A low-altitude unmanned aerial vehicle conflict detection method and system based on trajectory prediction, a terminal, and a storage medium

Abstract

The application relates to the technical field of unmanned planes, and discloses a low-altitude unmanned plane conflict detection method and system based on trajectory prediction, a terminal and a storage medium, the method comprises the following steps: acquiring space-time flow trajectory data of an unmanned plane, pre-processing the space-time flow trajectory data to obtain target space-time flow trajectory data, and dividing the target space-time flow trajectory data according to a preset division rule to obtain behavior characteristic constraint information; fusing the target space-time flow trajectory data and the behavior characteristic constraint information to obtain a space-time fusion input sequence; inputting the space-time fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain a plurality of trajectory prediction points; and establishing a plurality of three-dimensional hybrid bounding box models according to the plurality of trajectory prediction points, performing conflict detection on the plurality of three-dimensional hybrid bounding box models, and obtaining a conflict detection result. The application detects the collision of a plurality of unmanned planes according to the trajectories of the unmanned planes, and realizes accurate detection of the collision of the unmanned planes.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) technology, and in particular to a method, system, terminal, and computer-readable storage medium for low-altitude UAV conflict detection based on trajectory prediction. Background Technology

[0002] With the rapid development of drone technology and its widespread application in urban logistics, emergency response, and aerial inspection, the frequency of various drones performing tasks such as logistics delivery, medical supply transportation, and urban security patrols is increasing daily. This has created a complex airspace environment characterized by high density, multiple types, and mixed operations. During flight, drones are susceptible to multiple factors, including electromagnetic interference, satellite (GPS, Global Positioning System) signal deception, complex weather conditions, and hardware / software malfunctions. This can lead to abnormal behaviors such as flight path deviation, uncontrolled hovering, and unexpected maneuvers, potentially causing spatial conflicts between drones and buildings or manned aircraft, and even disrupting airspace operations and causing major safety accidents. To ensure airspace safety, high-precision prediction of drone flight trajectories and real-time detection and early warning of flight conflicts have become critical technical issues that urgently need to be addressed. In complex low-altitude urban environments, although drones plan their paths in advance before takeoff, non-cooperative drones lack cross-departmental collaborative communication mechanisms. Relying solely on single-drone obstacle avoidance and pre-planned flight paths is insufficient to cope with dynamic and complex environments, and electromagnetic interference further exacerbates the difficulty of real-time monitoring.

[0003] Existing conflict detection methods include rule-based detection methods based on fixed thresholds, trajectory prediction methods based on physical models, and prediction detection methods based on traditional machine learning. These traditional trajectory prediction methods are limited by weak maneuverability and insufficient prediction accuracy, making them unable to effectively support conflict early warning. Consequently, using existing conflict detection methods cannot reliably predict UAV trajectories continuously, and it is difficult to accurately detect conflicts involving multiple UAVs. Therefore, existing UAV trajectory prediction and conflict detection technologies still need improvement and optimization. Summary of the Invention

[0004] The main objective of this invention is to provide a low-altitude UAV conflict detection method, system, terminal, and computer-readable storage medium based on trajectory prediction, aiming to solve the problems that existing conflict detection methods have difficulty in estimating the current trajectory of UAVs in real time and stably, and have difficulty in accurately detecting the potential conflict of multiple UAVs colliding.

[0005] To achieve the above objectives, the present invention provides a low-altitude UAV conflict detection method based on trajectory prediction, the method comprising the following steps:

[0006] Acquire the spatiotemporal flow trajectory data of the UAV, preprocess the spatiotemporal flow trajectory data to obtain the target spatiotemporal flow trajectory data, and divide the target spatiotemporal flow trajectory data into behavioral features according to a preset division rule to obtain behavioral feature constraint information;

[0007] The target spatiotemporal flow trajectory data and the behavioral feature constraint information are fused to obtain a spatiotemporal fusion input sequence. The spatiotemporal fusion input sequence is then input into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period.

[0008] Multiple 3D hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results. Collision detection is performed on multiple 3D hybrid bounding box models to obtain collision detection results.

[0009] Optionally, in the low-altitude UAV conflict detection method based on trajectory prediction, the preprocessing includes standardization processing, completion processing, anomaly detection, and rejection processing.

[0010] The process of acquiring the spatiotemporal flow trajectory data of the UAV, preprocessing the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and dividing the target spatiotemporal flow trajectory data into behavioral features according to preset segmentation rules to obtain behavioral feature constraint information specifically includes:

[0011] Acquire the spatiotemporal trajectory data and current timestamp of the UAV, and standardize the geographic coordinate system of the spatiotemporal trajectory data according to the current timestamp to obtain the projected coordinates;

[0012] The spatiotemporal flow trajectory data is completed by using a linear interpolation method based on the projected coordinates to obtain the first spatiotemporal flow trajectory data;

[0013] Anomaly detection is performed on the first spatiotemporal flow trajectory data using the local outlier factor algorithm to obtain abnormal trajectories. The first spatiotemporal flow trajectory data is then removed based on the abnormal trajectories to obtain target spatiotemporal flow trajectory data. The target spatiotemporal flow trajectory data is then divided into behavioral features according to a preset partitioning rule to obtain behavioral feature constraint information.

[0014] The geographic coordinate system includes longitude, latitude, and altitude;

[0015] The behavioral characteristic constraint information includes takeoff characteristics, landing characteristics, rapid ascent characteristics, dive characteristics, cruise characteristics, turning characteristics, sharp turn characteristics, hovering characteristics, emergency stop characteristics, and circling characteristics.

[0016] Optionally, the low-altitude UAV conflict detection method based on trajectory prediction, wherein fusing the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, and inputting the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period, specifically includes:

[0017] The behavioral feature constraint information is mapped according to a preset one-hot encoding to obtain semantic features;

[0018] The target spatiotemporal flow trajectory data and the semantic features are fused according to the preset one-hot encoding to obtain a spatiotemporal fusion input sequence;

[0019] A hybrid trajectory prediction neural network model is constructed by fusing a temporal convolutional perception layer, a temporal evolution parsing layer, and a behavioral semantic fusion layer. The spatiotemporal fusion input sequence is then input into the hybrid trajectory prediction neural network model for prediction, resulting in predictions of multiple continuous trajectories within a future time period.

[0020] Optionally, the low-altitude UAV conflict detection method based on trajectory prediction, wherein fusing the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, and inputting the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period, further includes:

[0021] Obtain the joint loss function, and perform constraint optimization on the hybrid trajectory prediction neural network model based on the position error loss and motion continuity loss of the joint loss function to obtain the first hybrid trajectory prediction neural network model.

[0022] The first hybrid trajectory prediction neural network model is subjected to gradient optimization according to the preset dynamic correction rule to obtain the optimal hybrid trajectory prediction neural network model.

[0023] The optimal hybrid trajectory prediction neural network model is used to make high-precision predictions of the future trajectory of the UAV.

