a unified world coordinate system and time reference, spatiotemporal interpolation and alignment are performed on obstacle trajectories, environmental wind speeds, and other data to ensure all data reside on the same spatiotemporal grid. Finally, the data for each modality at the same time are organized into a tensor form. For example, the data for obstacle modality O at time t can be organized as follows: ,in The feature dimensions of each obstacle.

[0069] Step S3: Construct a dynamic obstacle avoidance decision model that includes a spatiotemporal attention network, a model predictive control optimizer, and a safe corridor generator.

[0070] In this embodiment, the dynamic obstacle avoidance decision-making model is the core of the invention. First, a spatiotemporal attention network is constructed, the specific structure of which includes:

[0071] Vehicle status encoder: Employs a multilayer perceptron (MLP) to input standardized vehicle status. Output state features A three-layer multilayer perceptron (MLP) was used, with the number of neurons in each layer being [128, 64, ...]. The ReLU activation function is used. The input is a normalized state vector. The output is the feature. .

[0072] Obstacle Dynamic Encoder: To effectively model the interactions and spatiotemporal evolution among multiple obstacles, a combination of Graph Attention Network (GAT) and Long Short-Term Memory Network (LSTM) is employed. A spatiotemporal graph is constructed with each obstacle as a node, and GAT is used to aggregate information about neighboring obstacles. Then, the processed feature sequence of each obstacle is input into the LSTM to extract its dynamic spatiotemporal features, and the output is... .

[0073] First, a spatiotemporal graph is constructed, with nodes representing obstacles and edges determined by a distance threshold between obstacles. A two-layer Graph Attention Network (GAT) is used, with 4 attention heads per layer and an output feature dimension of 64. The output sequence of the GAT is then fed into a unidirectional two-layer LSTM, with 128 hidden units per layer, to capture the temporal dynamics of each obstacle. The final output features... .

[0074] Environment encoder: Employs a convolutional neural network (CNN) to take rasterized or feature-based environmental data as input. Extract environmental features such as wind field and visibility, and output After rasterizing environmental parameters such as wind speed and density, the data is input into a lightweight CNN containing two convolutional layers (3x3 kernels) and one fully connected layer to extract environmental features. .

[0075] Target encoder: Employs MLP, inputs route target information Extract task intent features and output A two-layer MLP is used to process the target information and output... .

[0076] Cross-spatial-temporal attention fusion module: This module calculates the vehicle's features. Characteristics of each obstacle Cross-attention between different points allows the vehicle to focus on obstacles that pose the greatest threat to it. Simultaneously, it incorporates environmental features. With target features By incorporating gating mechanisms or weighted summation methods, a unified spatiotemporal perception feature vector containing multimodal information is ultimately output. .

[0077] The core of this module is calculating vehicle features. (As a query) with all obstacle features Cross-attention (as key and value). Specifically, firstly, Projected onto dimension through linear transformation ,Will Projected into Key and Value matrices respectively. Attention weights are... The calculated context-aware vehicle features are then compared with... and The data is concatenated and then fused and reduced in dimensionality using a fully connected layer to output unified spatiotemporal awareness features. .

[0078] Secondly, a safe corridor generator is constructed. This module is based on the dynamic obstacle features output by a spatiotemporal attention network. (Contains predicted trajectory and uncertainties) In the prediction time domain The internal computation of the spatio-temporal safe corridor (ST-SC) is as follows:

[0079] For each obstacle Calculate its future moments Spatiotemporal occupied area: ,in, Predict the location of obstacles. Where is the radius of the obstacle. For static safety boundaries, It represents the radius of motion uncertainty that increases over time.

[0080] Calculate the free space at time t: .

[0081] right Perform convex decomposition (e.g., using Voronoi diagrams or fast convex partitioning algorithms) and connectivity checks to form a series of convex polyhedra. .

[0082] Connecting adjacent convex regions at adjacent moments generates a continuous spatiotemporal safety corridor: ;

[0083] Finally, a model predictive control (MPC) optimizer is constructed. The prediction time domain is defined. With control time domain (generally State prediction equations are constructed based on simplified dynamic models of UAVs (such as double integrators or more refined models): ,in For state, To control inputs (such as acceleration or throttle commands), construct an optimization objective function. ,Include:

[0084] Tracking Costs Penalty State and Reference Path The deviation.

