A deep learning-based unmanned aerial vehicle low-altitude collision avoidance early warning and path planning system

Abstract

The application discloses a kind of based on deep learning's unmanned plane low altitude anti-collision early warning and path planning system, it is related to unmanned plane technical field, including: perception module, based on sensor component real-time acquisition unmanned plane flight process in the three-dimensional space in the periphery of multi-dimensional environment data, after pre-processing transmission to the signal receiving end of depth learning early warning module;Depth learning early warning module, receive the multi-dimensional environment data, based on YOLOv8 target detection model to the multi-dimensional environment data carries out multi-source fusion identification processing, determine whether the three-dimensional space in the periphery of unmanned plane exists barrier target.The execution module of the application is based on PID control algorithm, cooperates with the regulation and control unmanned plane power, steering component, ensure smooth flight, compared with traditional scheme, the system has greatly improved barrier identification precision, early warning response speed and path rationality, adapts to many low altitude operation scene, reduces collision risk and operation cost, improves the intelligent level of flight.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, specifically to a deep learning-based UAV low-altitude collision avoidance warning and path planning system. Background Technology

[0002] With the rapid iteration and popularization of drone technology, its applications in various fields such as aerial surveying and mapping, power line inspection, emergency rescue, agricultural plant protection, and logistics distribution are becoming increasingly widespread. Low-altitude flight scenarios (usually referring to airspace below 1000 meters above sea level) have become the core scenarios for drone operations. However, the low-altitude flight environment is complex and changeable, with various static and dynamic obstacles such as buildings, trees, cables, birds, and other aircraft. Some of these obstacles are small, move quickly, and are irregularly distributed, posing a serious challenge to drone flight safety. Cases of drone collisions and crashes, equipment damage, operation interruptions, and even safety accidents occur frequently, severely restricting the large-scale and standardized development of the drone industry. Currently, existing UAV low-altitude collision avoidance and path planning technologies still have many shortcomings, making it difficult to meet the high-precision, high-real-time, and high-reliability flight requirements in complex low-altitude environments. Traditional technologies mostly use single sensors (such as ultrasonic sensors and infrared sensors) to collect environmental data, which has defects such as limited sensing range, single data dimension, and weak anti-interference ability. It is difficult to comprehensively capture obstacle information in the surrounding three-dimensional space, especially in scenarios with complex obstruction, changing light, and bad weather, which easily leads to the problem of missed or false detection of obstacles. At the same time, there is a lack of accurate identification and distance calculation capabilities for obstacle targets, and the early warning response is delayed, often failing to allow sufficient time for UAV obstacle avoidance operations, resulting in poor collision avoidance warning effects. It is also unable to cope with dynamic obstacles that suddenly appear during flight and lacks dynamic adaptability. To address these issues, we propose a deep learning-based UAV low-altitude collision avoidance warning and path planning system. Summary of the Invention

[0003] To address the aforementioned technical problems, a deep learning-based low-altitude collision avoidance warning and path planning system for unmanned aerial vehicles (UAVs) is provided. This technical solution resolves the aforementioned technical issues.

[0004] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A deep learning-based low-altitude collision avoidance warning and path planning system for unmanned aerial vehicles (UAVs) includes: The perception module collects multi-dimensional environmental data in the surrounding three-dimensional space during the flight of the drone in real time based on sensor components, and transmits the pre-processed data to the signal receiving end of the deep learning early warning module. The deep learning early warning module receives the multi-dimensional environmental data, performs multi-source fusion recognition processing on the multi-dimensional environmental data based on the YOLOv8 target detection model, and determines whether there are obstacle targets in the three-dimensional space around the drone. At the same time, through the built-in distance measurement unit, it calculates the real-time distance between the obstacle target and the drone body based on the preset ranging algorithm. When the real-time distance is less than the preset safety threshold, it determines that there is a flight safety hazard and immediately sends a standardized obstacle avoidance operation command to the execution module. The path dynamic planning module uses the SAC reinforcement learning algorithm, which combines the real-time positioning coordinates of the UAV, the coordinates of the obstacle target, and the preset flight destination coordinates to perform dynamic replanning calculations on the original flight path. Through multi-objective optimization, the optimal flight path that meets the dual requirements of flight efficiency and obstacle avoidance safety is obtained, and the optimal flight path parameters are transmitted to the execution module. The execution module receives the optimal flight path and, based on the PID control algorithm, coordinates and regulates the power and steering components of the UAV to execute corresponding obstacle avoidance operations, ensuring that the UAV flies smoothly and safely along the optimal flight path.

[0005] Preferably, the sensor components within the perception module include a binocular camera, a lidar, and an ultrasonic sensor. When the UAV is performing a flight mission, the sensor components collect multi-dimensional environmental data of the three-dimensional space surrounding the UAV in real time according to a preset acquisition frequency. The multi-dimensional environmental data includes obstacle contour data, raw distance data, and spatial coordinate data. The acquired multi-dimensional environmental data is preprocessed, including noise reduction, filtering correction, and normalization.

