An unmanned aerial vehicle conflict detection method under an incomplete perception state

Abstract

The application discloses a kind of unmanned aerial vehicle conflict detection methods under the condition of incomplete perception.First, based on the history trajectory database of unmanned aerial vehicle, large-scale trajectory segment is constructed, the behavior characteristics of unmanned aerial vehicle is extracted from history trajectory segment by space-time trajectory clustering, and a typical motion state category set is constructed.Second, the motion sequence of surrounding non-cooperative unmanned aerial vehicle in short-time observation time window is obtained, and the current motion state category is identified by the trajectory state recognition model of pre-trained convolutional neural network.Then, short-time observation motion sequence and current motion state category are input into trajectory prediction model based on Transformer, and the probability that non-cooperative unmanned aerial vehicle falls into each horizontal grid and each vertical height interval in future prediction time window is predicted.Finally, combined with the known track or planned track of target unmanned aerial vehicle, horizontal interval loss probability and vertical interval loss probability are calculated to determine whether there is conflict risk.The application improves the reliability of unmanned aerial vehicle conflict detection under the condition of incomplete perception.

Description

Technical Field

[0001] This invention belongs to the field of UAV operation safety and conflict detection technology, specifically relating to a UAV conflict detection method under incomplete perception conditions. Background Technology

[0002] With the increasing density of low-altitude airspace operations, drones are being used more and more widely in scenarios such as logistics delivery, urban patrol, emergency rescue, and security monitoring. Detection of aerial conflicts between target drones and surrounding drones has become a crucial technical aspect of ensuring safe low-altitude operations. Existing conflict detection methods typically rely on relatively complete target status information, such as real-time position, speed, heading, altitude, identity information, communication coordination data, or pre-set flight plans. However, in actual low-altitude operating environments, surrounding drones may not possess communication coordination capabilities or be unable to continuously provide the target drone with complete identity, flight plan, and motion intent information. The target drone can only obtain short-term, localized observation data through onboard sensors or external detection equipment. That is, in an incomplete perception state, when a target drone detects surrounding non-cooperative drones, it cannot obtain their complete flight status, flight intent, and future trajectory information; it can only infer their current motion state and future position changes based on local motion sequences within a short observation window.

[0003] To address the aforementioned issues, this invention proposes a UAV conflict detection method under incomplete perception conditions. By identifying the current motion state of non-cooperative UAVs, predicting their future spatiotemporal position probability distribution, and calculating the risk of interval loss based on horizontal and vertical interval constraints, the reliability and applicability of conflict detection are improved. Summary of the Invention

[0004] To address the aforementioned problems, this invention discloses a method for detecting drone conflicts under incomplete perception conditions. This method obtains a set of typical drone motion state categories through historical trajectory clustering, identifies the current motion state of non-cooperative drones using a convolutional neural network, and inputs short-term observed motion sequences and the current motion state category into a Transformer-based trajectory prediction model. The model outputs the probability that a non-cooperative drone will fall into each horizontal grid and each vertical altitude interval within a future prediction time window. Furthermore, based on the target drone's trajectory, the horizontal interval loss probability, the vertical interval loss probability, and the comprehensive interval loss risk are calculated to complete the conflict detection between the target drone and surrounding non-cooperative drones.

[0005] This invention aims to solve the following technical problems:

[0006] Under incomplete perception conditions, the target UAV can only obtain limited motion observation information such as the position, altitude, speed and time of surrounding non-cooperative UAVs, and it is difficult to directly obtain their flight plan, motion intention and future trajectory information;

[0007] The current motion status of surrounding non-cooperative drones cannot be directly given by cooperative information and needs to be inferred based on the motion sequence within a short observation window, which affects the accuracy of subsequent trajectory prediction.

[0008] The future movement of non-cooperative drones is uncertain. If only deterministic predicted trajectories are output, it is difficult to fully express the spatial range and probability differences that may occur within the future prediction time window.

[0009] For conflict detection between a target UAV and surrounding non-cooperative UAVs, it is necessary to transform the uncertainty prediction results of the future position of non-cooperative UAVs into horizontal interval loss probability, vertical interval loss probability, and comprehensive interval loss risk to support the determination of interval loss or conflict risk.

