Unmanned aerial vehicle autonomous obstacle avoidance and path planning system based on deep learning and planning method thereof

Abstract

This invention relates to a deep learning-based autonomous obstacle avoidance and path planning system and method for unmanned aerial vehicles (UAVs). The system includes: a perception simulation module, which uses multi-sensor joint modeling to achieve dynamic environmental reconstruction; a decision control and dynamics modeling module, which generates a series of control commands after acquiring fused perception data; and a scene interaction module, which realizes dynamic obstacle behavior modeling and multi-scene library access. The perception simulation module, decision control and dynamics modeling module, and scene interaction module are integrated with a ROS-MATLAB joint interface platform to achieve real-time synchronization and visualization analysis of multi-source heterogeneous data. This invention proposes a dynamic obstacle prediction model driven by multimodal data, using attention weights to dynamically allocate sensor confidence, reducing obstacle velocity prediction error to 0.12 m / s; and designs a hierarchical reward function and a priority experience replay strategy, improving the DRL training convergence speed to 40% and achieving a Pareto optimal state for path length and energy consumption.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) control technology, specifically to a deep learning-based UAV autonomous obstacle avoidance and path planning system and its planning method. Background Technology

[0002] With the rapid penetration of drone technology into logistics, disaster relief, and agricultural inspection, its autonomous operation capability in complex and dynamic environments has become a major bottleneck restricting its large-scale application. Traditional drones mostly rely on pre-set paths or rule-based obstacle avoidance methods. However, in scenarios involving randomly moving obstacles, sensor interference from noise, and multi-drone collaborative operations, they often face problems such as slow real-time decision-making and poor environmental adaptability. Especially in tasks like urban low-altitude logistics and earthquake search and rescue, drones must achieve millimeter-level obstacle avoidance and dynamic path adjustments in confined spaces. This places extremely high demands on the overall coordination of the perception-decision-control chain. How to overcome the technical challenges of autonomous obstacle avoidance and efficient path planning in dynamic and uncertain environments has become a focus of attention for both academia and industry.

[0003] Deep learning technology offers a new approach to solving these challenges. Deep reinforcement learning (DRL), relying on end-to-end environmental interaction and reward mechanisms, can autonomously explore optimal obstacle avoidance strategies in complex environments, greatly enhancing the ability to predict dynamic obstacles and make real-time decisions. However, current research still faces three limitations: insufficient sufficiency of multimodal sensor data fusion, making it difficult to achieve ideal environmental perception accuracy; single algorithms struggle to simultaneously achieve optimal global path performance and real-time response in local obstacle avoidance; and significant differences exist between the simulation environment and the modeling of real physical systems, hindering the engineering implementation of algorithms. Furthermore, in multi-machine collaborative scenarios, conflict resolution and resource allocation issues have not yet been universally resolved, necessitating the development of a more realistic high-fidelity simulation verification system. Summary of the Invention

[0004] To address the aforementioned issues, this invention proposes a deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles (UAVs) and its planning method.

[0005] A deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles includes: The perception simulation module uses multi-sensor joint modeling to achieve dynamic reconstruction of the environment; The decision control and dynamics modeling module acquires fused sensing data and generates a series of control commands. The scene interaction module enables dynamic obstacle behavior modeling and multi-scene library calls.

[0006] The perception simulation module, decision control and dynamics modeling module, and scene interaction module are integrated with a ROS-MATLAB joint interface platform to achieve real-time synchronization and visualization analysis of multi-source heterogeneous data.

[0007] The perception simulation module collects and integrates data streams from multiple sensors in real time, and outputs a unified environmental state tensor using a spatiotemporal alignment module and a DGA attention mechanism; the decision control and dynamics modeling module loads the PPO model that has been trained through multiple simulations and derives the desired action vector by combining dynamic constraints.