[0006] Preferably, the multi-source fusion identification processing steps within the deep learning early warning module are as follows: Start the YOLOv8 object detection model and initialize its configuration; train the model using obstacle sample data from drone flight scenarios to obtain weight values; import preset obstacles into the detection model for recognition and weight training; set the model's recognition threshold, confidence threshold, and fusion weight parameters. Feature extraction is performed on multidimensional environmental data to obtain feature vectors for various types of data, resulting in a standardized feature set. The standardized feature set is input into the initialized YOLOv8 target detection model. The multi-source feature vectors are weighted and fused through the feature fusion layer. Combined with the preset fusion weights, the features of different types of data are fused to construct a comprehensive feature map of the three-dimensional spatial environment around the UAV. When determining whether there are obstacle targets in the three-dimensional space around the drone, the system is based on a comprehensive feature map. Suspected obstacle targets in the map are identified and classified, and targets with a confidence level higher than a preset confidence threshold are preliminarily identified as potential obstacle targets.

[0007] Preferably, the steps for determining a flight safety hazard within the deep learning early warning module are as follows: The distance measurement unit is based on laser ranging fusion with visual ranging algorithm. It uses the center of gravity of the UAV body as the measurement reference point and the geometric center of the obstacle target as the target reference point. Combined with the synchronously imported real-time coordinates of the UAV and the coordinates of the obstacle target, it calculates the straight-line distance between the two reference points in real time through the three-dimensional spatial distance calculation formula, and obtains the distance between the obstacle target and the UAV body. The calculated spacing data is used for safety assessment. A preset safety threshold is called, and the spacing is compared with the preset safety threshold to determine the current spacing's safety level range. If the distance is greater than or equal to the preset safety threshold, it is determined that there is no flight safety hazard; if the distance is less than the preset safety threshold, it is determined that there is a flight safety hazard, and an obstacle avoidance operation command is sent.

[0008] Preferably, the three-dimensional spatial distance calculation formula uses the center of gravity of the UAV fuselage as the measurement reference point. The geometric center of the obstacle target is used as the target reference point. Combining the real-time coordinates of both points in the three-dimensional coordinate system, the real-time distance is calculated using the formula for the straight-line distance between two points in three-dimensional space. The specific formula is as follows: in The distance between the obstacle target and the drone's fuselage; Reference point for the center of gravity of the drone fuselage Real-time 3D coordinates; Reference point for the geometric center of the obstacle target Real-time 3D coordinates; in the formula , , The sum of the squared differences between the two reference points along the x, y, and z axes of the three-dimensional coordinate system is used to calculate the straight-line distance between the two reference points, i.e., the real-time spacing. ; Spacing is compared with the safety threshold by calling the preset safety threshold. The calculated real-time spacing and By comparing the data, the safety level range of the current distance is determined, and the specific level classification is as follows: No safety hazards: When At that time, it was determined that there were no safety hazards in the current flight environment of the drone; There is a safety hazard: When If the system determines that the drone poses a flight safety hazard, it immediately triggers the obstacle avoidance command generation process.

[0009] Preferably, the path dynamic planning module analyzes the original flight path currently being executed by the UAV, extracts the coordinates of the path points, flight trajectory parameters and preset flight speed of the original path, and clarifies the spatial relationship between the original path and the obstacle target; The original flight path currently being executed by the UAV is analyzed to extract the coordinates of the key path points, flight trajectory parameters and preset flight speed of the original path, and to clarify the spatial relationship between the original path and the obstacle target; based on the coordinates of the obstacle target and the preset obstacle avoidance safety distance threshold, the no-fly zone around the obstacle is marked. Based on the SAC reinforcement learning algorithm, a three-dimensional spatial flight environment model is constructed with the real-time coordinates of the UAV as the starting point, the preset flight endpoint coordinates as the endpoint, and the no-fly zone as the constraint. Flight efficiency and obstacle avoidance safety are taken as dual optimization objectives and embedded in the reward function design of the environment model. The algorithm generates multiple path candidate schemes through the Actor network, forming a path candidate set.

[0010] Preferably, the SAC reinforcement learning algorithm is based on the current flight environment state. Input, output path action distribution Multiple path actions are generated based on the action distribution sampling. Each path action corresponds to a complete path candidate solution, and the flight environment status... The formula to be entered is: in The preset flight speed for the original drone flight path; Preset endpoint coordinates; Actor network output path action distribution The expression is: in For Actor network parameters, output path action The probability distribution, This is the mean vector of the path actions. The standard deviation vector of the path actions. The identity matrix is ​​used to ensure the standardization of action distribution. N is the preset number of path candidate schemes, with a value range of 5 to 15, which is adaptively adjusted based on the complexity of the flight scenario. Each path action The core parameters corresponding to a candidate path include the path point sequence, flight trajectory slope, and flight speed adjustment coefficient, expressed as: in The coordinates of the critical path points for the candidate path solutions. The number of path points is adaptively adjusted based on the path length. The slope of the flight trajectory between adjacent path points is used to characterize the smoothness of the path; Path-based actions Generate complete path candidate solutions The formula is expressed as: in, A path generation function based on path actions. and 3D flight environment model It generates candidate path schemes containing a complete sequence of path point coordinates and flight trajectory equations.

