A low-altitude UAV traffic operation risk warning method and system

Abstract

The present invention proposes a low-altitude UAV traffic operation risk warning method and system, the method comprising: collecting UAV status data and external environment data, performing data preprocessing, and generating fused environmental perception data; inputting the fused environmental perception data into a lightweight multi-channel convolutional network, and outputting a local risk score tensor; generating a risk map sequence and a risk drift point sequence for future time periods through a dual-input gated recurrent neural network; intercepting an ROI region starting from the current position of the aircraft and extending with the current speed direction as a vector in the risk map sequence, extracting the average risk value of each frame in the ROI region to calculate the future regional risk level and generate a decision recommendation flag; generating executable flight control instructions based on the decision recommendation flag to drive the UAV to execute. The present invention significantly improves the real-time response capability and safety of low-altitude UAVs in complex flight environments.

Description

Technical Field

[0001] The present invention belongs to the field of low-altitude UAV traffic, and in particular relates to a low-altitude UAV traffic operation risk early warning method and system. Background Art

[0002] With the rapid development of drone technology, low-altitude drones have become a vital tool in diverse fields, including transportation, logistics, agriculture, and environmental monitoring. The widespread use of low-altitude drones has led to intense competition for airspace resources, particularly in densely populated areas like cities and industrial areas, where the risks of low-altitude flight are increasing. To ensure drone flight safety in complex environments, various low-altitude drone risk warning and collision avoidance systems have been proposed both domestically and internationally. These systems typically rely on real-time drone perception data, combined with traditional flight path planning and control methods, to achieve risk avoidance and flight safety management.

[0003] Existing low-altitude drone traffic management technologies primarily rely on two approaches: one is sensor-based environmental perception and localized path planning, using, for example, visual sensors, lidar, and infrared sensors to monitor the surrounding environment in real time and adjust flight trajectories using localized obstacle avoidance algorithms. The other approach involves airspace monitoring and traffic flow management for aircraft scheduling and risk warnings, typically based on static flight paths and fixed rules. However, existing technologies still have significant limitations when dealing with the highly dynamic environments faced by low-altitude drones. First, sensor-based risk warning systems lack the flexibility to adapt to rapidly changing flight environments. For example, existing obstacle avoidance algorithms mostly rely on simple static rules and are unable to respond to emergencies or complex airspace environments (such as complex airflow and temporary obstacles). Airspace management systems, on the other hand, rely on fixed flight paths and regular scheduling, failing to adjust and optimize routes in real time and failing to respond promptly to emergencies. These technologies cannot adequately address the dynamic risks encountered by drones during actual flight, such as high flight density, weather changes, and the presence of unexpected obstacles.

[0004] Furthermore, existing low-altitude drone risk warning technologies largely ignore the specific needs of different flight missions and airspace environments, failing to dynamically optimize based on real-time airspace changes and the drone's mission objectives. This results in delayed system responses in complex flight scenarios and even an inability to provide timely and effective warnings in high-risk situations. Overall, existing technologies are unable to provide comprehensive, flexible, and real-time risk warnings and behavioral adjustments in complex and changing low-altitude environments, and there remains significant room for improvement in adaptability, efficiency, and accuracy. Summary of the Invention

[0005] The purpose of this invention is to propose a low-altitude UAV traffic operation risk warning method and system, which is based on real-time environmental perception data and an efficient dynamic risk prediction model. It can flexibly respond to sudden changes in the low-altitude flight environment, adjust the flight path in real time and provide timely warnings.

[0006] In order to achieve the above objectives, a first aspect of the present invention provides a low-altitude UAV traffic operation risk warning method, the method comprising:

[0007] S1. Collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein the drone status data includes position, speed, and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar, and infrared sensors;

[0008] S2. Input the fused environmental perception data into a lightweight multi-channel convolutional network, combine it with the spatial attention module to enhance the regional significant risk features, and output a local risk score tensor, which represents the risk at time Corresponding risk level scores for spatial locations, dimensions and environmental conditions Figure 1 That is, the local risk score tensor is a risk score map, and the high-risk area coordinates are identified in the risk score map to construct a high-risk area coordinate set;

[0009] S3. Based on the risk score map and drone status data, a historical risk map sequence, a historical flight status sequence, and a historical sensor dynamic weight sequence are obtained. Combined with the current risk score map and the high-risk area coordinate set, a risk map sequence and a risk drift point sequence for future time periods are generated through a dual-input gated recurrent neural network.

[0010] S4. A ROI region is intercepted from the risk map sequence, starting from the current position of the aircraft and extending along the current speed direction as a vector. The average risk value of each frame in the ROI region is extracted to calculate the future regional risk level. The minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the period is calculated. The number of high-risk points entering the ROI region is counted to construct a high-risk point threat coefficient. The decision recommendation flag is generated based on the future regional risk level, the current speed of the aircraft, and the urgency of the mission.

[0011] S5. Generate an executable flight control instruction according to the decision suggestion flag to drive the UAV to execute.

[0012] Optionally, the structure of the lightweight multi-channel convolutional network is a three-layer convolutional structure, where the convolution kernel size of each layer is , the number of channels are 16, 32, and 64 respectively; among them, the first layer processes image data, the second layer processes the point cloud mapping of the lidar, and the third layer combines the thermal imaging channel of the infrared sensor to realize multi-modal feature joint encoding;

[0013] Among them, a spatial attention module is added after each convolutional layer to enhance the regional significant risk characteristics; and the risk score map is smoothed through spatial gradient regularization constraints to suppress noise interference.