[0008] The perception simulation module includes a lidar that provides high-precision three-dimensional spatial information in environmental perception, a visual sensor for acquiring images, video data and providing semantic information in environmental perception, an ultrasonic sensor that measures the distance to obstacles by emitting ultrasonic waves and receiving their reflected waves, an inertial measurement unit for providing attitude and acceleration information of the UAV, and a multimodal database processed by a dynamic gating attention fusion algorithm.

[0009] The planning method of the deep learning-based UAV autonomous obstacle avoidance and path planning system comprises the following specific steps: Step 1: Construct a mathematical model. Relying on the UAV kinematic model, define the state space as a combination of position, heading angle and control commands. Use differential equations to characterize the motion trajectory. Given the uncertainty of dynamic obstacles, use LSTM network and Kalman filter to jointly predict the time-varying trajectory of obstacles. Then, combine the noise distribution model corresponding to multiple sensors (LiDAR / vision / IMU) to set dynamic safety distance constraints. Step 2: At the path optimization level, a multi-objective cost function is established, listing path length, energy consumption, and safety as game variables. A dynamic weight adjustment method is adopted to adjust the priority of each objective in real time. If a nearby obstacle is detected, the weight coefficient of the safety item is adaptively increased so that the drone can avoid collision risks first. If in an open area, the path smoothness is optimized to reduce energy consumption. Step 3: Using a deep reinforcement learning framework, environmental perception information (such as predicted trajectories of obstacles and sensor data) is mapped into a state space. Speed ​​control commands are generated through a hierarchical decision network. The reward function is designed to comprehensively consider path tracking accuracy, energy efficiency, and collision penalties. The generalization ability of the model is improved by combining offline training with online fine-tuning.

[0010] In step three, a feature extraction mode based on spatiotemporal alignment is designed: LiDAR data generates a 3D occupancy grid map using voxelized mesh, visual data uses a lightweight YOLOv5 network to identify the contours of dynamic obstacles, and IMU data uses Kalman filtering to compensate for its attitude drift. Timestamp synchronization and coordinate system calibration are used to eliminate spatiotemporal deviations between multiple sensors and construct a unified environmental representation. This preprocessing process can reduce the alignment error of the original data to within 0.05m, meeting the accuracy requirements of real-time obstacle avoidance. This deep reinforcement learning framework, based on Proximal Policy Optimization (PPO), leverages multi-dimensional algorithms to enhance the comprehensive capabilities of obstacle avoidance and path planning. From the perspective of network architecture, it presents a dual-channel spatiotemporal feature fusion mechanism: the perception channel uses a convolutional network to extract spatial semantic information (such as obstacle distribution and feasible region contours) from the multimodal fusion data, while the temporal channel uses a recurrent neural network to capture the motion state of dynamic obstacles. The two channels are then fused after attention weighting to generate a joint feature representation. This design effectively solves the problem of sudden action changes caused by the discretization of environmental states in traditional methods, and greatly enhances the continuity and smoothness of control commands. The planning method also includes simulation verification steps, the specific steps of which are as follows: S1. Scene Setup: Built using the Unity3D engine; the core area is a 30m×30m×20m cube space; the coordinate system axis range is set to X: [-50,90], Y: [-30,70], Z: [-20,50], unit: m; spatial parameters are quantitatively represented by discretized mesh division. S2. Scene setup: The initial position of the drone is set at the origin (0,0,0); the global target point is located at (80,60,30); path planning needs to achieve safe navigation in a dynamic trajectory interweaving environment. Four types of dynamic obstacles are set, and their movement paths are represented by red, yellow, blue and green trajectories respectively. S3. Multimodal Data Generation and Visualization: The experiment generates LiDAR point clouds, visual images, and IMU data streams in real time using a physics engine. The LiDAR uses a 16-line scanning configuration with a horizontal angular resolution of 0.2°, a vertical angular resolution of 2°, and a maximum detection distance of 50m. The point cloud density is dynamically adjusted according to the distance to obstacles. The visual sensor outputs 1280×720 resolution images at a frequency of 30Hz. Motion blur and illumination noise are introduced to simulate real imaging defects. The 3D trajectory data is generated by B-spline curve interpolation. The position error of dynamic obstacles follows a Gaussian distribution with a mean of zero and a standard deviation of 0.15m. The velocity fluctuation range is precisely adjusted by a parametric controller. The spatiotemporal state of all moving entities in the scene is synchronized with a 0.01s timestamp. S4. Scene Validity Verification: By comparing the actual trajectory of the UAV with the theoretical optimal path, the dynamic complexity of the scene is quantitatively evaluated. The static scene baseline is 0.15m. The strong interference characteristics of dynamic obstacles are verified, and the Fréchet distance is used to analyze the trajectory similarity.