[0011] Preferably, the optimal flight path in the path dynamic planning module has the dual objectives of maximizing flight efficiency and optimizing obstacle avoidance safety. Combining the preset flight efficiency threshold and obstacle avoidance safety weight, a comprehensive score is calculated for each valid candidate path; the candidate path with the highest comprehensive score is selected as the optimal flight path; and the obtained optimal flight path is input into the execution module.

[0012] Preferably, the PID control algorithm is started within the execution module, the parameters of the PID controller are initialized and configured, the preset proportional coefficient, integral coefficient, and derivative coefficient are imported, the trajectory parameters of the optimal flight path are used as the control target, and the real-time flight status parameters of the UAV are used as the feedback signal to construct a closed-loop control system.

[0013] Preferably, based on the PID control algorithm, the power and steering components of the UAV are coordinated and controlled. For the power component, the PID algorithm is used to adjust the speed of each rotor motor and the thrust of the thruster to control the flight speed and altitude of the UAV. For the steering component, the PID algorithm is used to adjust the deflection angle of the rudder and the attitude parameters of the attitude adjuster to control the flight direction and body attitude of the UAV, avoid obstacles and targets, and fly based on the optimal flight path.

[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention integrates multiple advanced algorithms and modules to achieve safe and efficient management of UAVs flying at low altitudes. The perception module collects and preprocesses surrounding 3D environmental data through sensors, then transmits it to the deep learning early warning module. This module uses the YOLOv8 model to identify obstacles and calculates real-time distances using the ranging unit. When a potential hazard is detected, an obstacle avoidance command is immediately sent. The path dynamic planning module uses the SAC reinforcement learning algorithm, integrating real-time positioning, obstacle coordinates, and endpoint coordinates to dynamically replan the optimal path that balances efficiency and safety. The execution module, based on the PID control algorithm, coordinates and regulates the UAV's power and steering components to ensure stable flight. Compared to traditional solutions, this system significantly improves obstacle recognition accuracy, early warning response speed, and path rationality, adapts to various low-altitude operation scenarios, reduces collision risks and operating costs, and enhances the level of flight intelligence. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0016] Figure 1 This is a schematic diagram of the system framework of the present invention; Figure 2 This is a flowchart of the multi-source fusion recognition processing steps within the deep learning early warning module of this invention; Figure 3 This is a flowchart of the steps for determining potential flight safety hazards within the deep learning early warning module of this invention. Detailed Implementation

[0017] 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. 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.

[0018] It should be understood that, when used in this specification, the terms “comprising” and “including” indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0019] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” are intended to include the plural forms.

[0020] It should also be further understood that the term "and / or" as used in this specification refers to any combination of one or more of the associated listed items, as well as all possible combinations, and includes such combinations.

[0021] Reference Figure 1 As shown, a deep learning-based low-altitude collision avoidance warning and path planning system for unmanned aerial vehicles includes: The perception module collects multi-dimensional environmental data in the surrounding three-dimensional space during the flight of the drone in real time based on sensor components, and transmits the pre-processed data to the signal receiving end of the deep learning early warning module. The deep learning early warning module receives the multi-dimensional environmental data, performs multi-source fusion recognition processing on the multi-dimensional environmental data based on the YOLOv8 target detection model, and determines whether there are obstacle targets in the three-dimensional space around the drone. At the same time, through the built-in distance measurement unit, it calculates the real-time distance between the obstacle target and the drone body based on the preset ranging algorithm. When the real-time distance is less than the preset safety threshold, it determines that there is a flight safety hazard and immediately sends a standardized obstacle avoidance operation command to the execution module. The path dynamic planning module uses the SAC reinforcement learning algorithm, which combines the real-time positioning coordinates of the UAV, the coordinates of the obstacle target, and the preset flight destination coordinates to perform dynamic replanning calculations on the original flight path. Through multi-objective optimization, the optimal flight path that meets the dual requirements of flight efficiency and obstacle avoidance safety is obtained, and the optimal flight path parameters are transmitted to the execution module. The execution module receives the optimal flight path and, based on the PID control algorithm, coordinates and regulates the power and steering components of the UAV to execute corresponding obstacle avoidance operations, ensuring that the UAV flies smoothly and safely along the optimal flight path.

[0022] This application's deep learning early warning module integrates target recognition and distance calculation. While completing obstacle recognition, it calculates the real-time distance to obstacles in real time through the built-in ranging unit. This eliminates the need for cross-system interaction between multiple modules, significantly reducing the time delay of early warning judgment and meeting the millisecond-level response requirements of UAVs flying at low altitudes and high speeds. The quantitative early warning judgment based on preset safety thresholds replaces the traditional fuzzy obstacle avoidance triggering logic. The triggering conditions for collision avoidance warnings are clear and controllable. The threshold parameters can be flexibly adjusted according to different low-altitude operation scenarios, making it more adaptable and effectively avoiding the problems of ineffective obstacle avoidance or untimely obstacle avoidance. It breaks through the limitations of traditional planning algorithms with a single objective. While ensuring that the drone stays away from obstacles, it minimizes the flight path after replanning, avoiding the problems of increased energy consumption and extended operation time caused by excessively detours, and taking into account both the operational performance and safety performance of the drone. Precise collision avoidance warnings and intelligent path planning can effectively reduce drone damage caused by collisions, as well as the cost of replacing and repairing parts. At the same time, it can reduce the risk of operation interruption due to untimely obstacle avoidance and improve the continuity of drone operations.