[0014] Optionally, identifying high-risk area coordinates in the risk score map to construct a high-risk area coordinate set specifically includes:

[0015] The risk score map is used as a high-risk candidate area, and combined with the aircraft speed direction and predicted route, it is determined whether there is a high-risk intersection in the path; if so, its spatial coordinates and corresponding risk value are recorded as the current high-risk area coordinates.

[0016] Optionally, the dual-input gated recurrent neural network is composed of two parallel GRU modules, specifically:

[0017] The first GRU module processes the risk map sequence, and the second GRU module processes the multivariate vector sequence, outputs intermediate features, performs splicing and fusion, and then outputs a predicted risk map sequence for several future time steps through the fusion layer. The multivariate vector sequence is a combined flight status, sensor response, and high-risk point density sequence.

[0018] The first GRU module and the second GRU module are two groups of three-layer 128-unit GRU sub-networks.

[0019] Optionally, the step of generating a risk drift point sequence includes:

[0020] A current high-risk point set and a current horizontal velocity component of the aircraft are obtained, and drift mapping based on the current horizontal velocity component of the aircraft is performed on the current high-risk point set to generate a risk drift point sequence.

[0021] Optionally, the future regional risk level The calculation is as follows:

[0022]

[0023] Where T is the time window, For the The set of grid positions contained in the ROI area in the frame, is the risk prediction value of the point; (x, y) is the current horizontal velocity component of the aircraft;

[0024] The threat factor of the high-risk point , constructed as follows:

[0025]

[0026] in, It is an indicator function, which is used to count the high-risk points falling into the ROI area. The value of is the highest, indicating that the future flight path will have the highest degree of intersection with the high-risk area; is the current time window; is the current risk drift point sequence.

[0027] Optionally, the generating of a decision suggestion flag based on the future regional risk level, the current speed of the aircraft, and the urgency of the mission specifically includes:

[0028] The regular tree method is used for division, where Suggesting signs for decision making, For the urgency of the task:

[0029] like High ,but , indicating to continue flying;

[0030] like and ,but , it means to keep the current state and continue monitoring;

[0031] like or ,but:

[0032] like : , it is recommended to hover or wait;

[0033] like : , detour route planning is recommended.

[0034] Optionally, generating an executable flight control instruction according to the decision suggestion flag to drive the UAV to execute specifically includes:

[0035] like , the drone maintains the current velocity vector direction, sets the target acceleration to zero, and only makes speed maintenance and attitude stabilization adjustments;

[0036] like , then with Same as above, but with the high-frequency status monitoring flag turned on, the system enters high-frequency sampling mode and rechecks the risk level every 0.2 seconds.

[0037] like , then issue deceleration and altitude lock instructions; apply reverse acceleration in the current direction of motion to decay the velocity vector to zero, and activate the altitude lock module at the same time to keep the current altitude unchanged;

[0038] like , then the aircraft is offset by a certain angle from its current position away from the high-risk point, an avoidance heading vector is constructed, and the acceleration direction is adjusted.

[0039] Optionally, each generated control instruction will be encapsulated as a standard MAVLink control frame and sent to the flight control system for execution via the communication module; the control instruction includes acceleration and avoidance heading vector.

[0040] Another aspect of the present invention provides a low-altitude UAV traffic operation risk warning system, the system comprising:

[0041] Data acquisition module: used to collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein, the drone status data includes position, speed and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar and infrared sensors;

[0042] Risk identification module: used to input the fused environmental perception data into a lightweight multi-channel convolutional network, combined with the spatial attention module to enhance the regional significant risk features, and output a local risk score tensor, which represents the risk at time Corresponding to the risk level score of spatial location, dimension and environment Figure 1 That is, the local risk score tensor is a risk score map, and the high-risk area coordinates are identified in the risk score map to construct a high-risk area coordinate set;

[0043] Risk prediction module: used to obtain a historical risk map sequence, a historical flight state sequence, and a historical sensor dynamic weight sequence based on the risk score map and drone status data. Combined with the current risk score map and the high-risk area coordinate set, a risk map sequence and a risk drift point sequence for the future time period are generated through a dual-input gated recurrent neural network.

[0044] Decision support module: used to intercept an ROI area from the risk map sequence, starting from the current position of the aircraft and extending from the current speed direction as the vector, extract the average risk value of each frame in the ROI area to calculate the future regional risk level, and simultaneously calculate the minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the period, and count the number of high-risk points entering the ROI area to construct the high-risk point threat coefficient. In combination with the future regional risk level, the current speed of the aircraft and the urgency of the mission, a decision recommendation flag is generated;

[0045] Flight control module: used to generate executable flight control instructions according to the decision suggestion flag to drive the drone to execute.