[0085] Control Cost Punishment controls the amount of energy used to conserve energy.

[0086] Security Cost Encourage trajectories to move away from obstacles (this can be achieved as a soft constraint or through a hard constraint on the corridor).

[0087] Comfort Cost Punishment is applied to control drastic changes in the amount of force used to ensure smooth flight.

[0088] Specifically, the formula is explained in detail:

[0089] ;

[0090] Total cost. The goal of the optimizer is to find the optimal solution. Minimize the sequence of control instructions.

[0091] : Prediction time domain. Indicates how many time steps the optimizer predicts ahead.

[0092] : Control time domain. Indicates how many time steps the optimizer directly calculates the control instructions (usually) ).

[0093] : at time step The system status typically includes the drone's position, speed, etc.

[0094] : at time step Control inputs, such as acceleration or throttle commands.

[0095] : Control the amount of change in the input, i.e. .

[0096] Specifically, tracking costs :

[0097] Function: Penalize drone status The deviation from the expected reference path or target point.

[0098] Significance: To ensure the generated trajectory stays on track and ultimately accomplishes the core mission of reaching the target point. Completely deviating from the flight path would incur significant costs.

[0099] Control Cost :

[0100] Function: To punish the magnitude of the control quantity itself.

[0101] Significance: Encourages the use of smaller, more energy-efficient control commands and avoids unnecessary aggressive operations, thereby saving drone energy.

[0102] Comfort Cost :

[0103] Function: To punish drastic changes in the amount of control (i.e.) ).

[0104] Significance: Ensuring smooth flight and avoiding sudden acceleration changes. This is crucial for missions carrying sophisticated equipment (such as cameras and sensors) or requiring stable flight (such as logistics delivery and manned flights).

[0105] Security Cost :

[0106] Function: Penalize drone status Too close to the obstacle.

[0107] Significance: By proactively taking "avoiding obstacles" as one of the optimization objectives, a "soft" safety margin optimization is added to the rigid safety corridor constraint (as claimed in the claim), so that the generated trajectory not only stays within the corridor, but also tends to walk in a safer position in the center of the corridor.

[0108] Ultimately, a time-space safety corridor As a hard constraint (requiring the predicted state sequence to be within the corridor), a constrained finite-time optimal control problem is constructed using the UAV dynamic limit and the actuator physical limit as boundary constraints.

[0109] Dynamic obstacle avoidance decision-making models require supervised learning training. The training process includes the following steps:

[0110] Training dataset construction: Generate a large amount of multimodal perception data using a high-fidelity drone simulation environment (such as AirSim) or collected real flight logs. , , , and its corresponding expert teaching trajectory or optimized safety trajectory: ideal control command sequence and ideal state trajectory sequence Sample pairs, The ideal / optimal control command for the k-th time step (such as throttle, yaw, take-off and landing commands for the UAV). This represents the ideal / optimal UAV state at time step k (e.g., core state parameters such as 3D position, velocity, and attitude). The data must cover various obstacle densities, motion patterns, and weather conditions.

[0111] Loss function design: The total loss function consists of multiple parts:

[0112] Control command regression loss L ctrl The average error between the control commands output by the calculation model (such as throttle) and the "standard answer" is calculated (using the second norm to calculate the distance), so that the operation commands output by the model are close to the level of experts;

[0113] Trajectory prediction loss L traj The average error between the model's predicted drone position / velocity (state) and the "standard answer" trajectory is calculated, making the model's predicted flight path more accurate;

[0114] Safety corridor constraint loss L safe If the trajectory predicted by the model is close to the "danger zone" (beyond the safety corridor S) Ck safety margin If the model is in a safe zone, a penalty score is added; if it is in a safe zone, the penalty is 0, forcing the model to avoid danger and ensuring flight safety.

[0115] Feature learning loss L for MPC solution mpc Let the model output features (Z) st It can help the MPC module (real-time trajectory optimizer) find the optimal solution faster and improve the response speed of the model in actual use.