[0024] Optionally, the low-altitude UAV conflict detection method based on trajectory prediction, wherein the step of establishing multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, and performing conflict detection on the multiple three-dimensional hybrid bounding box models to obtain conflict detection results, specifically includes:

[0025] Multiple three-dimensional hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results;

[0026] Obtain the outer axis-aligned static bounding boxes of multiple 3D hybrid bounding box models. If any one of the outer axis-aligned static bounding boxes has a collision risk, then perform collision detection on the multiple 3D hybrid bounding box models to obtain the collision detection result.

[0027] Optionally, the low-altitude UAV conflict detection method based on trajectory prediction, wherein obtaining the outer axis-aligned static bounding boxes of multiple 3D hybrid bounding box models, and if any one of the outer axis-aligned static bounding boxes has a collision risk, then performing collision detection on the multiple 3D hybrid bounding box models to obtain a conflict detection result, specifically includes:

[0028] Obtain the outer axis-aligned static bounding box and multiple oriented bounding boxes of the multiple 3D hybrid bounding box models;

[0029] If any one of the multiple outer axis-aligned static bounding boxes has a collision risk, then the multiple directed bounding boxes are dynamically adjusted according to the semantic features to obtain multiple inner dynamic semantic deformation bounding boxes.

[0030] Collision detection is performed on multiple three-dimensional hybrid bounding box models based on multiple inner-layer dynamic semantic deformation bounding boxes to obtain conflict detection results.

[0031] Optionally, the low-altitude UAV conflict detection method based on trajectory prediction, wherein the step of establishing multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to the multiple consecutive trajectory prediction results, performing conflict detection on the multiple three-dimensional hybrid bounding box models to obtain conflict detection results, further includes:

[0032] The spatial intersection pattern and heading angle relationship of the continuous trajectory prediction results are obtained, and the scene judgment is performed on the conflict detection results based on the spatial intersection pattern and the heading angle relationship to obtain the target conflict scene type;

[0033] Obtain the minimum spatiotemporal distance and conflict duration of the conflict detection results, and obtain the conflict level based on the minimum spatiotemporal distance and the conflict duration;

[0034] Warnings are issued based on the conflict level and the target conflict scenario type to notify users to conduct real-time monitoring of the drone;

[0035] The target conflict scenario types include any one of head-on collision, cross-traffic, and same-direction rear-end collision;

[0036] The conflict level includes any one of high, medium, and low.

[0037] Furthermore, to achieve the above objectives, the present invention also provides a low-altitude UAV conflict detection system based on trajectory prediction, wherein the low-altitude UAV conflict detection system based on trajectory prediction is:

[0038] The feature segmentation module is used to acquire the spatiotemporal flow trajectory data of the UAV, preprocess the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and perform behavioral feature segmentation on the target spatiotemporal flow trajectory data according to preset segmentation rules to obtain behavioral feature constraint information.

[0039] The trajectory prediction module is used to fuse the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, and input the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period.

[0040] The collision detection module is used to establish multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, perform collision detection on the multiple three-dimensional hybrid bounding box models, and obtain collision detection results.

[0041] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a low-altitude UAV conflict detection program based on trajectory prediction, and the low-altitude UAV conflict detection program based on trajectory prediction implements the steps of the low-altitude UAV conflict detection method based on trajectory prediction as described above when executed by a processor.

[0042] In this invention, spatiotemporal flow trajectory data of a UAV is acquired, preprocessed to obtain target spatiotemporal flow trajectory data, and then divided into behavioral features according to preset segmentation rules to obtain behavioral feature constraint information. The target spatiotemporal flow trajectory data and behavioral feature constraint information are fused to obtain a spatiotemporal fusion input sequence, which is then input into a hybrid trajectory prediction neural network model for prediction, resulting in multiple trajectory prediction points. Multiple three-dimensional hybrid bounding box models are established based on these prediction points, and collision detection is performed on these models to obtain collision detection results. This invention first acquires UAV spatiotemporal flow trajectory data, preprocesses the trajectory data, and divides the UAV trajectory into motion features. Second, a deep learning trajectory prediction model with spatiotemporal feature fusion is established to predict the UAV trajectory. Finally, a hierarchical, progressive hybrid bounding box collision detection method is designed to discriminate collision detection results and provide rapid feedback, significantly improving the accuracy and real-time performance of multi-UAV flight collision detection in complex airspace. Attached Figure Description

[0043] Figure 1This is a flowchart of a preferred embodiment of the low-altitude UAV conflict detection method based on trajectory prediction of the present invention;

[0044] Figure 2 This is a flowchart of a preferred embodiment of the low-altitude UAV conflict detection method based on trajectory prediction of the present invention.

[0045] Figure 3 This is a flowchart of the spatiotemporal flow trajectory data of a preferred embodiment of the low-altitude UAV conflict detection method based on trajectory prediction of the present invention;

[0046] Figure 4 This is a flowchart of a hybrid trajectory prediction neural network model, representing a preferred embodiment of the low-altitude UAV conflict detection method based on trajectory prediction of the present invention.

[0047] Figure 5 This is a flowchart of UAV bounding box collision detection, a preferred embodiment of the low-altitude UAV collision detection method based on trajectory prediction of the present invention.

[0048] Figure 6 This is a structural diagram of a preferred embodiment of the low-altitude UAV conflict detection system based on trajectory prediction of the present invention;

[0049] Figure 7 This is a structural diagram of a preferred embodiment of the terminal of the device of the present invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0051] Existing conflict detection methods include rule-based detection methods based on fixed thresholds, trajectory prediction methods based on physical models, and prediction detection methods based on traditional machine learning. These traditional trajectory prediction methods are limited by weak maneuverability and insufficient prediction accuracy, and cannot effectively support conflict early warning. As a result, existing conflict detection methods cannot accurately predict the trajectory of UAVs, nor can they accurately detect conflicts between multiple UAVs. Therefore, a low-altitude UAV conflict detection method based on trajectory prediction is needed. This method can accurately predict the trajectory of UAVs through a hybrid trajectory prediction neural network model and detect conflicts between multiple UAVs with high accuracy, thus avoiding the problem of not being able to accurately detect conflicts between multiple UAVs.

[0052] The preferred embodiment of the low-altitude UAV conflict detection method based on trajectory prediction described in this invention, such as... Figure 1 and Figure 2As shown, the low-altitude UAV conflict detection method based on trajectory prediction includes the following steps:

[0053] Step S10: Obtain the spatiotemporal flow trajectory data of the UAV, preprocess the spatiotemporal flow trajectory data to obtain the target spatiotemporal flow trajectory data, and divide the target spatiotemporal flow trajectory data into behavioral features according to the preset division rules to obtain behavioral feature constraint information.

[0054] Step S10 includes:

[0055] Step S11: Obtain the spatiotemporal trajectory data and current timestamp of the UAV; standardize the geographic coordinate system of the spatiotemporal trajectory data according to the current timestamp to obtain the projected coordinates.