[0046] The beneficial technical effects of the present invention are at least as follows:

[0047] To address the aforementioned issues in existing low-altitude drone risk warning systems, this paper proposes a novel low-altitude drone traffic operation risk warning method and system. Based on real-time environmental perception data and an efficient dynamic risk prediction model, this system can flexibly respond to sudden changes in the low-altitude flight environment, adjust flight paths in real time, and provide timely warnings. The key innovation of this invention lies in its integration of real-time environmental perception and dynamic airspace risk prediction, overcoming the limitations of traditional static path planning and local obstacle avoidance. First, by acquiring detailed environmental information from multiple sensor sources (such as vision, lidar, and infrared sensors), the system can accurately identify real-time environmental factors such as aircraft, obstacles, and weather changes. Furthermore, the system incorporates a dynamic airspace risk prediction module. Based on real-time changes in flight density, airspace status, and meteorological conditions, it continuously optimizes flight paths and decision-making strategies, providing early warnings of potential risks and preventing collisions in high-risk environments.

[0048] Unlike existing technologies, the present invention can predict changes in traffic density, aircraft behavior, weather impacts, etc. that may occur in the next few minutes to hours through deep fusion analysis of real-time sensor data and historical flight data, and combine these prediction results with real-time perception information to provide comprehensive risk assessment and real-time decision support. This innovation enables drones to not only make quick decisions in the current environment, but also make risk avoidance preparations in advance and optimize flight efficiency. Based on this dynamic adaptation and predictive optimization, the flight system can automatically adjust the flight path according to changes in the environment, improving the safety and mission execution efficiency of drones in complex airspace. Through this method, the present invention solves the limitations of the static and single early warning mechanism in the existing technology, and significantly improves the real-time response capability and safety of low-altitude drones in complex flight environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The present invention is further described with reference to the accompanying drawings. However, the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative effort.

[0050] Figure 1 This is a flow chart of a low-altitude UAV traffic operation risk warning method according to an embodiment of the present invention.

[0051] Figure 2This is a framework diagram of a low-altitude UAV traffic operation risk warning system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0053] like Figure 1 As shown, an embodiment of the present invention provides a low-altitude UAV traffic operation risk warning method, the method comprising:

[0054] S1. Collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein the drone status data includes position, speed, and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar, and infrared sensors.

[0055] In low-altitude flight environments, accurate environmental perception is paramount for ensuring safe drone flight. Environmental perception and data acquisition utilize a combination of multiple sensors (such as vision, lidar, and infrared sensors) to acquire real-time data about the aircraft's surroundings. This data provides essential input for subsequent risk identification, prediction, and flight path optimization. Especially during flight, where environmental changes can be rapid, the real-time and accurate nature of data acquisition and preprocessing is crucial.

[0056] The input data for this step includes the aircraft's status data and the perception data of the external environment. Specifically:

[0057] Aircraft status data:

[0058] Location( ): The real-time coordinates of the aircraft in the airspace, usually obtained through the GPS system.

[0059] speed( ): The speed of the aircraft in each axis direction can be obtained through the inertial navigation system (INS) or speed sensor.

[0060] Flight altitude ( ): The altitude of the aircraft relative to the ground or sea level, usually provided by a barometer or GPS system.

[0061] Perception data of the external environment:

[0062] Image data from vision cameras: This data is crucial for identifying obstacles, aircraft, and ground objects, especially in low-light conditions.

[0063] Depth map from LiDAR: used to calculate the distance between the aircraft and surrounding objects, as well as the precise location of obstacles.

[0064] Data from infrared sensors: Provides additional information in low-light or nighttime environments, helping to sense temperature differences and identify potential obstacles or aircraft.

[0065] The collected data will be transmitted to the data processing module in real time and integrated into unified environmental perception data after preprocessing.

[0066] Specifically, the steps include:

[0067] First, aircraft use a variety of sensors to perceive their environment. During low-altitude flight, the types and locations of obstacles can be extremely complex, requiring the aircraft to make decisions based on the surrounding environment at all times. Therefore, the diversity and real-time nature of sensor data are crucial.

[0068] Among them, the multi-sensor deployment in low-altitude flight environments is designed to perceive the surrounding environment and obtain aircraft status data in real time. To this end, the drone is equipped with the following sensors:

[0069] Visual cameras: These provide image information of the environment. During daytime or in good lighting conditions, visual sensors can effectively identify surrounding obstacles, aircraft, and ground targets. However, at night or in low-light environments, the effectiveness of cameras can be significantly reduced, and infrared sensors can provide a powerful supplement.

[0070] LiDAR: LiDAR is a crucial sensor for low-altitude flight. It accurately measures the distance between an aircraft and surrounding objects and generates a highly accurate three-dimensional image of the environment. LiDAR data can provide real-time support for flight path planning, especially in environments with dense obstacles.

[0071] Infrared sensor: When the aircraft is in special environments such as low light, at night, or in haze, the infrared sensor can sense temperature differences and identify possible obstacles.

[0072] Furthermore, after sensor data is collected, it needs to be preprocessed. This includes operations such as denoising, time synchronization, and data normalization. Denoising eliminates random errors and external interference in sensor output to ensure data quality. Time synchronization ensures that data from different sensors is time-aligned, allowing comparison and fusion of data from multiple sensors within the same time dimension. For example, the sampling frequencies of lidar and visual cameras may differ, so their data needs to be time-aligned. Data normalization unifies the scale of sensor data, allowing data from different sensor types to be processed within the same computing framework.