[0116] The total loss is: L = λ1L ctrl +λ2L traj +λ3L safe +λ4L mpc λ1 to λ4 are the trade-off hyperparameters, used to weigh the importance of different loss terms in the overall objective. Their specific values ​​can be determined through automated hyperparameter tuning methods such as GridSearch, Random Search, or Bayesian Optimization on an independent validation set to maximize the overall performance of the model (e.g., a combined score of obstacle avoidance success rate and trajectory smoothness). In a specific training example, the value ranges are λ1∈[0.5,2.0], λ2∈[0.5, 1.5], λ3∈[0.2, 1.0], and λ4∈[0.1, 0.5]. After tuning, the preferred hyperparameter values ​​are: λ1 = 1.0, λ2 = 0.8, λ3 = 0.5, λ4 = 0.2. Those skilled in the art should understand that these values ​​are examples and can be adapted to different drone platforms, mission scenarios, and sensor configurations.

[0117] Training parameters and process: The Adam optimizer was used with an initial learning rate of 1e-4, batch gradient descent was employed, and the batch size was 32. Training lasted for at least 50 epochs, with loss monitored on independent validation sets. Early stopping was used to prevent overfitting. The model was finally evaluated for obstacle avoidance success rate and trajectory quality in unseen test scenarios.

[0118] Step S4: Input the standardized multimodal data into the dynamic obstacle avoidance decision model, extract and fuse spatiotemporal perception features through a spatiotemporal attention network, calculate the spatiotemporal safe corridor through a safe corridor generator, and solve the optimal obstacle avoidance trajectory and control command under the constraint of the safe corridor through a model predictive control optimizer.

[0119] In this embodiment, at each decision cycle, the multimodal data tensor processed in step S2 is input into the trained dynamic obstacle avoidance decision model. The model first obtains fused features through a spatiotemporal attention network. and obstacle dynamic characteristics Subsequently, the safe corridor generator is based on Real-time generation of the future During the time period Finally, the MPC optimizer uses the current state... As initial conditions, in Solve the optimal control problem under hard constraints and other boundary constraints.

[0120] Specifically, the solution process employs the Sequential Quadratic Programming (SQP) method: the original problem is transformed into a series of quadratic programming (QP) subproblems to be solved iteratively. In each iteration, at the current trajectory prediction point, the nonlinear dynamics and constraints are linearized / approximated quadratically, solving a QP subproblem to obtain the control increment, updating the trajectory prediction, until the convergence condition is met. After iterative convergence, the optimal control command sequence is output. and the corresponding predicted state sequence ,Pick As the control command to be executed at the current moment.

[0121] Step S5: Generate a continuous spatiotemporal safety window sequence based on the optimal obstacle avoidance trajectory.

[0122] In this embodiment, the safety window is a "slice" of the safety corridor at discrete time points, providing a more intuitive representation of the safety area for monitoring and replanning. From the optimal predicted state sequence... Extracting position sequences For each moment and their corresponding positions Combined with the free space convex region at that moment Generate a Centered on, contained in The geometry as a security window For ease of calculation, ellipsoids or convex polyhedra (such as cuboids) along the trajectory direction are often used. Ensure... ,and Finally, connect adjacent security windows to ensure... This forms a continuous sequence of spatiotemporal safety windows. This sequence can be provided to monitoring systems to determine in real time whether the drone is within a safe window and to provide a spatial basis for triggering replanning in case of possible emergencies.

[0123] Step S6: Execute the optimal control command and perform dynamic route monitoring and replanning based on the safety window sequence.

[0124] In this embodiment, the optimal control command obtained in step S4 is... The data is sent to the drone flight control system for execution. Simultaneously, the monitoring system continuously receives new sensor data and determines whether the drone's actual state deviates from the safe window sequence. Or, is there a new, unexpected obstacle? If a deviation occurs or a new, high-threat obstacle appears, a replanning is immediately triggered, starting a new decision-making cycle from step S1.

[0125] Furthermore, this method also includes an emergency obstacle avoidance mode. Its trigger condition is: the presence of an obstacle is detected in real time. This makes it possible to predict collision time. ,in A preset emergency threshold (e.g., 0.5 seconds) is set. Once triggered, the system will temporarily switch to emergency obstacle avoidance mode. In this mode, an analytical repulsive field method, which has a faster computation speed, is used to generate obstacle avoidance correction commands in real time. ,in This is the original instruction. This mode prioritizes collision avoidance and can temporarily relax some comfort constraints or precise corridor constraints. Once the emergency situation is resolved, it will revert to the MPC-based fine-grained planning mode.