[0023] The sensor components within the perception module include a binocular camera, a lidar, and an ultrasonic sensor. When the drone is performing a flight mission, the sensor components collect multi-dimensional environmental data of the three-dimensional space surrounding the drone in real time according to a preset acquisition frequency. The multi-dimensional environmental data includes obstacle contour data, raw distance data, and spatial coordinate data. The acquired multi-dimensional environmental data is preprocessed, including noise reduction, filtering correction, and normalization.

[0024] This application's perception module employs a combination of binocular cameras, LiDAR, and ultrasonic sensors. These three components work synergistically and complement each other. The binocular cameras capture obstacle outlines, the LiDAR achieves precise ranging, and the ultrasonic sensors fill in blind spots in near-range perception. They collect obstacle outlines, distance data, and multi-dimensional spatial coordinate information at preset frequencies. After noise reduction, filtering correction, and normalization preprocessing, redundant noise is removed, and the data format is standardized, ensuring accurate and standardized data transmitted to subsequent modules, thus improving the accuracy and reliability of obstacle recognition from the source. The deep learning early warning module uses the YOLOv8 model to identify obstacles and measure distances in real time, quickly sending obstacle avoidance commands when a hazard is triggered. The path planning module uses the SAC reinforcement learning algorithm to dynamically generate the optimal path that balances efficiency and safety. The execution module relies on PID control for precise regulation. The system is adaptable to various low-altitude operation scenarios, effectively reducing collision risks and operating costs, and significantly improving the safety, intelligence, and efficiency of UAV flight.

[0025] Reference Figure 2 As shown, the multi-source fusion identification processing steps within the deep learning early warning module are as follows: Start the YOLOv8 target detection model and perform comprehensive initialization configuration. Combine the actual operation scenarios of UAV low-altitude flight (such as urban streets, farmland and forest areas, park inspection, etc.) and set the core initialization parameters such as input resolution, batch size, and number of iterations to ensure that the model can stably adapt to the input specifications of UAV multi-dimensional environmental data after startup. At the same time, optimize the model running speed to meet the real-time detection requirements during low-altitude flight and avoid model lag and recognition delay due to unreasonable initialization parameters. To ensure the model's ability to accurately identify obstacles around the drone, model weight training was conducted. This involved two main steps: First, collecting obstacle sample data from various drone flight scenarios, including static obstacles (building walls, trees, utility poles, ground protrusions, etc.) and dynamic obstacles (birds, pedestrians, other low-altitude aircraft, moving vehicles, etc.). Samples were also collected under different lighting conditions and weather environments (overcast skies, light winds, light haze) to ensure data diversity and comprehensiveness, reducing the risk of model overfitting. Second, the collected obstacle sample data was labeled and classified, including obstacle type, contour coordinates, and size parameters. The labeled sample data was then input into the YOLOv8 model for iterative training. The backpropagation algorithm continuously adjusted the model's internal parameters, gradually optimizing its recognition accuracy, ultimately yielding the optimal model weight values ​​suitable for low-altitude drone scenarios. Furthermore, pre-set standard obstacle samples were imported into the detection model for targeted training, further enhancing the model's sensitivity to common obstacles and optimizing weight parameters to ensure the model can quickly match various obstacle targets encountered in actual flight. Feature extraction processing is performed on the multidimensional environmental data (obstacle contour data, raw distance data, and spatial coordinate data) transmitted from the perception module. Using the feature extraction network built into the YOLOv8 model, targeted feature mining is conducted on each type of environmental data to extract the core feature vectors for each data type. For example, contour edge and shape feature vectors are extracted from obstacle contour data, distance change and gradient feature vectors are extracted from raw distance data, and position and orientation feature vectors are extracted from spatial coordinate data. Subsequently, the extracted feature vectors are standardized to eliminate dimensional differences and redundant information between different types of data, integrating them into a standardized feature set to ensure the standardization and uniformity of the feature data, laying the foundation for subsequent feature fusion and obstacle recognition. The standardized feature set is input into the initialized YOLOv8 target detection model. The multi-source feature vectors are weighted and fused through the feature fusion layer. Combined with the preset fusion weights, the features of different types of data are fused to construct a comprehensive feature map of the three-dimensional spatial environment around the UAV. When determining whether there are obstacles in the three-dimensional space surrounding the drone, a comprehensive feature map is used. The target recognition and classification module of the YOLOv8 model identifies and classifies suspected obstacles one by one. The model matches and compares the suspected targets in the comprehensive feature map with the trained standard features of obstacles, analyzes the feature similarity and contour matching degree of the suspected targets, and filters out targets with confidence scores higher than the preset confidence threshold, which are initially identified as potential obstacles. For suspected targets with confidence scores lower than the preset confidence threshold, the model performs a second verification to eliminate false detection interference, ensuring that the initially identified potential obstacles have high reliability. This provides accurate target basis for subsequent distance calculation and safety warning, ensuring the safety of drones flying at low altitudes from the source.