[0011] The beneficial effects of this invention are as follows: This invention proposes a dynamic obstacle prediction model driven by multimodal data, and uses attention weights to dynamically allocate sensor confidence, reducing the obstacle speed prediction error to 0.12m / s; it designs a hierarchical reward function and a priority experience replay strategy, which improves the DRL training convergence speed to 40% and also achieves the Pareto optimal situation of path length and energy consumption; it establishes a joint simulation system that includes motor loss and communication delay, and compared with the traditional Gazebo platform, the accuracy of sensor noise modeling is increased by 35 percentage points, creating a reliable verification benchmark for the engineering deployment of the algorithm. Attached Figure Description

[0012] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0013] Figure 1 This is a schematic diagram of the system of the present invention; Figure 2 This is a schematic diagram illustrating the scenario effectiveness verification of the present invention; Figure 3 This is a schematic diagram of the mathematical model of the present invention; Figure 4 This is a schematic diagram of the deep learning framework of the present invention. Detailed Implementation

[0014] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below.

[0015] like Figures 1 to 4 As shown, the deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles includes: The perception simulation module uses multi-sensor joint modeling to achieve dynamic reconstruction of the environment; The decision control and dynamics modeling module acquires fused sensing data and generates a series of control commands. The scene interaction module enables dynamic obstacle behavior modeling and multi-scene library calls.

[0016] The perception simulation module, decision control and dynamics modeling module, and scene interaction module are integrated with a ROS-MATLAB joint interface platform to achieve real-time synchronization and visualization analysis of multi-source heterogeneous data.

[0017] The perception simulation module collects and integrates data streams from multiple sensors in real time, and outputs a unified environmental state tensor using a spatiotemporal alignment module and a DGA attention mechanism; the decision control and dynamics modeling module loads the PPO model that has been trained through multiple simulations and derives the desired action vector by combining dynamic constraints.

[0018] The perception simulation module includes a lidar that provides high-precision three-dimensional spatial information in environmental perception, a visual sensor for acquiring images, video data and providing semantic information in environmental perception, an ultrasonic sensor that measures the distance to obstacles by emitting ultrasonic waves and receiving their reflected waves, an inertial measurement unit for providing attitude and acceleration information of the UAV, and a multimodal database processed by a dynamic gating attention fusion algorithm.

[0019] The aforementioned visual sensor can be a camera, such as a binocular camera, which is mainly used in environmental perception to acquire image and video data and provide semantic information, such as object classification, target recognition and tracking. The visual sensor has high resolution and rich detail information, and can accurately identify the contour, color and texture features of objects. The main task of the camera is to identify dynamic obstacles through a lightweight YOLOv5 network and extract their semantic features.

[0020] The inertial measurement unit can be an IMU sensor. In dynamic and complex environments, IMU data can provide stable navigation information and make up for the shortcomings of other sensors. The main function of the IMU is to compensate for the drift problem in visual and LiDAR data caused by rapid movement or attitude changes. Through the Kalman filter algorithm, IMU data can be effectively fused to reduce the attitude drift of the system and ensure the accuracy of the multi-sensor fusion results.