[0073] In practical applications, denoising techniques such as Kalman filtering and particle filtering are employed. Kalman filtering can be used to optimize the estimation of sensor data, particularly when subject to noise and dynamic changes, providing more stable data output. Particle filtering can be used to estimate uncertainty in complex dynamic systems, particularly for predicting and tracking aircraft states.

[0074] For example, the Kalman filter formula is as follows:

[0075]

[0076] in, is the state estimate at the current moment, is the Kalman gain, is the current observation value, is the observation matrix. is the state estimate of the previous step. is the Kalman gain, which determines the weight between the measured data and the predicted data. is the current observation data. is the observation matrix, which is used to convert the state prediction value into the observation space.

[0077] Furthermore, data fusion technology combines data from different sensors to form a unified, accurate environmental perception dataset. The innovation of this step lies in its use of an adaptive weighted fusion method, dynamically adjusting the weights of various sensor data based on flight conditions detected by the environmental perception module (such as light intensity, visibility, and weather type). For example, at night or in low-light environments, where visual sensor performance degrades, the system will correspondingly reduce the weight of visual data and increase the weight of infrared sensors. Conversely, during clear daytime conditions, where visual image clarity is high, the system prioritizes visual data. In complex weather conditions with low visibility, such as dense fog or strong winds, lidar, due to its superior penetrating power, is significantly weighted, enhancing the accuracy of obstacle location.

[0078] To implement this weighting mechanism, the system continuously analyzes data from the aircraft's auxiliary environmental sensors (such as the brightness sensor, barometer, and hygrometer) and categorizes the current environmental conditions. Based on the classification results (e.g., clear sky / daytime, dense fog and strong wind at night), the system automatically calls a configured weighting strategy table and dynamically sets the weighting coefficients.

[0079] For example:

[0080] When daytime / clear weather conditions are detected: , , ;

[0081] When detecting night / no streetlight conditions: , , ;

[0082] When dense fog is detected: , , ;

[0083] in, Represents the weighting coefficient of the visual sensor; Represents the weighting coefficient of the lidar sensor; Represents the weighted coefficient of the infrared sensor; specifically, the sensor weighted fusion calculation formula is as follows:

[0084]

[0085] in, is the fused environmental data, For the Data from sensors (vision, lidar, infrared), For the system according to time The weighted coefficient of the dynamic selection of the environment type (where i is or or ).

[0086] Through this mechanism, this method not only enhances the utilization efficiency of different sensors under their optimal working conditions, but also improves the system's adaptability to sudden environmental changes, ensuring that high-precision environmental perception capabilities can be maintained even in complex or sudden low-altitude flight environments.

[0087] Furthermore, the fusion environment perception data is output ( ), that is, the fused aircraft surrounding environment data, including the real-time status of the aircraft and environmental characteristics, will be passed as input data to the subsequent step (step 2) for feature extraction and risk identification.

[0088] S2. Input the fused environmental perception data into a lightweight multi-channel convolutional network, combine it with the spatial attention module to enhance the regional significant risk features, and output a local risk score tensor, which represents the risk at time Corresponding risk level scores for spatial locations, dimensions and environmental conditions Figure 1 That is, the local risk score tensor is a risk score map, and the high-risk area coordinates are identified in the risk score map to construct a high-risk area coordinate set.

[0089] The goal of this step is to integrate the environmental perception data output from step 1 into the real-time flight process. Environmental features highly correlated with flight risk are extracted from the data and used to identify potential risk areas. The environment in which low-altitude drones fly is complex and changeable, containing static obstacles (such as buildings and trees), dynamic targets (such as other drones and birds), and unstable factors (such as sudden changes in airflow and weather). Therefore, this step requires not only extracting geometric and motion features from the data, but also estimating the risk level based on the spatiotemporal context. To achieve high robustness and real-time performance, the system uses a method that combines a lightweight convolutional network with a time-series window sliding analysis to extract local features of the environment and conduct immediate risk assessments.

[0090] Specifically, this step directly inherits the output of step 1 , which is the fused multimodal environmental perception data. First, the system It is fed into a lightweight multi-channel convolutional feature extraction network. The network structure is a three-layer convolution structure, where the convolution kernel size of each layer is , the number of channels are 16, 32, and 64 respectively. The first layer processes the original visual image The second layer processes the point cloud map, and the third layer combines thermal imaging The key to this architecture is the addition of a spatial attention module after each convolutional layer to enhance regionally significant risk features (such as areas with high-density obstacle clusters or hotspots of rapidly approaching objects).

[0091] Furthermore, in order to improve the accuracy of risk identification and the foresight of the system, this step introduces a sliding time analysis mechanism based on a local dynamic window. Next, not only consider the current , and also quoted the past Frame data Perform short-term behavior change trend analysis. This mechanism is completed through the time feature fusion layer, outputting a local risk score tensor , indicating that at time The risk level score for the corresponding spatial location. The core formula for calculating the risk score is as follows:

[0092]

[0093] in, represents the output of the multi-channel convolutional feature extraction module, It is a timing weighting factor, which is adjusted exponentially based on the time distance from the current frame (e.g. ), is a normalized activation function, which is used to normalize the risk value to between.