[0056] Step S12: Complete the spatiotemporal flow trajectory data according to the projected coordinates using a linear interpolation method to obtain the first spatiotemporal flow trajectory data;

[0057] Step S13: Anomaly detection is performed on the first spatiotemporal flow trajectory data using the local outlier factor algorithm to obtain abnormal trajectories. The first spatiotemporal flow trajectory data is then removed based on the abnormal trajectories to obtain target spatiotemporal flow trajectory data. The target spatiotemporal flow trajectory data is then divided into behavioral features according to preset division rules to obtain behavioral feature constraint information.

[0058] Specifically, such as Figure 3As shown, the preprocessing includes standardization, completion, anomaly detection, and removal. It acquires the UAV's spatiotemporal trajectory data and current timestamp. Based on the current timestamp, it standardizes the geographic coordinate system of the spatiotemporal trajectory data (mainly including UAV ID, timestamp, latitude, longitude, and altitude information; subsequent extensions include pitch angle, yaw angle, and roll angle; the trajectory data arrives sequentially in timestamp order) to obtain projected coordinates. This is achieved by standardizing the latitude, longitude, and altitude information recorded in the UAV flight record using the geographic coordinate system (WGS-84, World Geodetic System 1984) according to the current timestamp, converting the trajectory coordinate system from the geographic coordinate system to the projected coordinate system (UTM, Universal Transverse). Mercator (a method for adding projected coordinates to a buffer, where the geographic coordinate system includes longitude, latitude, and altitude) is used to complete the spatiotemporal trajectory data based on the projected coordinates using linear interpolation, resulting in first spatiotemporal trajectory data (the trajectory coordinate system is converted from the geographic coordinate system to the projected coordinate system, and missing data points are completed using linear interpolation). Anomaly detection is then performed on the first spatiotemporal trajectory data using the Local Outlier Factor (LOF) algorithm to obtain abnormal trajectories. These abnormal trajectories are then used to remove the first spatiotemporal trajectory data, resulting in target spatiotemporal trajectory data (using the Local Outlier Factor (LOF) algorithm). Factor) performs anomaly detection on the spatiotemporal trajectory data, identifying and removing trajectory segments that significantly deviate from the normal flight mode, such as abrupt jumps and stationary shifts, thereby improving the quality of model training data. It also performs behavioral feature segmentation on the target spatiotemporal trajectory data according to preset segmentation rules, obtaining behavioral feature constraint information (based on kinematic laws and trajectory change trends, extracting the UAV's behavioral features within a sliding window and classifying them into ten types: takeoff features, landing features, rapid ascent features, dive features, cruise features, turning features, sharp turn features, hovering features, and circling features, enabling the use of UAV behavioral features as constraints for subsequent deep learning trajectory prediction). The geographic coordinate system includes longitude, latitude, and altitude; the behavioral feature constraint information includes takeoff features, landing features, rapid ascent features, dive features, cruise features, turning features, sharp turn features, hovering features, sharp stop features, and circling features.

[0059] In this embodiment, longitude ( ) and latitude ( The projection transformation formula for () is as follows:

[0060] ;

[0061] ;

[0062] in, Represents the x-coordinate after longitude projection. Represents the ordinate after latitude projection. Represents the standard scale factor. Indicates the radius of curvature of the zonal loop. Indicates the arc length of the meridian. Indicates longitude. Indicates latitude, This indicates the corresponding central meridian.

[0063] In this embodiment, because unstable digital transmission signals or communication delays can lead to missing spatiotemporal flow trajectory data, it is necessary to perform continuity detection on the timestamps of the spatiotemporal flow trajectory data. First, the timestamps are standardized and sorted, and then the time points at any given time ( ) and the previous moment ( The time intervals between the two points are calculated to construct a time difference sequence. Under the premise of the theoretical sampling interval, if at a certain moment If a point is found to be larger than the sampling interval, it is considered to be missing. For the missing point, a linear interpolation method based on the time axis is used to reconstruct the actual location data. The trajectory is defined at the specified time point. and Missing If there are 10 data points, the formula for calculating the interpolation coordinates of the missing points is: , The calculation formula is as follows:

[0064] ;

[0065] in, This represents the total number of missing data points. express The missing point One location is missing. express Location coordinates at a given time point express Location coordinates at a given time point Indicates the first Interpolated coordinates of the missing points;

[0066] By reconstructing the spatiotemporal flow trajectory data at time intervals, the smoothness and rationality of the reconstructed trajectory in physical space can be guaranteed, which helps to improve the temporal consistency and prediction robustness of deep models.

[0067] Furthermore, in UAV flight data acquisition, anomalies often appear in the trajectory due to GPS signal jumps, inertial navigation bias, magnetic interference, or data communication interruptions. These anomalies manifest as abrupt jumps, stationary drift, sudden speed increases, and other abnormal behaviors. If these anomalies are not removed, interference trajectory pattern recognition and deep model training will be hindered. For example, first define the time-space neighborhood of the trajectory points, and then... trajectory points By the Three-dimensional coordinates of a trajectory point ( ) and timestamp ( Composed of, and the first trajectory points By the Three-dimensional coordinates of a trajectory point ( ) and timestamp ( Composed of, for each Define its Nearest neighbor set, containing the nearest neighbors in the trajectory. The closest The distance between points can be measured using weighted Euclidean spacetime distance. The calculation formula is as follows:

[0068] ;

[0069] in, Indicates the first The point in time, Indicates the spatial distance between two points. This is a time-scale adjustment coefficient used to unify the weights of time dimension and spatial distance. Through this spatiotemporal distance metric, the neighborhood not only considers spatial location similarity but also flight time consistency to ensure that local pattern judgments are more physically meaningful.

[0070] For example, such as Figure 3 As shown, the specific process of the removal process (removal of abnormal trajectories) is as follows: Then, based on the spatiotemporal neighborhood, the local outlier factor is calculated, and the trajectory points of the first spatiotemporal flow trajectory data ( Select the closest point from all trajectory points. Each point as Distance, to avoid overestimation of density due to some points being too close to their neighbors, for each trajectory point ( ) and neighboring points ( Its reachability distance is defined as: The reachability distance metric reflects the structural distance between two points, avoiding over-reliance on locally dense points and improving sensitivity to anomalously sparse points. (Trajectory points) The local reachability density of a point can describe the degree of clustering of that point in the local space, where, This indicates the degree of clustering. Higher density suggests the point is located in a densely populated local trajectory region, while lower density may indicate an anomaly or boundary point. Based on these definitions, the local outlier factor can be derived. ), The calculation formula is as follows:

[0071] ;

[0072] The calculation formula is as follows:

[0073] ;

[0074] in, Represents the distance to the spatial neighborhood. Indicates distance point Recent One point;

[0075] Finally, the local outlier factors are judged. If If the point is identified as a local outlier, it is suspected of exhibiting jumps, drifting, or abnormal errors, and the point is directly removed.