[0021] Ultrasonic sensors, as short-range sensing devices, are widely used in environmental detection and obstacle avoidance systems. They measure the distance to obstacles by emitting ultrasonic waves and receiving their reflected waves. Although their accuracy is not as high as LiDAR, ultrasonic sensors are less expensive and perform well in short-range obstacle detection. In multimodal sensing systems, ultrasonic sensors are primarily used for short-range obstacle detection, especially when drones are flying at low altitudes. They can supplement the blind spots of LiDAR and cameras, further improving the reliability of obstacle recognition.

[0022] This invention utilizes the collaborative operation of LiDAR, a camera, an ultrasonic sensor, and an IMU (Insulated Unit), enabling the system to efficiently perceive and accurately identify obstacles in various complex environments. Under good lighting conditions, the visual sensor provides rich semantic information; under poor lighting conditions, LiDAR data provides precise spatial structure information; and the ultrasonic sensor further enhances close-range detection capabilities. The IMU ensures the system's stability and accuracy in dynamic environments.

[0023] This invention, through the rational configuration and integration of these hardware devices, can fully leverage the advantages of each sensor, improve the UAV's perception capabilities in complex and dynamic environments, reduce errors and delays, enhance the system's real-time performance and reliability, and provide more accurate environmental data for subsequent path planning and decision-making.

[0024] This invention proposes a dynamic obstacle prediction model driven by multimodal data. It employs attention weights to dynamically allocate sensor confidence, reducing obstacle speed prediction error to 0.12 m / s. A hierarchical reward function and a priority experience replay strategy are designed to improve the convergence speed of DRL training by 40% and achieve Pareto optimality for path length and energy consumption. A joint simulation system including motor loss and communication delay is constructed. Compared with the traditional Gazebo platform, the accuracy of sensor noise modeling is increased by 35 percentage points, creating a reliable verification benchmark for the engineering deployment of the algorithm.

[0025] The decision control and dynamics modeling module includes a PID controller that serves as the underlying flight control system to perform trajectory tracking tasks, an outer loop that generates attitude commands according to the Lyapunov stability criterion, and an inner loop that uses quaternion calculations to drive the motor speed.

[0026] The scenario library includes three typical behaviors: random walking, cooperative avoidance, and sudden intrusion.

[0027] The planning method of the deep learning-based UAV autonomous obstacle avoidance and path planning system comprises the following specific steps: Step 1: Construct a mathematical model. Relying on the UAV kinematic model (also known as the quadrotor rigid body model), the state space is defined as a combination of position, heading angle, and control commands. Differential equations are used to characterize the motion trajectory. Given the uncertainty of dynamic obstacles, LSTM network and Kalman filter are used to jointly predict the time-varying trajectory of obstacles. Then, combined with the noise distribution model corresponding to multiple sensors (LiDAR / vision / IMU), dynamic safe distance constraints are established. Specifically, the principle of the mathematical model is as follows: the throttle signal is given to the dynamic model, which in turn generates a propeller speed signal that is given to the control efficiency model. For the quadcopter rigid body model, this will generate force, torque, velocity, and angular velocity feedback signals, thereby completing the control of position and attitude.

[0028] Step 2: At the path optimization level, a multi-objective cost function is established, listing path length, energy consumption, and safety as game variables. A dynamic weight adjustment method is adopted to adjust the priority of each objective in real time. If a nearby obstacle is detected, the weight coefficient of the safety item is adaptively increased so that the drone can avoid collision risks first. If in an open area, the path smoothness is optimized to reduce energy consumption. Step 3: Using a deep reinforcement learning framework, environmental perception information (such as predicted trajectories of obstacles and sensor data) is mapped into a state space. Speed ​​control commands are generated through a hierarchical decision network. The reward function is designed to comprehensively consider path tracking accuracy, energy efficiency, and collision penalties. The generalization ability of the model is improved by combining offline training with online fine-tuning.