[0094] in, For in time Fusion perception data at all times; It is a three-layer convolutional neural network feature extractor; is the time sequence weighting factor, which controls the weight of near time frames to be high and the weight of far time frames to be low; The risk score tensor for the current moment, with dimensions and environment Figure 1 It indicates the risk level of each spatial location.

[0095] Subsequently, the system extracts the local maximum area in the risk score map as the high-risk candidate area, and combines the aircraft speed direction The system uses the predicted route to determine whether there are high-risk intersections within the route. If so, its spatial coordinates and corresponding risk values ​​are recorded as the current instant risk identification output.

[0096] Finally, to prevent the system from over-responding to noise interference, the system introduces a regularization constraint based on spatial gradient to smooth the high-frequency fluctuations in the risk score graph caused by noise. The regularization term is defined as follows:

[0097]

[0098] in, and Respectively represent the risk graph in and The gradient in the direction is approximated using the Sobel operator, and the regularization term is used to suppress the sudden change of local risk value, thereby enhancing the stability and interpretability of the recognition results. The risk score diagram is The risk value of the position; To smooth the regularization term, it is ultimately used in the system as part of the confidence score to correct the risk results.

[0099] Furthermore, the risk score graph is output : Used to calibrate the risk level of different spatial locations within the current flight area; coordinate set of high-risk areas : The system outputs a list of risk points after threshold segmentation and spatial screening for use by subsequent prediction and path planning modules.

[0100] This step constructs a risk feature extraction mechanism centered around local temporal information. This mechanism leverages the multimodal fusion data output from step one, and through joint modeling of spatial attention and temporal windows, enables real-time identification of high-risk points along the flight path. Compared to traditional methods, this approach demonstrates greater adaptability and robustness in identifying dynamic risks and intra-path collision trends. It is particularly well-suited for low-altitude environments with rapidly emerging targets, frequent occlusions, and unstable lighting.

[0101] S3. Based on the risk score map and drone status data, a historical risk map sequence, a historical flight status sequence, and a historical sensor dynamic weight sequence are obtained. Combined with the current risk score map and the high-risk area coordinate set, a risk map sequence and a risk drift point sequence for the future time period are generated through a dual-input gated recurrent neural network.

[0102] In order to improve the foresight of the system, this step uses the multivariate time series analysis method to predict the potential risk change trend in the future flight process. Unlike traditional time series modeling, this approach integrates multiple heterogeneous temporal data (including the aircraft's own state sequence, environmental sensor response sequences, and local risk dynamics) to construct a higher-dimensional, more interactive time series analysis model. On this basis, a multi-input gated recurrent neural network is designed to model the joint evolution of multiple variables, thereby improving the accuracy and robustness of future risk predictions. This mechanism is crucial for handling multi-target interference and sudden meteorological changes in complex and dynamic low-altitude scenarios.

[0103] Specifically, this step uses the output of step 2 as the main input and simultaneously integrates the flight status and sensor behavior variables to form the following multivariate time series input:

[0104] Current Risk Score Graph ;

[0105] High-risk point collection ;

[0106] Historical Risk Map Series , each frame is a gridded and normalized two-dimensional risk distribution;

[0107] Historical flight status sequence ,in ;

[0108] Historical sensor dynamic weight sequence ,in ;

[0109] High-risk point quantity sequence .

[0110] Furthermore, these multivariate sequences together constitute the prediction input tensor group, and the time steps are aligned through the synchronized time window processing mechanism.

[0111] In order to integrate the above multi-modal, multi-source, and multi-dimensional temporal features, a dual-input gated recurrent neural network structure (Dual-GRU) is designed in this step. The structure consists of two parallel GRU modules: one for processing risk graph sequences and the other for , and another one that handles multivariate vector sequences , after outputting the intermediate features, they are spliced ​​and fused, and then the predicted risk map sequence for several time steps in the future is output through the fusion layer.

[0112] The core expression of the risk graph prediction of this structure is:

[0113]

[0114] in, and There are two groups of three-layer 128-unit GRU sub-networks, Represents the fusion layer, which consists of a fully connected layer plus an activation function.

[0115] in, To predict the future A two-dimensional risk score map at each moment; GRU model for processing risk graph sequences; GRU model for processing multivariate state vector sequences; It is a function that performs fusion transformation after splicing the output of two sequences; for the combined flight status, sensor response, and high-risk point density sequence.

[0116] The fusion output risk prediction map can identify high-risk areas that may appear on the future flight trajectory. However, considering that low-altitude aircraft have continuous motion characteristics, only the predicted spatial distribution map is not enough to locate the precise conflict point. Therefore, the system simultaneously adopts the forward risk point drift mechanism to calculate the current high-risk point set. Do drift mapping based on flight speed:

[0117]

[0118] This mechanism assumes that high-risk areas have a derivable spatial offset path under the action of aircraft inertia, which further improves prediction accuracy and reaction time.

[0119] in, To predict the time The position of the high-risk point after drifting; High-risk points identified at the current moment; is the current horizontal velocity component of the aircraft; is the forecast time step offset value.

[0120] Output: sequence of risk graphs for future time periods ; Risk drift point sequence : Used to accurately predict potential high-collision areas in the future.