[0026] Reference Figure 3 As shown, the steps for determining a flight safety hazard within the deep learning early warning module are as follows: The distance measurement unit is based on laser ranging fusion with visual ranging algorithm. It uses the center of gravity of the UAV body as the measurement reference point and the geometric center of the obstacle target as the target reference point. Combined with the synchronously imported real-time coordinates of the UAV and the coordinates of the obstacle target, it calculates the straight-line distance between the two reference points in real time through the three-dimensional spatial distance calculation formula, and obtains the distance between the obstacle target and the UAV body. The calculated spacing data is used for safety assessment. A preset safety threshold is called, and the spacing is compared with the preset safety threshold to determine the current spacing's safety level range. If the distance is greater than or equal to the preset safety threshold, it is determined that there is no flight safety hazard; if the distance is less than the preset safety threshold, it is determined that there is a flight safety hazard, and an obstacle avoidance operation command is sent.

[0027] The formula for calculating three-dimensional spatial distance uses the center of gravity of the UAV as the measurement reference point. The geometric center of the obstacle target is used as the target reference point. Combining the real-time coordinates of both points in the three-dimensional coordinate system, the real-time distance is calculated using the formula for the straight-line distance between two points in three-dimensional space. The specific formula is as follows: in The distance between the obstacle target and the drone's fuselage; Reference point for the center of gravity of the drone fuselage Real-time 3D coordinates; Reference point for the geometric center of the obstacle target Real-time 3D coordinates; in the formula , , The sum of the squared differences between the two reference points along the x, y, and z axes of the three-dimensional coordinate system is used to calculate the straight-line distance between the two reference points, i.e., the real-time spacing. ; Spacing is compared with the safety threshold by calling the preset safety threshold. The calculated real-time spacing and By comparing the data, the safety level range of the current distance is determined, and the specific level classification is as follows: No safety hazards: When At that time, it was determined that there were no safety hazards in the current flight environment of the drone; There is a safety hazard: When If the system determines that the drone poses a flight safety hazard, it immediately triggers the obstacle avoidance command generation process.

[0028] This application employs a hybrid algorithm combining laser ranging and visual ranging in its determination step, overcoming the limitations of a single ranging algorithm. Laser ranging boasts advantages such as high accuracy and strong anti-interference capabilities, enabling precise capture of distance information for distant obstacles. Visual ranging, on the other hand, supplements the detailed ranging of nearby obstacles. The fusion of these two methods significantly improves the accuracy and adaptability of real-time distance calculation. It can quickly process coordinate data transmitted from the sensing module and, combined with the parallel computing capabilities of the distance calculation unit, achieve real-time distance calculation, perfectly matching the millisecond-level early warning response requirements of UAVs flying at low altitudes and high speeds. This avoids untimely obstacle avoidance and flight accidents caused by excessively long ranging calculation times. When a flight safety hazard is detected, the system immediately triggers the obstacle avoidance command generation process, sending standardized obstacle avoidance commands directly to the execution module without additional intermediate verification steps. This ensures the timely triggering of commands and buys valuable time for the drone to quickly execute obstacle avoidance operations. When no safety hazard is detected, the system does not trigger unnecessary commands, ensuring that the drone flies smoothly along the original planned path and avoiding invalid commands from interfering with flight efficiency.

[0029] The path dynamic planning module analyzes the original flight path currently being executed by the UAV, extracts the coordinates of the path points, flight trajectory parameters and preset flight speed of the original path, and clarifies the spatial relationship between the original path and the obstacle target. The original flight path currently being executed by the UAV is analyzed to extract the coordinates of the key path points, flight trajectory parameters and preset flight speed of the original path, and to clarify the spatial relationship between the original path and the obstacle target; based on the coordinates of the obstacle target and the preset obstacle avoidance safety distance threshold, the no-fly zone around the obstacle is marked. Based on the SAC reinforcement learning algorithm, a three-dimensional spatial flight environment model is constructed with the real-time coordinates of the UAV as the starting point, the preset flight endpoint coordinates as the endpoint, and the no-fly zone as the constraint. Flight efficiency and obstacle avoidance safety are taken as dual optimization objectives and embedded in the reward function design of the environment model. The algorithm generates multiple path candidate schemes through the Actor network, forming a path candidate set.