[0029] In step three, a feature extraction mode based on spatiotemporal alignment is designed: LiDAR data generates a 3D occupancy grid map using voxelized mesh, visual data uses a lightweight YOLOv5 network to identify the contours of dynamic obstacles, and IMU data uses Kalman filtering to compensate for its attitude drift. Timestamp synchronization and coordinate system calibration are used to eliminate spatiotemporal deviations between multiple sensors and construct a unified environmental representation. This preprocessing process can reduce the alignment error of the original data to within 0.05m, meeting the accuracy requirements of real-time obstacle avoidance. This deep reinforcement learning framework, based on Proximal Policy Optimization (PPO), leverages multi-dimensional algorithms to enhance the comprehensive capabilities of obstacle avoidance and path planning. From the perspective of network architecture, it presents a dual-channel spatiotemporal feature fusion mechanism: the perception channel uses a convolutional network to extract spatial semantic information (such as obstacle distribution and feasible region contours) from the multimodal fusion data, while the temporal channel uses a recurrent neural network to capture the motion state of dynamic obstacles. The two channels are then fused after attention weighting to generate a joint feature representation. This design effectively solves the problem of sudden action changes caused by the discretization of environmental states in traditional methods, and greatly enhances the continuity and smoothness of control commands. Perceptual channel: Convolutional Neural Network (CNN), input 32×32 pixel point cloud raster image + 224×224 visual features, 5 convolutional layers (3×3 convolutional kernels per layer, ReLU activation).

[0030] Temporal channel: LSTM network with 128 hidden units, capturing obstacle motion trajectories in the first 500ms. Priority experience replay is used (priority weights α=0.6, β=0.4), with 100,000 training iterations. The convergence speed is improved by 40% compared to the standard PPO, and the average reward value is increased from -500 to +800.

[0031] The reward function design adopts a tiered architecture, progressively refining from basic path following and active environmental exploration to safety risk elimination. The basic layer reward propels the UAV to fly along the global reference path, the mid-level reward incentivizes reasonable exploration of unknown areas to expand navigation boundaries, and the high-level penalty mechanism uses a dynamic potential energy field to quantify collision risks and ensure priority in emergency obstacle avoidance. To address the problem of policy convergence under sparse reward conditions, a priority experience playback method is introduced, dynamically adjusting the weights of training samples according to action value and risk level, prioritizing the learning of high-value decision fragments, and enhancing the efficiency of policy iteration.

[0032] In terms of training strategy optimization, a heterogeneous simulation environment cluster is constructed to simulate complex scenarios with different obstacle densities, movement speeds, and sensor noise levels, thereby improving the overall generalization ability of the algorithm. By using a transfer learning mechanism, the model parameters trained in the simulation environment are transferred to the physical platform. Combined with online adaptive fine-tuning measures, the perception-control deviation between simulation and reality is dynamically filled. In scenarios with dynamic obstacles and random movement, the algorithm can achieve sub-meter-level path tracking accuracy and millisecond-level real-time response capability. Compared with traditional methods, it significantly reduces the frequency of emergency braking and energy consumption, providing a highly robust solution for autonomous navigation in complex environments.

[0033] In complex and dynamic environments, UAV obstacle avoidance and path planning need to achieve synergy across multiple performance metrics.

[0034] The core optimization objectives are: (1) Path movement efficiency, that is, to minimize the distance and time of travel while ensuring safety; (2) Economic efficiency of energy consumption: smooth speed control is adopted to reduce the fluctuation of motor power and enhance the range; (3) Safety level: dynamically adjust the minimum distance to obstacles to eliminate the risk of collision with sudden moving targets.

[0035] The three factors must be weighed in real time through a dynamic weighting mechanism. For example, in a narrow area, priority is given to ensuring a safe distance, while in an open airspace, the focus is on optimizing the smoothness of the path. The constraints are mainly divided into two categories: environment and body. Environmental constraints involve the uncertainty of the movement of dynamic obstacles, and the boundary of obstacle avoidance needs to be corrected in real time using a trajectory prediction model. The positioning deviation caused by sensor noise also needs to be considered. Body constraints include the maximum speed and angular velocity limits of the UAV, as well as the delay in the response of the power system.