[0076] like Figure 3 As shown, the specific process of behavioral feature segmentation is as follows: To improve the trajectory prediction model's ability to perceive flight status, the behavioral states of the UAV during flight are segmented by combining flight kinematic features and trajectory change trends. This behavioral feature segmentation not only has flight semantic expression capabilities but also provides basic semantic support for subsequent behavioral guidance prediction modeling and bounding box dynamic adjustment. First, trajectory features are extracted and time windows are defined. The trajectory data is sliced ​​according to a fixed time window, and the defined window is set. and step length Extract the corresponding flight speed for each window segment. acceleration , Rate of change of direction angle , height change rate Key dynamic features such as horizontal turning rate are combined with parameters such as mean, standard deviation, and extreme values ​​of multiple time window steps as feature quantities.

[0077] Subsequently, based on the extracted multidimensional dynamic features, behavioral feature classification rules are defined, and the following ten typical UAV behavioral features and their judgment logic are designed: 1. Takeoff feature: 2. Landing characteristics: 3. Rapid Ascend characteristic: 4. Diving characteristics: 5. Cruise characteristics: 6. Steering characteristics: 7. Abrupt Turning: 8. Hovering feature: 9. Sudden stop: 10. Circling: ;

[0078] in, The height threshold indicating a significant decrease The velocity threshold representing a significant change This indicates a higher acceleration threshold. This indicates a smaller threshold for height variation. Indicates the threshold for significant changes in direction angle. Indicates the negative acceleration threshold. Represents the curvature of a two-dimensional plane trajectory. The curvature threshold represents the degree of curvature.

[0079] Step S20: The target spatiotemporal flow trajectory data and the behavioral feature constraint information are fused to obtain a spatiotemporal fusion input sequence. The spatiotemporal fusion input sequence is then input into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period.

[0080] Step S20 includes:

[0081] Step S21: Map the behavioral feature constraint information according to the preset one-hot encoding to obtain semantic features;

[0082] Step S22: The target spatiotemporal flow trajectory data and the semantic features are fused according to the preset one-hot encoding to obtain a spatiotemporal fusion input sequence;

[0083] Step S23: The temporal convolutional perception layer, temporal evolution parsing layer and behavior semantic fusion layer are fused to construct a hybrid trajectory prediction neural network model. The spatiotemporal fusion input sequence is input into the hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results in the future time period.

[0084] Specifically, such as Figure 4As shown, the behavioral feature constraint information is mapped according to a preset one-hot encoding to obtain semantic features. The target spatiotemporal trajectory data and the semantic features are fused according to the preset one-hot encoding to obtain a spatiotemporal fusion input sequence (the preprocessed target spatiotemporal trajectory data is constructed into an input sequence using a sliding window method, and the behavioral features are mapped to a low-dimensional semantic space using behavioral one-hot encoding, converting discrete UAV spatial features into continuous semantic features, achieving deep fusion of semantic features and target spatiotemporal trajectory data). The temporal convolutional perception layer, temporal evolution parsing layer, and behavioral semantic fusion layer are fused to construct a hybrid trajectory prediction neural network model. The spatiotemporal fusion input sequence is input into the hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period (using a temporal convolutional perception layer...). TCP Temporal ConvolutionPerception), Temporal Evolution Analysis Layer ( TEA Temporal Evolution Analysis), behavioral semantic fusion layer ( BSF A hybrid trajectory prediction neural network architecture (using Behavior Semantic Fusion) is constructed to obtain a hybrid trajectory prediction neural network model. A behavioral attention mechanism is introduced to achieve deep fusion of spatial state and flight semantics. The spatiotemporal fusion input sequence is input into the hybrid trajectory prediction neural network model for prediction, and finally, multiple continuous trajectory prediction results are output.

[0085] In this embodiment, the first step is to capture the temporal dependencies of the UAV trajectory, mainly addressing the dependence of UAV trajectory data on time order, the impact of historical information on future predicted trajectories, and the dynamic evolution patterns. This requires constructing a time-series input using a sliding window method, assuming the original trajectory data is a multidimensional sequence arranged in ascending time order. The calculation formula is as follows:

[0086] ;

[0087] in, Represents historical trajectory data points. Indicates the drone at a certain time The three-dimensional position space coordinates of the point. Indicates the duration or quantity of time.

[0088] Using a length of The sliding window constructs the input sequence ( Each window constitutes a model training sample, and each input sequence includes... From 1 to Historical trajectory points, The calculation formula is as follows:

[0089] ;

[0090] in, Indicates a sliding window. Indicates the first A historical trajectory point, Indicates the first A historical trajectory point, Indicates the first A historical trajectory point.

[0091] Subsequently, one-hot encoding of the drone's behavioral features is performed to enable them to directly participate in neural network training, and behavioral feature constraint information is used. These ten predefined categories—takeoff features, landing features, rapid ascent features, dive features, cruise features, turning features, sharp turn features, hovering features, sudden stop features, and circling features—are represented as vectors using one-hot encoding.

[0092] ;

[0093] Finally, the semantic features of the drone's behavior and trajectory features are fused. The one-hot variables of behavior at each time step are directly concatenated with the position vector. The behavior features and spatial position features are then fused into a single feature through vector dimension expansion, forming the fused input. ), The formula is as follows:

[0094] ;

[0095] Thus, the input dimensions at each time step are expanded from 3 to 13 (where, Represents spatial coordinates, =3D coordinate points, behavioral feature constraint information is =10-dimensional one-hot encoding, where each dimension is represented as: , , To represent one dimension, Representing two dimensions, (representing ten dimensions), ultimately the spatiotemporal fusion input sequence of each training sample ( )for, The formula is as follows:

[0096] ;

[0097] in, Indicates the first A fusion input, Indicates the first A fusion input, Indicates the first A fusion input, Represents spatial coordinates, Represents the real vector space, Representing behavioral feature constraint information, the completed spatiotemporal fusion input sequence is constructed. This input serves as the input to the subsequent spatiotemporal neural network, allowing the hybrid neural network model to extract dynamic spatial features step by step and combine them with behavioral semantics for attention weighting. Through this input design, the model can learn the rules of position evolution while capturing the guiding effect of behavioral features on future trajectories, thus achieving a deep spatiotemporal-semantic fusion of trajectory prediction.

[0098] Furthermore, the output of the temporal convolutional perceptron layer is , The formula is as follows:

[0099] ;

[0100] in, The number of channels represents the local feature representation extracted at each time step. The temporal convolutional perceptual layer, as the first layer of the network, utilizes causal convolution and dilated convolution to extract short-term local dynamic features from the trajectory sequence, enabling rapid perception of local changes such as acceleration, turning, and altitude changes in the drone trajectory. The temporal evolution parsing layer, after receiving the output from the temporal convolutional perceptual layer, models the evolution of the trajectory over the entire time dimension. Its input sequence... This is the output of a temporal convolutional network, capturing long-term trajectory evolution trends. The formula is as follows:

[0101] ;

[0102] The temporal evolution parsing layer retains long-term valid information through gating mechanisms (input gate, forget gate, output gate) and outputs the hidden state at each time step. ), It has a future predicted state expressed through features of historical trajectory data, and this information provides the basis for the behavioral attention module with temporal contextual representation.