[0121] As can be understood, this step, through joint modeling of multivariate time series, not only incorporates the evolutionary trends of the spatial risk map but also incorporates auxiliary information such as the aircraft's own behavior, sensor status, and local risk density, forming a highly coupled dynamic prediction model. The parallel design of the dual GRU architecture demonstrates the flexibility of time series information processing. Combined with the drift inference strategy, this approach enhances the practicality and interpretability of risk prediction. This solution effectively addresses the inability of single-variable modeling to address complex interacting factors and is a core component for implementing dynamic early warning and proactive avoidance strategies.

[0122] S4. An ROI region is intercepted from the risk map sequence, starting from the current position of the aircraft and extending along the current speed direction as a vector. The average risk value of each frame is extracted from the ROI region to calculate the future regional risk level. At the same time, the minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the calculation period is calculated, and the number of high-risk points entering the ROI region is counted to construct a high-risk point threat coefficient. In combination with the future regional risk level, the current speed of the aircraft and the urgency of the mission, a decision recommendation mark is generated.

[0123] This step aims to aggregate and model the future risk information predicted in step 3, and form quantitative risk levels and strategic recommendations for the current flight status. During low-altitude UAV flight, the system not only needs to know whether there is a risk somewhere, but also needs to form a comprehensive judgment on whether to take evasive action at this moment. To this end, this step designs a dynamic risk accumulation grading mechanism based on a multi-step prediction score map and high-risk point density, and introduces the current aircraft state parameters and mission parameters for joint decision-making, outputting two indicators: regional risk level ( ) and the action suggestion flag ( ).

[0124] Specifically, the core idea of ​​this step is to map the future risk The risk scores of the area along the aircraft's expected flight path are aggregated to form a regional risk score, which is then combined with information such as the density trend of high-risk points and the current flight speed to output two quantifiable indicators.

[0125] First, the system In the

[15] , a dynamic region of interest (ROI) is captured, starting from the current position of the aircraft and extending along the current velocity direction. This region simulates the natural forward flight of the drone without evasive maneuvers. The average risk value of each frame in this ROI is then extracted to construct the future regional risk level:

[0126]

[0127] in, For the The set of grid positions contained in the ROI area in the frame, is the risk prediction value at that point. This indicator reflects the average risk level faced by the aircraft in the next few frames if it continues to fly in the current direction.

[0128] Then, the system calculates all high risk points during the forecast period The minimum distance from the current trajectory of the aircraft, and the number of high-risk points entering the ROI area are counted to construct the high-risk point threat coefficient:

[0129]

[0130] in, It is an indicator function, which is used to count the high-risk points falling into the ROI area. The higher the value, the higher the degree of intersection between the future flight path and the high-risk area.

[0131] Finally, the system combines 、 , current speed and task urgency , entering a hierarchical strategy model. The model is divided into the following parts using a rule tree:

[0132] like and ,but (Continue flying);

[0133] like and ,but (maintain current status and continue monitoring);

[0134] like or ,but:

[0135] like : (It is recommended to hover or wait in place);

[0136] like : (A detour route planning is recommended).

[0137] Further, output: future regional risk level : Used to quantify the risk aggregation value of the future track area; decision recommendation mark : Action flag with a value of 0 to 3, which is called by the subsequent path planning module.

[0138] This step uses future multi-frame risk maps and high-risk point predictions, combined with the current flight direction and mission requirements, to construct a comprehensive risk level using spatiotemporal ROI risk integration and risk point density assessment. Action recommendations are then generated based on a rule-based approach. Unlike traditional rule-based threshold-based approaches, this solution considers both risk maps and drift point information, comprehensively assessing persistent risk trends across the predicted space and creating a precise and flexible decision support module.

[0139] S5. Generate an executable flight control instruction according to the decision suggestion flag to drive the UAV to execute.

[0140] Specifically, this step is responsible for the behavioral recommendations output from step 4 based on the completion of all risk perception, analysis and prediction. and regional risk level These instructions are converted into executable flight control commands, enabling actions such as path avoidance, speed adjustment, or hovering. This step eliminates modeling or evaluation, but directly calls the action control module to implement the flight strategy, thus completing the execution of the entire low-altitude risk warning closed loop.

[0141] System based The value of enters different control strategy templates and generates corresponding control instruction sets under each strategy The control instructions are expressed in the form of acceleration vector and angular velocity instructions, which are compatible with conventional UAV flight control interfaces.

[0142] like (Continue flying): The system maintains the current velocity vector direction and sets the target acceleration Zero, only speed maintenance and attitude stabilization adjustments are performed:

[0143]

[0144] in, is the attitude angular velocity control quantity, The current pitch and roll angles are calculated by the built-in stabilization module.

[0145] like (Continue to monitor): Same as above, but with the high-frequency status monitoring flag turned on, the flight control system enters high-frequency sampling mode, and the system rechecks the risk level every 0.2s.

[0146] like (Hovering is recommended): The system issues deceleration and altitude lock commands. Reverse acceleration is applied in the current direction of motion to reduce the velocity vector to zero. At the same time, the altitude lock module is activated to maintain the current altitude.

[0147]

[0148] in, is the deceleration coefficient, generally ranging from 0.8 to 1.2, is the current velocity vector, Indicates the altitude change rate. Setting it to zero means locking the altitude.