[0036] The planning method also includes simulation verification steps, namely: S1. Scene Setup: Built using the Unity3D engine; the core area is a 30m × 30m × 20m cube; the coordinate system axis range is set as X: [-50, 90], Y: [-30, 70], Z: [-20, 50] Unit: m, spatial parameters are quantitatively characterized by discretization of the grid, such as a resolution of 0.5 m; S2. Scene setup: The initial position of the drone is set at the origin (0,0,0), and the global target point is located at (80,60,30). Path planning needs to achieve safe navigation in a dynamic trajectory interweaving environment. Four types of dynamic obstacles are set, and their movement paths are represented by red, yellow, blue and green trajectories respectively. S3. Multimodal Data Generation and Visualization: The experiment uses a physics engine to generate LiDAR point clouds, visual images, and IMU data streams in real time. The LiDAR uses a 16-line scanning configuration with a horizontal angular resolution of 0.2°, a vertical angular resolution of 2°, and a maximum detection distance of 50m. The point cloud density is dynamically adjusted according to the distance to obstacles, ranging from 5 to 100 points / m². The visual sensor outputs 1280×720 resolution images at a frequency of 30Hz, and introduces motion blur (blur kernel size 5×5) and illumination noise (SNR=25dB) to simulate real imaging defects. The 3D trajectory data is generated by B-spline curve interpolation. The dynamic obstacle position error follows a Gaussian distribution with a mean of zero and a standard deviation of 0.15m. The velocity fluctuation range is precisely adjusted by a parametric controller. The spatiotemporal state of all moving entities in the scene is synchronized with a 0.01s timestamp to ensure that the time alignment accuracy of multi-source data is better than 10ms. S4. Scenario validity verification: such as Figure 2As shown, the dynamic complexity of the scenario is quantitatively evaluated by comparing the actual trajectory of the UAV with the theoretical optimal path. The trajectory deviation δ is defined, and the average deviation measured experimentally is 0.82m (the baseline value for the static scenario is 0.15m), verifying the strong interference characteristics of dynamic obstacles. Furthermore, Fréchet distance analysis is used to analyze trajectory similarity. In the dynamic scenario, this index reaches 4.73m, an improvement of 316% compared to the static scenario, indicating that the designed scenario can effectively reproduce motion conflicts in complex environments. In addition, the obstacle trajectory intersection frequency of 1.2 times / s and the UAV obstacle avoidance action trigger rate of 0.9 times / s show a high positive correlation, with a Pearson coefficient of 0.87, proving that the scenario dynamics model can accurately reflect the spatiotemporal coupling characteristics of real obstacle avoidance tasks.

[0037] In step S2, the red trajectory corresponds to uniform linear motion with a speed of 2.5m / s ± 0.3m / s, the yellow trajectory is a random disturbance mode with a speed standard deviation of 1.2m / s, and the blue-green dual-color trajectory simulates cooperative avoidance behavior with a relative distance constraint of ≥3m.

[0038] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely prisms of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles (UAVs), characterized by: include: The perception simulation module uses multi-sensor joint modeling to achieve dynamic reconstruction of the environment; The decision control and dynamics modeling module acquires fused sensing data and generates a series of control commands. The scene interaction module enables dynamic obstacle behavior modeling and multi-scene library calls.

2. The deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles according to claim 1, characterized in that: The perception simulation module, decision control and dynamics modeling module, and scene interaction module are integrated with a ROS-MATLAB joint interface platform to achieve real-time synchronization and visualization analysis of multi-source heterogeneous data.

3. The deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles according to claim 1, characterized in that: The perception simulation module includes a lidar, a visual sensor, an ultrasonic sensor, an inertial measurement unit for providing attitude and acceleration information of the UAV, and a multimodal database processed by a fusion algorithm.