[0103] In this embodiment, in order to introduce the guiding effect of behavioral state on prediction, a behavioral attention mechanism is designed. First, feature constraint information ( ) through the embedding matrix ( Mapped to the behavior embedding vector space ( ), The calculation formula is as follows:

[0104] ;

[0105] Then, use The function behavior embedding vector space is fused with the intermediate hidden states output by the temporal evolution parsing layer to form behavior-aware features. ), The calculation formula is as follows:

[0106] ;

[0107] in, Represents the fusion function. express Transpose of;

[0108] Fusion state ( At the same time, it preserves the temporal structure and behavioral semantic information of the trajectory. This state not only includes the temporal sequence structural features of the trajectory, but also injects the guiding tendency of the UAV's behavioral features, thereby enhancing the ability of the prediction output to adapt to changes in the UAV's behavioral features.

[0109] Finally, merge the states of all time steps ( The input is fed into the fully connected prediction layer to predict the coordinates of future locations. The prediction can be expanded over multiple future time steps and used for spatial judgment in the collision detection module. The calculation formula is as follows:

[0110] .

[0111] Furthermore, after step S20, the method further includes obtaining a joint loss function (designing a joint loss function that comprehensively considers position error and motion continuity error as indicators, using the Adam (Adaptive Moment Estimation) optimizer for parameter training, and employing a piecewise decay strategy for the learning rate; where position error refers to the Euclidean distance error between the UAV trajectory points predicted by the model and the actual trajectory points, and motion continuity error refers to the smoothness and continuity of the trajectory measured by changes in velocity and acceleration); constraining and optimizing the hybrid trajectory prediction neural network model based on the position error loss and motion continuity loss of the joint loss function to obtain a first hybrid trajectory prediction neural network model; and performing gradient optimization on the first hybrid trajectory prediction neural network model according to a preset dynamic correction rule to obtain an optimal hybrid trajectory prediction neural network model, wherein the optimal hybrid trajectory prediction neural network model is used to make high-precision predictions of the future trajectory of the UAV.

[0112] In this embodiment, after processing the spatiotemporal fusion input sequence using a hybrid neural network model, the output result of the prediction stage is obtained. This output sequence represents a sequence of three-dimensional spatial location points within a future time step, used to reflect the future trajectory of the UAV. Let the prediction step size be... Then the output sequence The definition is as follows:

[0113] ;

[0114] in, This indicates that the model predicts the drone's performance in the future. Similarly, the position coordinates at each time step... This indicates that the model predicts the drone's performance in the future. The position coordinates of each time step. This indicates that the model predicts the drone's performance in the future. Position coordinates at each time step ( ), middle Pick , and .

[0115] Furthermore, to achieve the dual constraints of trajectory prediction accuracy and smoothness, a joint loss function is adopted, consisting of two parts: position error loss (position accuracy) and motion continuity loss (trajectory smoothness).

[0116] ;

[0117] in, , The loss weight hyperparameter is used to balance prediction error and trajectory continuity. This represents the total loss value. Indicates position error loss, Indicates continuous loss of motion;

[0118] Position error loss is based on the Euclidean distance error between the actual trajectory point and the predicted point, which measures the magnitude of the position deviation in three-dimensional space and maintains the spatial consistency between the predicted trajectory and the target trajectory. Motion continuity loss is achieved by performing first-order difference operation on the velocity vector changes of adjacent time steps to punish the acceleration fluctuations of the predicted trajectory and prevent the trajectory from experiencing severe jitter or unreasonable abrupt changes.

[0119] Step S30: Establish multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, perform conflict detection on multiple three-dimensional hybrid bounding box models, and obtain conflict detection results.

[0120] Step S30 includes:

[0121] Step S31: Establish multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results;

[0122] Step S32: Obtain the outer axis-aligned static bounding boxes of the multiple 3D hybrid bounding box models. If any one of the multiple outer axis-aligned static bounding boxes has a collision risk, then perform collision detection on the multiple 3D hybrid bounding box models to obtain the collision detection result.

[0123] Specifically, such as Figure 5 As shown, multiple 3D hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results (taking a single 3D hybrid bounding box model as an example, the trajectory prediction point is the center point of the UAV to establish a 3D hybrid bounding box model: adopting a modeling mechanism that integrates static bounding boxes and dynamic semantic deformation bounding boxes to form a hybrid bounding box structure with hierarchical representation capabilities and behavioral adaptation; the trajectory prediction point refers to the predicted position point of the UAV, which includes timestamps and xyz axis position points). The outer axis-aligned static bounding boxes of the multiple 3D hybrid bounding box models are obtained (first layer: fast outer bounding box overlap detection). If any one of the multiple outer axis-aligned static bounding boxes has a collision risk, collision detection is performed on the multiple 3D hybrid bounding box models (second layer: precise matching detection guided by behavioral semantics) to obtain the conflict detection result.

[0124] In this embodiment, the first layer is fast bounding box overlap detection. Utilizing the fast speed of static bounding box detection, it enables rapid spatial overlap filtering among large-scale UAV groups. First, predictive timing encapsulation is performed. For the predicted timing of each UAV, the minimum bounding static axis alignment at that moment is constructed. AABB The bounding box is defined and stored in six-element boundary coordinates, which include... , , The maximum and minimum coordinate values ​​of the axes are then used for rapid boundary filtering. For any two drones, the comparison is made solely based on the numerical range along the coordinate axis. On each coordinate axis, if the maximum boundary coordinate value of one drone's bounding box is less than the minimum boundary coordinate value of another bounding box, and its minimum boundary coordinate value is less than the maximum boundary coordinate value of another bounding box, then the two static bounding boxes are considered to overlap on that coordinate axis. AABB Enclosure box , , For a collision to occur, overlapping areas must exist on all three coordinate axes. This generates candidate drone pairs with potential collision risks. Only when all three axes overlap is a potential collision risk identified and added to the candidate drone pair set for subsequent refined detection. The second layer involves precise matching detection guided by behavioral semantics. For the candidate drone pairs with potential collision risks selected in the first layer, this stage further combines their trajectory evolution features with behavioral semantic information to perform time-series inner-layer dynamic semantic deformation bounding boxes. Fine-grained potential conflict identification and detection. First, behavior-driven modeling is performed, generating corresponding inner dynamic semantic deformation bounding boxes based on the predicted behavior labels of the UAV at the predicted time step. The bounding box size is dynamically adjusted based on the directional scaling factor corresponding to the behavior label to better match the current flight trend and behavior state. Then, discrimination is performed step by step. For candidate UAV pairs, the Fast Separating Axis Projection (SAT) algorithm is used step by step at each prediction time step to determine spatial intersection. SAT needs to perform projection overlap judgment on 9-15 directional axes, record the time segments in which bounding boxes overlap continuously, and judge whether there is a potential conflict risk.

[0125] Step S32 includes:

[0126] Step S321: Obtain the outer axis-aligned static bounding box and multiple oriented bounding boxes of the multiple 3D hybrid bounding box models;

[0127] Step S322: If any one of the multiple outer axis-aligned static bounding boxes has a collision risk, then the multiple directed bounding boxes are dynamically adjusted according to the semantic features to obtain multiple inner dynamic semantic deformation bounding boxes.