[0030] The analysis operation is based on the preset planning file of the original path and uses a path point extraction algorithm to accurately extract the coordinates of all path points of the original path (including the coordinates of the starting point, the coordinates of the intermediate turning points, the coordinates of the nodes of each path segment, and the coordinates of the ending point). At the same time, the core flight trajectory parameters of the original path are extracted simultaneously, covering the trajectory curvature, path segment length, turning angle, and flight attitude requirements of each path segment, and clarifying the flight status standards of the UAV at each node of the original path. A 3D spatial modeling algorithm is used to construct a spherical no-fly zone with the geometric center of the obstacle as the origin and the obstacle avoidance safety distance threshold as the radius. If there are multiple obstacle targets, the no-fly zone corresponding to each obstacle is marked separately. At the same time, overlapping no-fly zones are merged to form a complete set of no-fly zones, and the 3D coordinate range and boundary contour of the no-fly zone are clearly marked. During model building, the focus is on optimizing the reward function design, with flight efficiency and obstacle avoidance safety as dual optimization objectives, embedded in the core logic of the reward function. Among them, the reward items corresponding to obstacle avoidance safety are judged by the distance between the drone's path and the no-fly zone and whether it touches the no-fly zone. The higher the reward value, the better. The reward items corresponding to flight efficiency are judged by the path length, flight time, and speed adaptability. The shorter the path, the less the flight time, and the closer the speed is to the original path's preset speed, the higher the reward value.

[0031] The SAC reinforcement learning algorithm is based on the current flight environment state. Input, output path action distribution Multiple path actions are generated based on the action distribution sampling. Each path action corresponds to a complete path candidate solution, and the flight environment status... The formula to be entered is: in The preset flight speed for the original drone flight path; Preset endpoint coordinates; Actor network output path action distribution The expression is: in For Actor network parameters, output path action The probability distribution, This is the mean vector of the path actions. The standard deviation vector of the path actions. The identity matrix is ​​used to ensure the standardization of action distribution. N is the preset number of path candidate schemes, with a value range of 5 to 15, which is adaptively adjusted based on the complexity of the flight scenario. Each path action The core parameters corresponding to a candidate path include the path point sequence, flight trajectory slope, and flight speed adjustment coefficient, expressed as: in The coordinates of the critical path points for the candidate path solutions. The number of path points is adaptively adjusted based on the path length. The slope of the flight trajectory between adjacent path points is used to characterize the smoothness of the path; Path-based actions Generate complete path candidate solutions The formula is expressed as: in, A path generation function based on path actions. and 3D flight environment model It generates candidate path schemes containing a complete sequence of path point coordinates and flight trajectory equations.

[0032] This application clarifies the planning direction and constraint boundaries by using obstacle and endpoint coordinates, providing comprehensive and accurate basic data support for subsequent action distribution generation and path planning, ensuring more targeted algorithm decisions; the identity matrix I effectively guarantees the standardization of action distribution and avoids action sampling deviations caused by parameter disorder; the preset number N (5~15) of path candidate schemes can be adaptively adjusted according to the complexity of the flight scenario, taking a larger value when obstacles are dense and the scenario is complex to ensure the diversity of candidate schemes to select the optimal solution, and taking a smaller value when the scenario is simple to reduce the amount of computation and improve planning efficiency, achieving a balance between "diversity and efficiency"; the SAC algorithm adopts standardized formula design from environmental state input and action distribution output to path action definition and path generation, and every parameter and every step of the calculation has a clear logical link and traceability.

[0033] The optimal flight path in the path dynamic planning module aims to maximize flight efficiency and optimize obstacle avoidance safety. It combines preset flight efficiency thresholds and obstacle avoidance safety weights to calculate a comprehensive score for each valid candidate path. The candidate path with the highest comprehensive score is selected as the optimal flight path, and the obtained optimal flight path is input into the execution module.

[0034] When selecting the optimal flight path, the dynamic path planning module of this application adopts a dual objective of maximizing flight efficiency and optimizing obstacle avoidance safety. It combines preset flight efficiency thresholds and obstacle avoidance safety weights to perform standardized comprehensive scoring on all effective path candidate schemes. The selection logic avoids the drawbacks of traditional single-objective selection. It ensures uniform standards and objective results through quantitative scoring. The obstacle avoidance safety weights and flight efficiency thresholds can be flexibly adjusted according to flight scenarios to adapt to different operational needs. Finally, the candidate path with the highest comprehensive score is selected as the optimal flight path. This path balances safety and efficiency, and has good smoothness and speed adaptability. It can be directly input into the execution module without intermediate parameter conversion, achieving seamless connection between the two modules, shortening control time, improving the overall operational efficiency, and ensuring that the UAV can adapt to its own performance and flight environment, flying smoothly and safely along the optimal path.