4. The deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles according to claim 1, characterized in that: The decision control and dynamics modeling module includes a PID controller that serves as the underlying flight control system to perform trajectory tracking tasks, an outer loop that generates attitude commands according to the Lyapunov stability criterion, and an inner loop that uses quaternion calculations to drive the motor speed.

5. The deep learning-based autonomous obstacle avoidance and path planning system for unmanned aerial vehicles according to claim 1, characterized in that: The scenario library includes three typical behaviors: random walking, cooperative avoidance, and sudden intrusion.

6. The planning method for the autonomous obstacle avoidance and path planning system for unmanned aerial vehicles based on deep learning, as described in any one of claims 1 to 5, is characterized in that: The specific steps are as follows: Step 1: Construct a mathematical model. Relying on the UAV kinematic model, define the state space as a combination of position, heading angle and control commands. Use differential equations to characterize the motion trajectory. Given the uncertainty of dynamic obstacles, use LSTM network and Kalman filter to jointly predict the time-varying trajectory of obstacles. Then, combine the noise distribution model corresponding to multiple sensors to set dynamic safety distance constraints. Step 2: At the path optimization level, a multi-objective cost function is established, with path length, energy consumption, and safety listed as game variables. A dynamic weight adjustment method is adopted to adjust the priority of each objective in real time. Step 3: Using a deep reinforcement learning framework, environmental perception information is mapped into a state space. Speed ​​control commands are generated through a hierarchical decision network. The reward function is designed to comprehensively consider path tracking accuracy, energy efficiency, and collision penalties. The generalization ability of the model is improved by combining offline training with online fine-tuning.

7. The planning method according to claim 6, characterized in that: In step three, a feature extraction mode based on spatiotemporal alignment is designed: LiDAR data generates a 3D occupancy grid map using voxelized mesh, visual data uses a lightweight YOLOv5 network to identify the contours of dynamic obstacles, and IMU data uses Kalman filtering to compensate for its attitude drift. Timestamp synchronization and coordinate system calibration are used to eliminate spatiotemporal deviations between multiple sensors and construct a unified environmental representation.

8. The planning method according to claim 6, characterized in that: In step three, a deep reinforcement learning framework based on near-end policy optimization is used to enhance the comprehensive capabilities of obstacle avoidance and path planning through multi-dimensional algorithm optimization: the perception channel uses a convolutional network to extract spatial semantic information from the multimodal fusion data, and the temporal channel uses a recurrent neural network to capture the motion state of dynamic obstacles. The two are then fused after attention weighting to generate a joint feature representation.

9. The planning method according to claim 6, characterized in that: It also includes the methods and steps for simulation verification.

10. The planning method according to claim 9, characterized in that: The steps of the simulation verification method are as follows: S1. Scene Setup: Built using the Unity3D engine; the core area is a 30m×30m×20m cube space; the coordinate system axis range is set to X: [-50,90], Y: [-30,70], Z: [-20,50], unit: m; spatial parameters are quantitatively represented by discretized mesh division. S2. Scene setup: The initial position of the drone is set at the origin (0,0,0); the global target point is located at (80,60,30); path planning needs to achieve safe navigation in a dynamic trajectory interweaving environment. Four types of dynamic obstacles are set, and their movement paths are represented by red, yellow, blue and green trajectories respectively. S3. Multimodal data generation and visualization: The experiment generates LiDAR point clouds, visual images and IMU data streams in real time through a physics engine; the point cloud density is dynamically adjusted with the distance to obstacles; the visual sensor outputs 1280×720 resolution images at a frequency of 30Hz; motion blur and illumination noise are introduced to simulate real imaging defects; the spatiotemporal state of all moving entities in the scene is synchronized with a 0.01s timestamp. S4. Scene Validity Verification: By comparing the actual trajectory of the UAV with the theoretical optimal path, the dynamic complexity of the scene is quantitatively evaluated. The static scene baseline is 0.15m. The strong interference characteristics of dynamic obstacles are verified, and the Fréchet distance is used to analyze the trajectory similarity.

Images