[0128] Step S323: Perform collision detection on multiple three-dimensional hybrid bounding box models based on multiple inner-layer dynamic semantic deformation bounding boxes to obtain collision detection results.

[0129] In this embodiment, the outer axis-aligned static bounding boxes of multiple 3D hybrid bounding box models are obtained. AABB Axis-aligned bounding box and multiple oriented bounding boxes OBB If any one of the multiple outer axis-aligned static bounding boxes has a collision risk, then the multiple oriented bounding boxes are dynamically adjusted according to the semantic features to obtain multiple inner dynamic semantic deformation bounding boxes. Collision detection is performed on multiple three-dimensional hybrid bounding box models based on multiple inner-layer dynamic semantic deformation bounding boxes to obtain conflict detection results.

[0130] In this embodiment, the outer axis-aligned static bounding box is mainly used to encapsulate the spatial position of the UAV at each prediction time with a fixed size, which facilitates rapid preliminary spatial occupancy detection. First, the predicted trajectory points of the continuous trajectory prediction results are selected. As the center point of the current bounding box, it is then established based on the UAV laser point cloud. AABB The bounding box size is determined, and principal component analysis is used to extract the principal directions of the point cloud family. The covariance matrix is ​​calculated, and eigenvectors are obtained. The principal axis direction is determined by the eigenvector corresponding to the largest eigenvalue. The point cloud is then projected onto the principal axis coordinate system, and calculations are performed in the principal axis coordinate system. , , The bounding box dimensions are obtained from the minimum and maximum boundary values ​​in the three axes. , , Complete the construction of the axis-aligned static bounding box.

[0131] For example, considering the impact of drone behavior on future space occupation patterns, it is necessary to construct an inner dynamic semantic deformation bounding box based on behavior guidance on top of the static bounding box. Firstly, the dynamic bounding box adopts a directed bounding box (…). OBB The basic structure (oriented bounding box) rotates in space around the heading direction to closely follow the actual movement trend of the drone. Then, behavioral semantic injection and orientation adjustment are performed to obtain the prediction points (…). ) feature constraint information ( ) and the current velocity vector and heading angle of the motion direction, and finally, based on the behavior label and velocity direction, define the forward extension length factor of the bounding box in the main motion direction ( ) Different behaviors correspond to different deformation ratios; for example, for "sharp turns," the lateral buffer should be appropriately increased. , , The formula for calculating the three-axis adjustment ratio is as follows:

[0132] ;

[0133] ;

[0134] ;

[0135] in, express The dimensions of the enclosure box after shaft adjustment express The dimensions of the enclosure box after shaft adjustment express The dimensions of the enclosure after shaft adjustment;

[0136] Finally, static With dynamics Within each prediction time step, a three-dimensional hybrid bounding box structure is formed ( ).

[0137] Furthermore, after step S30, the method further includes obtaining the spatial intersection pattern and heading angle relationship of the continuous trajectory prediction results, performing scene judgment on the conflict detection results based on the spatial intersection pattern and heading angle relationship to obtain the target conflict scene type, obtaining the minimum spatiotemporal distance and conflict duration of the conflict detection results, obtaining the conflict level based on the minimum spatiotemporal distance and conflict duration (judging the conflict scene and evaluating the conflict level of the detection results, and providing rapid feedback through structured data units), and issuing an early warning based on the conflict level and the target conflict scene type to notify the user to perform real-time detection of the drone. The target conflict scene type includes any one of head-on collision, cross-traverse, and same-direction tail-end collision; the conflict level includes any one of high, medium, and low.

[0138] In this embodiment, the UAV pairs with collision risk identified in the above-mentioned layered conflict detection results are first classified into conflict scenarios and hazard levels. Based on the predicted spatial intersection patterns and heading angles of both sides, the conflict scenario type is determined, resulting in the target conflict scenario type. The target conflict scenario types mainly include: head-on collision (opposite heading angles, concentrated intersection points); cross-crossing (trajectories intersect in space, large angle between flight directions); and same-direction tail-end collision (trajectories are in the same direction, showing a chasing trend). Based on the minimum spatiotemporal distance and conflict duration, the conflict level is divided into high, medium, and low risk levels. The minimum spatiotemporal distance is determined by whether the closest approach distance is below a safety threshold, and the conflict duration is calculated by determining the duration of the predicted bounding box intersection within a continuous time step.

[0139] Furthermore, such as Figure 6 As shown, based on the above-mentioned trajectory prediction-based low-altitude UAV conflict detection method, the present invention also provides a trajectory prediction-based low-altitude UAV conflict detection system, wherein the trajectory prediction-based low-altitude UAV conflict detection system includes:

[0140] The feature segmentation module 51 is used to acquire the spatiotemporal flow trajectory data of the UAV, preprocess the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and perform behavioral feature segmentation on the target spatiotemporal flow trajectory data according to preset segmentation rules to obtain behavioral feature constraint information.

[0141] The trajectory prediction module 52 is used to fuse the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, and input the spatiotemporal fusion input sequence into the hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period.

[0142] The collision detection module 53 is used to establish multiple three-dimensional hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, perform collision detection on the multiple three-dimensional hybrid bounding box models, and obtain collision detection results.

[0143] Furthermore, such as Figure 7 As shown, based on the above-mentioned low-altitude UAV conflict detection method and system based on trajectory prediction, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 7 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0144] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard disk or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a low-altitude UAV conflict detection program 40 based on trajectory prediction, which can be executed by the processor 10 to implement the low-altitude UAV conflict detection method based on trajectory prediction in this application.

[0145] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the low-altitude UAV conflict detection method based on trajectory prediction.

[0146] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The terminals communicate with each other via a system bus.

[0147] In one embodiment, when the processor 10 executes the low-altitude UAV conflict detection program 40 based on trajectory prediction in the memory 20, the following steps are performed:

[0148] Acquire the spatiotemporal flow trajectory data of the UAV, preprocess the spatiotemporal flow trajectory data to obtain the target spatiotemporal flow trajectory data, and divide the target spatiotemporal flow trajectory data into behavioral features according to a preset division rule to obtain behavioral feature constraint information;

[0149] The target spatiotemporal flow trajectory data and the behavioral feature constraint information are fused to obtain a spatiotemporal fusion input sequence. The spatiotemporal fusion input sequence is then input into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period.

[0150] Multiple 3D hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results. Collision detection is performed on multiple 3D hybrid bounding box models to obtain collision detection results.

[0151] The preprocessing includes standardization, completion, anomaly detection, and rejection.

[0152] The process of acquiring the spatiotemporal flow trajectory data of the UAV, preprocessing the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and dividing the target spatiotemporal flow trajectory data into behavioral features according to preset segmentation rules to obtain behavioral feature constraint information specifically includes:

[0153] Acquire the spatiotemporal trajectory data and current timestamp of the UAV, and standardize the geographic coordinate system of the spatiotemporal trajectory data according to the current timestamp to obtain the projected coordinates;

[0154] The spatiotemporal flow trajectory data is completed by using a linear interpolation method based on the projected coordinates to obtain the first spatiotemporal flow trajectory data;

[0155] Anomaly detection is performed on the first spatiotemporal flow trajectory data using the local outlier factor algorithm to obtain abnormal trajectories. The first spatiotemporal flow trajectory data is then removed based on the abnormal trajectories to obtain target spatiotemporal flow trajectory data. The target spatiotemporal flow trajectory data is then divided into behavioral features according to a preset partitioning rule to obtain behavioral feature constraint information.