[0126] In one specific implementation, taking a city logistics drone delivery scenario as an example, the drone needs to fly between buildings and avoid other drones that suddenly appear (dynamic obstacles) and known static buildings. First, the drone's status is collected by sensors, three dynamic obstacles (parameters shown in Table 1) and the outlines of static buildings are identified, and wind speed information and preset target points are obtained at the same time.

[0127]

[0128] Secondly, the data was standardized and aligned. Then, the constructed spatiotemporal attention network effectively integrated information from the vehicle, obstacles, wind field, and target. The cross-attention mechanism ensured the vehicle focused on the oncoming obstacle with ID 1. The safe corridor generator, based on the predicted obstacle trajectory and uncertainty, generated a continuously changing convex polyhedral safe corridor over the next 10 seconds. Under this corridor constraint, the MPC optimizer solved for a smooth, energy-efficient, and safe obstacle avoidance trajectory and generated the corresponding safe window sequence. After executing control commands, the drone successfully avoided the obstacle. When simulating a sudden, rapidly approaching obstacle, the system triggered an emergency obstacle avoidance mode, applying an additional lateral repulsive force to ensure safety.

[0129] To verify the effectiveness of this method, it was compared with the traditional artificial potential field (APF) method and sampling-based real-time planning methods (such as RRT). The key performance indicators were compared in the same 100 random dynamic scenario simulations, as shown in Table 2.

[0130]

[0131] This method, through multimodal perception fusion and spatiotemporal safety corridor generation mechanism, can more comprehensively and proactively assess environmental risks, thereby achieving a higher obstacle avoidance success rate in complex dynamic scenarios. Traditional artificial potential field (APF) methods are prone to getting trapped in local minima, while sampled planning (RRT) struggles to guarantee real-time performance and trajectory consistency in dynamic environments, resulting in low success rates.

[0132] This method employs a Model Predictive Control (MPC) framework, explicitly introducing a comfort cost term (penalizing abrupt changes in control variables) into the optimization objective, thereby generating a smooth trajectory that conforms to UAV dynamics. In contrast, the APF method is prone to producing oscillating trajectories, and the trajectories generated by RRT-like methods are typically polygonal, requiring post-processing smoothing.

[0133] This method integrates an emergency obstacle avoidance mode. When the predicted collision time is below a threshold, it immediately switches to an analytical obstacle avoidance strategy based on a repulsive field, achieving a millisecond-level response. Traditional MPC or sampling methods are computationally inefficient enough to meet such extreme real-time requirements.

[0134] This method is the only one capable of outputting a continuous spatiotemporal safety window sequence, providing a structured and interpretable spatial basis for real-time route monitoring, dynamic replanning, and safety verification. Traditional methods typically output only a single trajectory, lacking a description of the surrounding safety space.

[0135] The Graph Attention Network (GAT) employed in this invention can explicitly model the spatial interaction relationships between multiple dynamic obstacles, and its attention mechanism can automatically identify the obstacles that pose the greatest threat to the vehicle. The Long Short-Term Memory Network (LSTM) is used to capture the temporal patterns of each obstacle's movement; the combination of these two technologies accurately characterizes the evolution of the dynamic environment. A cross-attention mechanism associates the vehicle's state with the characteristics of the obstacle group, achieving contextual awareness centered on the vehicle. Finally, Model Predictive Control (MPC) uses the safety corridor generated by these spatiotemporally semantically rich features as a hard constraint for trajectory optimization, ensuring the unity of physical feasibility and safety in decision-making. Therefore, the algorithmic and technical features of this invention functionally support each other and interact with each other, together forming a complete technical solution for solving the problem of low-altitude dense dynamic obstacle avoidance.