[0149] like (Avoidance recommended): The system constructs an avoidance heading vector by offsetting the aircraft's current position by a certain angle away from the high-risk point. , and gives the acceleration direction adjustment:

[0150]

[0151] in, is the maximum horizontal acceleration allowed by the system (defined by the flight control constraints), It is an avoidance direction vector constructed in the opposite direction of the reference drift point. For example, if the high-risk point is on the left front side of the current heading, the system will choose to correct the heading in the right rear direction.

[0152] Each generated control instruction It will be encapsulated into a standard MAVLink control frame and sent to the flight control system for execution via the communication module.

[0153] like Figure 2 As shown, another embodiment of the present invention provides a low-altitude UAV traffic operation risk warning system, the system comprising:

[0154] Data acquisition module 301: used to collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein the drone status data includes position, speed, and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar, and infrared sensors;

[0155] Risk identification module 302: used to input the fused environmental perception data into a lightweight multi-channel convolutional network, combined with the spatial attention module to enhance the regional significant risk features, and output a local risk score tensor, which represents the risk at time Corresponding risk level scores for spatial locations, dimensions and environmental conditions Figure 1 That is, the local risk score tensor is a risk score map, and the high-risk area coordinates are identified in the risk score map to construct a high-risk area coordinate set;

[0156] Risk prediction module 303: used to obtain a historical risk map sequence, a historical flight state sequence, and a historical sensor dynamic weight sequence based on the risk score map and drone status data, and generate a risk map sequence and a risk drift point sequence for the future time period through a dual-input gated recurrent neural network by combining the current risk score map and the high-risk area coordinate set;

[0157] Decision support module 304 is configured to intercept a ROI region from the risk map sequence, extending from the current position of the aircraft and the current speed direction as a vector, extract the average risk value of each frame in the ROI region to calculate the future regional risk level, and simultaneously calculate the minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the prediction period, and count the number of high-risk points entering the ROI region to construct a high-risk point threat coefficient. Furthermore, a decision recommendation flag is generated based on the future regional risk level, the current speed of the aircraft, and the urgency of the mission.

[0158] The flight control module 305 is used to generate executable flight control instructions according to the decision suggestion flag to drive the UAV to execute.

[0159] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of the present invention. In practical applications, technicians in this field can select part or all of it according to actual needs to achieve the purpose of the embodiment scheme, and no limitation is made here.

[0160] In addition, for technical details not fully described in this embodiment, please refer to the parameter operation method provided in any embodiment of the present invention, and will not be repeated here.

[0161] It should be noted that, in this document, the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, article, or system comprising a series of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or system. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or system comprising the element.

[0162] The serial numbers of the above embodiments of the present invention are for description only and do not represent the advantages or disadvantages of the embodiments.

[0163] Through the above description of the embodiments, those skilled in the art will clearly understand that the methods of the above embodiments can be implemented using software plus the necessary general-purpose hardware platform. Of course, hardware can also be used, but in many cases the former is a more preferred implementation method. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory / random access memory, a magnetic disk, or an optical disk) and includes a number of instructions for enabling a terminal device (such as a mobile phone, computer, server, air conditioner, or network device) to execute the methods described in the various embodiments of the present invention.

[0164] The above are only preferred embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention description and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A low-altitude UAV traffic operation risk warning method, characterized in that: The method comprises: S1. Collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein the drone status data includes position, speed, and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar, and infrared sensors; S2. Input the fused environmental perception data into a lightweight multi-channel convolutional network, combined with a spatial attention module to enhance regional significant risk features, and output a local risk score tensor representing the risk level score corresponding to the spatial location at time t. The dimension is consistent with the environmental map, that is, the local risk score tensor is a risk score map. The coordinates of high-risk areas are identified in the risk score map to construct a high-risk area coordinate set. The risk score map is used to calibrate the risk level of different spatial locations within the current flight area. S3. Based on the risk score map and drone status data, a historical risk map sequence, a historical flight status sequence, and a historical sensor dynamic weight sequence are obtained. Combined with the current risk score map and the high-risk area coordinate set, a risk map sequence and a risk drift point sequence for future time periods are generated through a dual-input gated recurrent neural network. S4. A ROI region is intercepted from the risk map sequence, starting from the current position of the aircraft and extending along the current speed direction as a vector. The average risk value of each frame in the ROI region is extracted to calculate the future regional risk level. The minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the period is calculated. The number of high-risk points entering the ROI region is counted to construct a high-risk point threat coefficient. The decision recommendation flag is generated based on the future regional risk level, the current speed of the aircraft, and the urgency of the mission. Among them, the future regional risk level R zone The calculation is as follows: Where T is the time window, ROI τ is the grid position set contained in the ROI area in the τth frame, is the risk prediction value of the current position; (x, y) is the coordinate of the current position of the aircraft; The high-risk point threat coefficient ω is constructed as follows: in, is an indicator function used to count the high-risk points falling into the ROI area. The highest value of ω indicates that the future flight path will have the highest degree of intersection with the high-risk area; τ is the current time window; is the current risk drift point sequence, which is used to predict the potential high-risk collision area in the future; S5. Generate an executable flight control instruction according to the decision suggestion flag to drive the UAV to execute.