[0156] The geographic coordinate system includes longitude, latitude, and altitude;

[0157] The behavioral characteristic constraint information includes takeoff characteristics, landing characteristics, rapid ascent characteristics, dive characteristics, cruise characteristics, turning characteristics, sharp turn characteristics, hovering characteristics, emergency stop characteristics, and circling characteristics.

[0158] Specifically, the process of fusing the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, and then inputting the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain multiple continuous trajectory prediction results within a future time period, includes:

[0159] The behavioral feature constraint information is mapped according to a preset one-hot encoding to obtain semantic features;

[0160] The target spatiotemporal flow trajectory data and the semantic features are fused according to the preset one-hot encoding to obtain a spatiotemporal fusion input sequence;

[0161] A hybrid trajectory prediction neural network model is constructed by fusing a temporal convolutional perception layer, a temporal evolution parsing layer, and a behavioral semantic fusion layer. The spatiotemporal fusion input sequence is then input into the hybrid trajectory prediction neural network model for prediction, resulting in predictions of multiple continuous trajectories within a future time period.

[0162] The process includes fusing the target spatiotemporal flow trajectory data and the behavioral feature constraint information to obtain a spatiotemporal fusion input sequence, inputting the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction, and obtaining multiple continuous trajectory prediction results within a future time period. The process further includes:

[0163] Obtain the joint loss function, and perform constraint optimization on the hybrid trajectory prediction neural network model based on the position error loss and motion continuity loss of the joint loss function to obtain the first hybrid trajectory prediction neural network model.

[0164] The first hybrid trajectory prediction neural network model is subjected to gradient optimization according to the preset dynamic correction rule to obtain the optimal hybrid trajectory prediction neural network model.

[0165] The optimal hybrid trajectory prediction neural network model is used to make high-precision predictions of the future trajectory of the UAV.

[0166] Specifically, the step of establishing multiple 3D hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, and performing collision detection on the multiple 3D hybrid bounding box models to obtain collision detection results includes:

[0167] Multiple three-dimensional hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results;

[0168] Obtain the outer axis-aligned static bounding boxes of multiple 3D hybrid bounding box models. If any one of the outer axis-aligned static bounding boxes has a collision risk, then perform collision detection on the multiple 3D hybrid bounding box models to obtain the collision detection result.

[0169] Specifically, the step of establishing multiple 3D hybrid bounding box models based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results, and performing collision detection on the multiple 3D hybrid bounding box models to obtain collision detection results includes:

[0170] Multiple three-dimensional hybrid bounding box models are established based on multiple trajectory prediction points corresponding to multiple continuous trajectory prediction results;

[0171] Obtain the outer axis-aligned static bounding boxes of multiple 3D hybrid bounding box models. If any one of the outer axis-aligned static bounding boxes has a collision risk, then perform collision detection on the multiple 3D hybrid bounding box models to obtain the collision detection result.

[0172] The process includes establishing multiple 3D hybrid bounding box models based on multiple trajectory prediction points corresponding to the multiple continuous trajectory prediction results, performing collision detection on the multiple 3D hybrid bounding box models to obtain collision detection results, and then further including:

[0173] The spatial intersection pattern and heading angle relationship of the continuous trajectory prediction results are obtained, and the scene judgment is performed on the conflict detection results based on the spatial intersection pattern and the heading angle relationship to obtain the target conflict scene type;

[0174] Obtain the minimum spatiotemporal distance and conflict duration of the conflict detection results, and obtain the conflict level based on the minimum spatiotemporal distance and the conflict duration;

[0175] Warnings are issued based on the conflict level and the target conflict scenario type to notify users to conduct real-time monitoring of the drone;

[0176] The target conflict scenario types include any one of head-on collision, cross-traffic, and same-direction rear-end collision;

[0177] The conflict level includes any one of high, medium, and low.

[0178] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a low-altitude UAV conflict detection program based on trajectory prediction, and the low-altitude UAV conflict detection program based on trajectory prediction implements the steps of the low-altitude UAV conflict detection method based on trajectory prediction as described above when executed by a processor.

[0179] In summary, this invention provides a method, system, terminal, and storage medium for low-altitude UAV conflict detection based on trajectory prediction. The method includes: acquiring spatiotemporal flow trajectory data of the UAV; preprocessing the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data; and dividing the target spatiotemporal flow trajectory data into behavioral features according to preset division rules to obtain behavioral feature constraint information; fusing the target spatiotemporal flow trajectory data and behavioral feature constraint information to obtain a spatiotemporal fusion input sequence; inputting the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain multiple trajectory prediction points; establishing multiple three-dimensional hybrid bounding box models based on the multiple trajectory prediction points; performing conflict detection on the multiple three-dimensional hybrid bounding box models to obtain conflict detection results. This invention first acquires UAV spatiotemporal flow trajectory data, performs trajectory data preprocessing, and divides the UAV trajectory into motion features; secondly, it establishes a deep learning trajectory prediction model with spatiotemporal feature fusion to predict the UAV trajectory; finally, it designs a hierarchical and progressive hybrid bounding box collision detection method to discriminate the conflict detection results and provide rapid feedback, greatly improving the accuracy and real-time performance of multi-UAV flight conflict detection in complex airspace.

[0180] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal system that includes that element.