[0136] As can be seen from the above, the embodiments of this application first collect and fuse multimodal perception data, and extract rich environmental features through a spatiotemporal attention network; then, based on obstacle prediction, a spatiotemporal safety corridor is dynamically generated, providing a clear safety constraint space for trajectory optimization; next, model predictive control is used to solve for the optimal trajectory within the corridor that balances tracking, safety, energy consumption, and comfort, and a continuous sequence of safety windows is derived; finally, through execution and monitoring closed loops, combined with an emergency obstacle avoidance mechanism, real-time flight safety is ensured. In this way, the method significantly improves the safety, reliability, and overall efficiency of UAV obstacle avoidance decision-making in low-altitude dynamic environments.

[0137] like Figure 3 As shown, this embodiment of the invention also provides a low-altitude route obstacle avoidance and safety window generation system, including:

[0138] The multi-source sensing and data acquisition unit is used to collect multimodal data on vehicle status, obstacle dynamics, environment, and route targets.

[0139] The data preprocessing and alignment unit is used to standardize and spatiotemporally align the data of each modality.

[0140] The dynamic obstacle avoidance decision-making model unit integrates a spatiotemporal attention network module, a safe corridor generation module, and a model prediction control optimization module;

[0141] The trajectory optimization and safety window generation unit is used to solve for the optimal trajectory and generate a safety window sequence.

[0142] The control execution and monitoring unit is used to execute control commands, monitor safety windows, and trigger replanning or emergency obstacle avoidance.

[0143] Furthermore, embodiments of this application also disclose an electronic device, including at least one processor, at least one memory, a communication interface, an input / output interface, and a communication bus. The memory stores a computer program, which is loaded and executed by the processor to implement the relevant steps in the low-altitude flight path obstacle avoidance and safety window generation method disclosed in any of the foregoing embodiments.

[0144] Furthermore, this application also discloses a computer-readable storage medium for storing a computer program; wherein, when executed by a processor, the computer program implements the aforementioned low-altitude flight path obstacle avoidance and safety window generation method. Specific steps of this method can be found in the corresponding content disclosed in the foregoing embodiments, and will not be repeated here.

[0145] The above embodiments are merely exemplary embodiments of this application and are not intended to limit this application. The scope of protection of this application is defined by the claims. Those skilled in the art can make various modifications or equivalent substitutions to this application within its substance and scope of protection, and such modifications or equivalent substitutions should also be considered to fall within the scope of protection of this application.

Claims

1. A method for low-altitude route obstacle avoidance and safety window generation, characterized in that, Includes the following steps: Step S1: Collect multimodal perception data of the low-altitude dynamic environment, including vehicle status modal data V, obstacle dynamic modal data O, environmental modal data E, and route target modal data G; Step S2: Standardize and spatiotemporally align the collected modal data to form a spatiotemporally aligned standardized data tensor. , where m represents the modal index; Step S3: Construct a dynamic obstacle avoidance decision model that includes a spatiotemporal attention network, a model predictive control optimizer, and a safe corridor generator; Step S4: Input the standardized multimodal data into the dynamic obstacle avoidance decision model, extract and fuse spatiotemporal perception features through the spatiotemporal attention network, calculate the spatiotemporal safe corridor through the safe corridor generator, and solve the optimal obstacle avoidance trajectory and control command under the safe corridor constraint through the model prediction control optimizer. Step S5: Generate a continuous spatiotemporal safety window sequence based on the optimal obstacle avoidance trajectory; Step S6: Execute the optimal control command and perform dynamic route monitoring and replanning according to the safety window sequence.

2. The method for low-altitude route obstacle avoidance and safety window generation according to claim 1, characterized in that: In step S1, the autonomous vehicle state modal data V includes the drone's location. ,speed acceleration and attitude angle data, for The three-dimensional components, for The three-dimensional components; The obstacle dynamic modal data O includes the obstacle position. ,speed acceleration ,category and motion uncertainty parameters ,in N is the number of obstacles. for The three-dimensional components; The environmental modal data E includes wind speed vectors. air density and visibility Data, including for The three-dimensional components; The route target modal data G includes the target point position. Expected arrival time and path reference point sequence ,in for The three-dimensional components.

3. The method for low-altitude route obstacle avoidance and safety window generation according to claim 2, characterized in that: The spatiotemporal attention network in step S3 includes: The vehicle state encoder employs a multilayer perceptron (MLP) to extract vehicle state features. ; The obstacle dynamic encoder employs a combination of a graph attention network (GAT) and a long short-term memory network (LSTM) to extract dynamic spatiotemporal features of obstacles. ; An environmental encoder, employing a convolutional neural network (CNN), is used to extract environmental features. ; The target encoder, employing an MLP (Multi-Level Processing) algorithm, is used to extract features of targets along the route. ; The cross-spatial attention fusion module is used for computation. and Cross-attention between them, and fusion and Information, outputting unified spatiotemporal perception features .