2. A low-altitude UAV traffic operation risk warning method according to claim 1, characterized in that: The lightweight multi-channel convolutional network has a three-layer convolutional structure, where the convolution kernel size of each layer is 3×3 and the number of channels is 16, 32, and 64 respectively. The first layer processes the image data, the second layer processes the point cloud mapping of the lidar, and the third layer combines the thermal imaging channel of the infrared sensor to achieve multimodal feature joint encoding. Among them, a spatial attention module is added after each convolutional layer to enhance the regional significant risk characteristics; and the risk score map is smoothed through spatial gradient regularization constraints to suppress noise interference.

3. A low-altitude UAV traffic operation risk warning method according to claim 2, characterized in that: The identifying high-risk area coordinates in the risk score map to construct a high-risk area coordinate set specifically includes: The risk score map is used as a high-risk candidate area, and combined with the aircraft speed direction and predicted route, it is determined whether there is a high-risk intersection in the path; if so, its spatial coordinates and corresponding risk value are recorded as the current high-risk area coordinates.

4. A low-altitude UAV traffic operation risk warning method according to claim 1, characterized in that: The dual-input gated recurrent neural network consists of two parallel GRU modules, specifically: The first GRU module processes the risk map sequence, and the second GRU module processes the multivariate vector sequence, outputs intermediate features, performs splicing and fusion, and then outputs a predicted risk map sequence for several future time steps through the fusion layer. The multivariate vector sequence is a combined flight status, sensor response, and high-risk point density sequence. The first GRU module and the second GRU module are two groups of three-layer 128-unit GRU sub-networks.

5. A low-altitude UAV traffic operation risk warning method according to claim 4, characterized in that: The step of generating a risk drift point sequence includes: A current high-risk point set and a current horizontal velocity component of the aircraft are obtained, and drift mapping based on the current horizontal velocity component of the aircraft is performed on the current high-risk point set to generate a risk drift point sequence.

6. A low-altitude UAV traffic operation risk warning method according to claim 1, characterized in that: The generating of a decision suggestion flag by combining the future regional risk level, the current speed of the aircraft and the urgency of the mission specifically includes: The regular tree method is used for partitioning, where F action is the decision suggestion sign, and η is the task urgency: If R zone <0.3 and high ω=0, then F action =0, indicating continued flight; If 0.3≤R zone <0.6 and ω≤2, then F action =1, it means to keep the current state and continue monitoring; If R zone ≥0.6 or ω≥3, then: If η<0.5: F action =2, it is recommended to hover or wait; If η≥0.5: F action =3, detour route planning is recommended.

7. A low-altitude UAV traffic operation risk warning method according to claim 6, characterized in that: Generating an executable flight control instruction according to the decision suggestion flag to drive the UAV to execute specifically includes: If F action =0, the drone maintains the current velocity vector direction, sets the target acceleration to zero, and only performs speed maintenance and attitude stabilization adjustments; If F action =1, then with F action =0 is the same as the above, but the high-frequency status monitoring flag is turned on, entering the high-frequency sampling mode, and the system rechecks the risk level every 0.2s; If F action =2, then the deceleration and altitude lock instructions are issued; the reverse acceleration is applied in the current direction of motion to decay the velocity vector to zero, and the altitude lock module is activated at the same time to keep the current altitude unchanged; If F action =3, then the aircraft is offset by a certain angle from its current position away from the high-risk point, an avoidance heading vector is constructed, and the acceleration direction is adjusted.

8. A low-altitude UAV traffic operation risk warning method according to claim 7, characterized in that: Each generated control command will be encapsulated into a standard MAVLink control frame and sent to the flight control system for execution via the communication module; the control command includes acceleration and avoidance heading vector.

9. A system for executing the low-altitude UAV traffic operation risk warning method according to claim 1, characterized in that: The system comprises: Data acquisition module: used to collect drone status data and external environment data, perform data preprocessing, and generate fused environmental perception data; wherein, the drone status data includes position, speed and flight altitude; the external environment data is sensor data, and the sensors include image data, lidar and infrared sensors; Risk Identification Module: This module is used to input the fused environmental perception data into a lightweight multi-channel convolutional network, combined with a spatial attention module to enhance regional significant risk features, and output a local risk score tensor representing the risk level score corresponding to the spatial location at time t. The dimension is consistent with the environmental map, that is, the local risk score tensor is a risk score map. The coordinates of high-risk areas are identified in the risk score map to construct a high-risk area coordinate set. The risk score map is used to calibrate the risk level of different spatial locations within the current flight area. Risk prediction module: used to obtain a historical risk map sequence, a historical flight state sequence, and a historical sensor dynamic weight sequence based on the risk score map and drone status data. Combined with the current risk score map and the high-risk area coordinate set, a risk map sequence and a risk drift point sequence for the future time period are generated through a dual-input gated recurrent neural network. Decision support module: used to intercept an ROI area from the risk map sequence, starting from the current position of the aircraft and extending from the current speed direction as the vector, extract the average risk value of each frame in the ROI area to calculate the future regional risk level, and simultaneously calculate the minimum distance between the coordinates of all high-risk areas and the current heading trajectory of the aircraft during the period, and count the number of high-risk points entering the ROI area to construct the high-risk point threat coefficient. In combination with the future regional risk level, the current speed of the aircraft and the urgency of the mission, a decision recommendation flag is generated; Flight control module: used to generate executable flight control instructions according to the decision suggestion flag to drive the drone to execute.

Images