[0181] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0182] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for detecting conflict of low-altitude unmanned aerial vehicle based on trajectory prediction, characterized in that, The low-altitude unmanned aerial vehicle conflict detection method based on trajectory prediction comprises: Obtain the spatiotemporal flow trajectory data of the unmanned aerial vehicle, preprocess the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and divide the target spatiotemporal flow trajectory data according to a preset division rule to obtain behavior feature constraint information; Fuse the target spatiotemporal flow trajectory data and the behavior feature constraint information to obtain a spatiotemporal fusion input sequence, input the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain a plurality of continuous trajectory prediction results in a future time period; Establish a plurality of three-dimensional hybrid bounding box models according to a plurality of trajectory prediction points corresponding to the plurality of continuous trajectory prediction results, perform conflict detection on the plurality of three-dimensional hybrid bounding box models to obtain a conflict detection result; The fusing of the target spatiotemporal flow trajectory data and the behavior feature constraint information to obtain a spatiotemporal fusion input sequence, the inputting of the spatiotemporal fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain a plurality of continuous trajectory prediction results in a future time period specifically comprises: Map the behavior feature constraint information according to a preset one-hot encoding to obtain semantic features; Fuse the target spatiotemporal flow trajectory data and the semantic features according to the preset one-hot encoding to obtain a spatiotemporal fusion input sequence; Fuse a time series convolution perception layer, a time series evolution analysis layer, and a behavior semantic fusion layer to construct a hybrid trajectory prediction neural network model, input the spatiotemporal fusion input sequence into the hybrid trajectory prediction neural network model for prediction to obtain a plurality of continuous trajectory prediction results in a future time period; The establishment of a plurality of three-dimensional hybrid bounding box models according to a plurality of trajectory prediction points corresponding to the plurality of continuous trajectory prediction results, the conflict detection on the plurality of three-dimensional hybrid bounding box models to obtain a conflict detection result specifically comprises: Establish a plurality of three-dimensional hybrid bounding box models according to a plurality of trajectory prediction points corresponding to the plurality of continuous trajectory prediction results; Obtain outer axis-aligned bounding boxes of the plurality of three-dimensional hybrid bounding boxes, and if there is a collision risk in any one of the outer axis-aligned bounding boxes, perform collision detection on the plurality of three-dimensional hybrid bounding boxes to obtain a conflict detection result. 2.The trajectory prediction based low altitude UAV conflict detection method according to claim 1, characterized in that, The preprocessing comprises standardization processing, completion processing, anomaly detection, and rejection processing; The obtaining of the spatiotemporal flow trajectory data of the unmanned aerial vehicle, the preprocessing of the spatiotemporal flow trajectory data to obtain target spatiotemporal flow trajectory data, and the division of the target spatiotemporal flow trajectory data according to a preset division rule to obtain behavior feature constraint information specifically comprises: Obtain the spatiotemporal flow trajectory data and a current timestamp of the unmanned aerial vehicle, standardize the geographic coordinate system of the spatiotemporal flow trajectory data according to the current timestamp to obtain projected coordinates; Complete the spatiotemporal flow trajectory data according to the projected coordinates by a linear interpolation method to obtain first spatiotemporal flow trajectory data; Anomaly detection is performed on the first spatio-temporal flow trajectory data by a local outlier factor algorithm to obtain an abnormal trajectory, the first spatio-temporal flow trajectory data is processed by removing the abnormal trajectory to obtain target spatio-temporal flow trajectory data, and the target spatio-temporal flow trajectory data is divided into behavior characteristics according to a preset division rule to obtain behavior characteristic constraint information; The geographic coordinate system includes longitude, latitude, and altitude. The behavior characteristic constraint information includes take-off characteristics, landing characteristics, rapid ascent characteristics, diving characteristics, cruising characteristics, turning characteristics, rapid turning characteristics, hovering characteristics, rapid stopping characteristics, and circling characteristics. 3.The trajectory prediction based low altitude UAV conflict detection method according to claim 1, characterized in that, The target spatio-temporal flow trajectory data and the behavior characteristic constraint information are fused to obtain a spatio-temporal fusion input sequence, the spatio-temporal fusion input sequence is input into a hybrid trajectory prediction neural network model for prediction to obtain a plurality of continuous trajectory prediction results in a future time period, and then the method further includes: A joint loss function is obtained, and the hybrid trajectory prediction neural network model is constrained and optimized according to a position error loss and a motion continuity loss of the joint loss function to obtain a first hybrid trajectory prediction neural network model; The first hybrid trajectory prediction neural network model is gradient-optimized according to a preset dynamic correction rule to obtain an optimal hybrid trajectory prediction neural network model. The optimal hybrid trajectory prediction neural network model is used to predict a future trajectory of a UAV with high precision. 4.The trajectory prediction based low altitude UAV conflict detection method according to claim 1, characterized in that, If any one of the plurality of outer axis-aligned bounding boxes has a collision risk, collision detection is performed on the plurality of three-dimensional hybrid bounding box models to obtain a conflict detection result, specifically including: The outer axis-aligned bounding boxes of the plurality of three-dimensional hybrid bounding box models and a plurality of directional bounding boxes are obtained. If any one of the plurality of outer axis-aligned bounding boxes has a collision risk, the plurality of directional bounding boxes are dynamically adjusted according to the semantic features to obtain a plurality of inner dynamic semantic deformation bounding boxes. Collision detection is performed on the plurality of three-dimensional hybrid bounding box models according to the plurality of inner dynamic semantic deformation bounding boxes to obtain a conflict detection result.

5. The trajectory prediction based low altitude UAV conflict detection method according to claim 1, wherein, The plurality of three-dimensional hybrid bounding box models are established based on a plurality of trajectory prediction points corresponding to the plurality of continuous trajectory prediction results, and conflict detection is performed on the plurality of three-dimensional hybrid bounding box models to obtain a conflict detection result, and then the method further includes: The spatial intersection mode and the heading angle relationship of the continuous trajectory prediction result are obtained, the conflict detection result is judged according to the spatial intersection mode and the heading angle relationship to obtain a target conflict scene type; The minimum spatio-temporal distance and the conflict duration of the conflict detection result are obtained, and a conflict level is obtained according to the minimum spatio-temporal distance and the conflict duration; The conflict level and the target conflict scene type are used for early warning to notify a user to perform real-time detection on the UAV. The target conflict scenario type includes any one of positive hedging, cross crossing, and same-direction rear-end collision. The conflict level includes any one of high, medium, and low. 6.A low altitude unmanned aerial vehicle collision detection system based on trajectory prediction, characterized in that, The low-altitude unmanned aerial vehicle conflict detection system based on trajectory prediction is applied to the low-altitude unmanned aerial vehicle conflict detection method based on trajectory prediction in any one of claims 1-5, and the low-altitude unmanned aerial vehicle conflict detection system based on trajectory prediction comprises: A feature division module is configured to obtain space-time flow trajectory data of an unmanned aerial vehicle, pre-process the space-time flow trajectory data to obtain target space-time flow trajectory data, and divide the target space-time flow trajectory data according to a preset division rule to obtain behavior feature constraint information. A trajectory prediction module is configured to fuse the target space-time flow trajectory data and the behavior feature constraint information to obtain a space-time fusion input sequence, input the space-time fusion input sequence into a hybrid trajectory prediction neural network model for prediction to obtain a plurality of continuous trajectory prediction results in a future time period. A conflict detection module is configured to establish a plurality of three-dimensional hybrid bounding box models according to a plurality of trajectory prediction points corresponding to the plurality of continuous trajectory prediction results, perform conflict detection on the plurality of three-dimensional hybrid bounding box models, and obtain a conflict detection result.

7. A terminal, characterized by comprising: The terminal comprises a memory, a processor, and a low-altitude unmanned aerial vehicle conflict detection program based on trajectory prediction stored on the memory and executable on the processor, and the low-altitude unmanned aerial vehicle conflict detection program based on trajectory prediction implements the steps of the low-altitude unmanned aerial vehicle conflict detection method based on trajectory prediction in any one of claims 1-5 when executed by the processor.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a low-altitude unmanned aerial vehicle conflict detection program based on trajectory prediction, and the low-altitude unmanned aerial vehicle conflict detection program based on trajectory prediction implements the steps of the low-altitude unmanned aerial vehicle conflict detection method based on trajectory prediction in any one of claims 1-5 when executed by the processor.

Images