4. The method for low-altitude route obstacle avoidance and safety window generation according to claim 3, characterized in that: In step S3, the method for constructing the model predictive control optimizer is as follows: Define the prediction time domain With control time domain ; Constructing state prediction equations based on UAV dynamics models: ; Build including tracking cost Control costs Security costs Cost of comfort Optimization objective function ,in For the total cost, To predict the time domain, To control the time domain, For the system state at time step k, For the control input at the k-th time step, To control the amount of change in the input; Using the aforementioned spatiotemporal safety corridor as a hard constraint and the dynamics and actuator physical limits as boundary constraints, a constrained finite-time optimal control problem is constructed.

5. The method for low-altitude route obstacle avoidance and safety window generation according to claim 4, characterized in that: In step S3, the calculation process of the safe corridor generator is as follows: based on Obstacle prediction trajectory and uncertainty parameters Calculate the spatiotemporal occupancy region of obstacles in the prediction time domain. ,in To predict the location, For radius, For safety boundaries; Calculate the free space at time t: ; Will Perform convex decomposition and connectivity checks to form a series of convex polyhedra. ,in L represents the number of convex regions; Connecting adjacent convex regions generates a continuous spatiotemporal safety corridor. ,in For the current moment, For time step.

6. The method for low-altitude route obstacle avoidance and safety window generation according to claim 5, characterized in that: The method for solving the optimal obstacle avoidance trajectory and control command in step S4 includes: The constrained finite-time optimal control problem is transformed into a sequential quadratic programming (SQP) problem. In the current forecast state With safety corridor The problem is solved iteratively under constraints, with each iteration solving a quadratic programming (QP) subproblem and updating the control sequence. After iterative convergence, the optimal control command sequence is output. and the corresponding predicted state sequence ,in ( () represents the optimal control input at the k-th time step. ( ) for the first Predicting the system state at each time step. , Let the optimal predicted position coordinates of the UAV at the k-th time step be... Let be the optimal predicted velocity of the UAV at the k-th time step.

7. The method for low-altitude route obstacle avoidance and safety window generation according to claim 6, characterized in that: The method for generating a continuous spatiotemporal safety window sequence in step S5 is as follows: Security Window Defined as within the time interval A safe and accessible area within the space; From the predicted state sequence Extracting position sequences ,in ( () represents the optimal predicted position coordinates of the UAV at the k-th time step; by Centered on, combined The geometry generates a safe window in the form of an ellipsoid or a convex polyhedron. ,satisfy ,and ; Connect adjacent security windows to ensure This forms a continuous sequence of spatiotemporal safety windows. .

8. The method for low-altitude route obstacle avoidance and safety window generation according to claim 7, characterized in that: Step S6 also includes an emergency obstacle avoidance mode, triggered by the presence of an obstacle. This makes it possible to predict collision time. ,in As an emergency threshold, Euclidean distance; In emergency obstacle avoidance mode, obstacle avoidance commands are generated in real time using an analytical repulsive field method: ,in This is the original instruction. This is the repulsive force gain, used to control the intensity of the obstacle avoidance response. It is the square of the Euclidean distance.

9. A low-altitude flight path obstacle avoidance and safety window generation system, characterized in that, The system applied to the low-altitude flight path obstacle avoidance and safety window generation method according to any one of claims 1-8, the system comprising: The multi-source sensing and data acquisition unit is used to collect multimodal data on vehicle status, obstacle dynamics, environment, and route targets. The data preprocessing and alignment unit is used to standardize and spatiotemporally align the data of each modality. The dynamic obstacle avoidance decision-making model unit includes a spatiotemporal attention network module, a safe corridor generation module, and a model prediction control optimization module; The trajectory optimization and safety window generation unit is used to solve for the optimal trajectory and generate a safety window sequence. The control execution and monitoring unit is used to execute control commands and monitor the security window.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 8.

Images