[0035] After receiving the optimal flight path parameters from the path dynamic planning module, the execution module immediately starts the built-in PID control algorithm and completes the parameter initialization configuration of the PID controller. Combining the operational scenario characteristics of the UAV's low-altitude flight, the airframe performance parameters, and the trajectory requirements of the optimal flight path, it accurately imports the preset proportional coefficient (P), integral coefficient (I), and derivative coefficient (D). These three types of coefficients have been debugged and optimized in multiple scenarios to adapt to the control requirements under different flight states. At the same time, with the trajectory parameters of the optimal flight path (including the path point coordinate sequence, flight trajectory equation, speed thresholds for each segment, and turning angle reference) as the core control target, and the flight status parameters (including current flight speed, airframe attitude angle, real-time three-dimensional coordinates, and rate of ascent and descent) collected in real time by the UAV flight control system as the closed-loop feedback signal, a full closed-loop control system of target setting, real-time feedback, deviation calculation, and dynamic adjustment is constructed to ensure that subsequent control actions always revolve around the optimal flight path and accurately track the path trajectory. Based on the aforementioned closed-loop control system, the PID control algorithm implements coordinated regulation of the UAV's power and steering components, achieving coordinated action and parameter matching between the two components. This avoids flight imbalance caused by controlling a single component, ensuring the smoothness of obstacle avoidance and path following operations. For the power component, the PID algorithm dynamically adjusts the output speed of each rotor motor by calculating the deviation between the UAV's current flight speed, altitude, and the preset parameters of the optimal path in real time. This achieves differentiated and precise control of the multi-rotor speed, while simultaneously regulating the thrust of the thrusters. Through coordinated matching of thrust and speed, it quickly corrects speed and altitude deviations, ensuring that the UAV's flight speed matches the speed thresholds of each segment of the optimal path, and that the altitude precisely matches the height changes required by the path, providing stable power support for path following. For the steering component, the PID algorithm adjusts the rudder deflection angle in real time based on the deviation between the UAV's current flight direction, fuselage attitude and optimal path steering parameters, and attitude reference. This precisely controls the UAV's horizontal steering direction and steering amplitude, while simultaneously adjusting the roll, pitch, and yaw attitude parameters of the attitude adjuster to quickly correct fuselage attitude deviations. This ensures the fuselage maintains a stable flight attitude, effectively avoiding obstacles while ensuring the UAV flies precisely along the optimal flight path. Throughout the entire control process, the parameters of the power and steering components are adjusted synchronously and in real time. Deviation data is continuously fed back to the PID controller, and the algorithm dynamically optimizes the control parameters based on the feedback deviations, forming a continuous closed-loop dynamic control until the UAV completes obstacle avoidance and flies stably along the optimal flight path.

[0036] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based low-altitude collision avoidance warning and path planning system for unmanned aerial vehicles (UAVs), characterized in that, include: The perception module collects multi-dimensional environmental data in the surrounding three-dimensional space during the flight of the drone in real time based on sensor components, and transmits the pre-processed data to the signal receiving end of the deep learning early warning module. The deep learning early warning module receives the multi-dimensional environmental data, performs multi-source fusion recognition processing on the multi-dimensional environmental data based on the YOLOv8 target detection model, and determines whether there are obstacle targets in the three-dimensional space around the drone. At the same time, through the built-in distance measurement unit, it calculates the real-time distance between the obstacle target and the drone body based on the preset ranging algorithm. When the real-time distance is less than the preset safety threshold, it determines that there is a flight safety hazard and immediately sends a standardized obstacle avoidance operation command to the execution module. The path dynamic planning module uses the SAC reinforcement learning algorithm, which combines the real-time positioning coordinates of the UAV, the coordinates of the obstacle target, and the preset flight destination coordinates to perform dynamic replanning calculations on the original flight path. Through multi-objective optimization, the optimal flight path that meets the dual requirements of flight efficiency and obstacle avoidance safety is obtained, and the optimal flight path parameters are transmitted to the execution module. The execution module receives the optimal flight path and, based on the PID control algorithm, coordinates and regulates the power and steering components of the UAV to execute corresponding obstacle avoidance operations, ensuring that the UAV flies smoothly and safely along the optimal flight path.

2. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that: The sensor components within the perception module include a binocular camera, a lidar sensor, and an ultrasonic sensor; When the drone is performing a flight mission, the sensor components collect multi-dimensional environmental data of the three-dimensional space around the drone in real time according to the preset acquisition frequency. The multi-dimensional environmental data includes obstacle outline data, raw distance data and spatial coordinate data. The acquired multi-dimensional environmental data is preprocessed, including noise reduction, filtering correction and normalization.

3. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that, The multi-source fusion identification processing steps within the deep learning early warning module are as follows: Start the YOLOv8 object detection model and initialize its configuration; train the model using obstacle sample data from drone flight scenarios to obtain weight values; import preset obstacles into the detection model for identification and weight training; Set the model's recognition threshold, confidence threshold, and fusion weight parameters; Feature extraction is performed on multidimensional environmental data to obtain feature vectors for various types of data, resulting in a standardized feature set. The standardized feature set is input into the initialized YOLOv8 target detection model. The multi-source feature vectors are weighted and fused through the feature fusion layer. Combined with the preset fusion weights, the features of different types of data are fused to construct a comprehensive feature map of the three-dimensional spatial environment around the UAV. When determining whether there are obstacle targets in the three-dimensional space around the drone, the system is based on a comprehensive feature map. Suspected obstacle targets in the map are identified and classified, and targets with a confidence level higher than a preset confidence threshold are preliminarily identified as potential obstacle targets.

4. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that, The steps for determining potential flight safety hazards within the deep learning-based early warning module are as follows: The distance measurement unit is based on a laser ranging fusion visual ranging algorithm. It uses the center of gravity of the UAV body as the measurement reference point and the geometric center of the obstacle target as the target reference point. Combining the synchronously imported real-time coordinates of the UAV and the coordinates of the obstacle target, it calculates the straight-line distance between the two reference points in real time through the three-dimensional spatial distance calculation formula, and obtains the distance between the obstacle target and the UAV body. The calculated spacing data is used for safety assessment. A preset safety threshold is called, and the spacing is compared with the preset safety threshold to determine the current spacing's safety level range. If the distance is greater than or equal to the preset safety threshold, it is determined that there is no flight safety hazard; if the distance is less than the preset safety threshold, it is determined that there is a flight safety hazard, and an obstacle avoidance operation command is sent.

5. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 4, characterized in that: The formula for calculating three-dimensional spatial distance uses the center of gravity of the UAV as the measurement reference point. The geometric center of the obstacle target is used as the target reference point. Combining the real-time coordinates of both points in the three-dimensional coordinate system, the real-time distance is calculated using the formula for the straight-line distance between two points in three-dimensional space. The specific formula is as follows: in The distance between the obstacle target and the drone's fuselage; Reference point for the center of gravity of the drone fuselage Real-time 3D coordinates; Reference point for the geometric center of the obstacle target Real-time 3D coordinates; in the formula , , The sum of the squared differences between the two reference points along the x, y, and z axes of the three-dimensional coordinate system is used to calculate the straight-line distance between the two reference points, i.e., the real-time spacing. ; Spacing is compared with the safety threshold by calling the preset safety threshold. The calculated real-time spacing and By comparing the data, the safety level range of the current distance is determined, and the specific level classification is as follows: No safety hazards: When At that time, it was determined that there were no safety hazards in the current flight environment of the drone; There is a safety hazard: When If the system determines that the drone poses a flight safety hazard, it immediately triggers the obstacle avoidance command generation process.

6. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that: The path dynamic planning module analyzes the original flight path currently being executed by the UAV, extracts the coordinates of the path points, flight trajectory parameters and preset flight speed of the original path, and clarifies the spatial relationship between the original path and the obstacle target. The original flight path currently being executed by the UAV is analyzed, and the coordinates of the key path points, flight trajectory parameters and preset flight speed of the original path are extracted to clarify the spatial relationship between the original path and the obstacle target. Based on the target coordinates of the obstacle and the preset obstacle avoidance safety distance threshold, the no-fly zone around the obstacle is marked; Based on the SAC reinforcement learning algorithm, a three-dimensional spatial flight environment model is constructed with the real-time coordinates of the UAV as the starting point, the preset flight endpoint coordinates as the endpoint, and the no-fly zone as the constraint. Flight efficiency and obstacle avoidance safety are taken as dual optimization objectives and embedded in the reward function design of the environment model. The algorithm generates multiple path candidate schemes through the Actor network, forming a path candidate set.

7. A deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 6, characterized in that: The SAC reinforcement learning algorithm is based on the current flight environment state. Input, output path action distribution Multiple path actions are generated based on the action distribution sampling. Each path action corresponds to a complete path candidate solution, and the flight environment status... The formula to be entered is: in The preset flight speed for the original drone flight path; Preset endpoint coordinates; Actor network output path action distribution The expression is: in For Actor network parameters, output path action The probability distribution, This is the mean vector of the path actions. The standard deviation vector of the path actions. The identity matrix is ​​used to ensure the standardization of action distribution. N is the preset number of path candidate schemes, with a value range of 5 to 15, which is adaptively adjusted based on the complexity of the flight scenario. Each path action The core parameters corresponding to a candidate path include the path point sequence, flight trajectory slope, and flight speed adjustment coefficient, expressed as: in The coordinates of the critical path points for the candidate path solutions. The number of path points is adaptively adjusted based on the path length. The slope of the flight trajectory between adjacent path points is used to characterize the smoothness of the path; Path-based actions Generate complete path candidate schemes The formula is expressed as: in, A path generation function based on path actions. and 3D flight environment model It generates candidate path schemes containing a complete sequence of path point coordinates and flight trajectory equations.

8. The deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that: The optimal flight path in the path dynamic planning module aims to maximize flight efficiency and optimize obstacle avoidance safety. It combines preset flight efficiency thresholds and obstacle avoidance safety weights to calculate a comprehensive score for each valid candidate path. The candidate path with the highest comprehensive score is selected as the optimal flight path, and the obtained optimal flight path is input into the execution module.

9. A deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 1, characterized in that: The execution module starts its PID control algorithm, initializes the PID controller parameters, imports preset proportional coefficients, integral coefficients, and derivative coefficients, uses the trajectory parameters of the optimal flight path as the control target, and uses the real-time flight status parameters of the UAV as the feedback signal to build a closed-loop control system.

10. A deep learning-based UAV low-altitude collision avoidance warning and path planning system according to claim 9, characterized in that: Based on the PID control algorithm, the power and steering components of the UAV are coordinated and controlled. For the power components, the speed of each rotor motor and the thrust of the thruster are adjusted by the PID algorithm to control the flight speed and altitude of the UAV. For the steering components, the PID algorithm is used to adjust the deflection angle of the rudder and the attitude parameters of the attitude regulator to control the UAV's flight direction and body attitude, avoid obstacles, and fly based on the optimal flight path.

Images