[0010] To achieve the above-mentioned technical objectives and effects, the present invention is implemented through the following technical solutions.

[0011] A method for UAV conflict detection under incomplete perception conditions includes the following steps:

[0012] S1. Construct large-scale trajectory segments based on the historical trajectory database of UAVs, and extract UAV behavior features from the historical trajectory segments through spatiotemporal trajectory clustering to construct a set of typical motion state categories:

[0013] The historical trajectories in the drone historical trajectory database are divided into multiple trajectory segments according to a preset sliding time window. Let the segment be... Each trajectory segment is represented as:

[0014]

[0015] in, Indicates the first The first trajectory segment Each trajectory point includes its horizontal position coordinates, height, velocity, and direction of motion.

[0016] Based on the three-dimensional position changes and time intervals between adjacent trajectory points, calculate the velocity components in three spatial directions and the corresponding cosines of the motion direction changes to obtain the first... Behavioral feature vectors corresponding to each trajectory segment:

[0017]

[0018] in, Indicates the first The first one in the behavioral feature vector corresponding to the trajectory segment 1 eigenvalue, The feature vector contains the dimension of the feature values. The feature vector is composed of the three-dimensional velocity components and the cosine of the change in motion direction of two adjacent motion units. To measure the motion similarity between different trajectory segments, the similarity between any two trajectory segments is calculated. and Correlation coefficients between corresponding behavioral feature vectors:

[0019]

[0020] In the formula, For the first The trajectory segment and the first The correlation coefficient between trajectory segments and The first The and the first The trajectory segment at the ... The values ​​that can be taken on each behavioral feature and The first The and the first The mean of each behavioral feature vector is used. A trajectory segment similarity matrix is ​​constructed based on the correlation coefficient, and spatiotemporal trajectory clustering is performed on multiple trajectory segments based on the similarity matrix. Spatiotemporal trajectory clustering includes: determining candidate cluster centers based on the similarity matrix; calculating the correlation coefficient between the remaining trajectory segments and each candidate cluster center; assigning each trajectory segment to the category corresponding to the candidate cluster center with the highest correlation coefficient; and determining the number of typical motion state categories based on the density and variation of similar trajectory segments near the cluster centers, thus constructing a set of typical motion state categories.

[0021]

[0022] in, The number of typical motion state categories (such as climbing, descending, turning, etc.) is used to represent different motion modes of the UAV obtained by spatiotemporal trajectory clustering.

[0023] S2. Obtain the motion sequence of surrounding non-cooperative drones within a short observation window, construct the current trajectory segment, and input the current trajectory segment into a pre-trained convolutional neural network trajectory state recognition model to obtain the current motion state category of the non-cooperative drones:

[0024] Historical trajectory segments and their typical motion state categories determined by spatiotemporal trajectory clustering are used as training samples. Each historical trajectory segment is converted into a trajectory segment feature matrix, and each typical motion state category is used as a supervision label to classify and train a convolutional neural network, thus obtaining a convolutional neural network trajectory state recognition model.

[0025] Let the short-time observation window length be... The information obtained regarding surrounding non-cooperative drones at the current moment The corresponding motion sequence is represented as:

[0026]

[0027] in, Indicates time The observation status includes at least the horizontal position coordinates, altitude, speed, and direction of motion of the non-cooperative UAV. The motion sequence... The features are organized chronologically into a trajectory segment feature matrix, and then input into a pre-trained convolutional neural network trajectory state recognition model to obtain the probability that the current trajectory segment belongs to each typical motion state category:

[0028]

[0029] In the formula, The current trajectory segment belongs to the first... The probability of a typical motion state category. This is the feature matrix corresponding to the current trajectory segment. This is the set of classification parameters for the output layer of a convolutional neural network. This is the feature representation extracted by the convolutional neural network from the current trajectory segment. This represents the total number of typical motion state categories. Based on the obtained probabilities, the typical motion state category with the highest probability is determined as the current motion state category of the non-cooperative drone.

[0030]

[0031] in, To identify the current motion state category.

[0032] S3. Input the motion sequence and current motion state category within the short-term observation window into the Transformer-based trajectory prediction model to predict the probability of the non-cooperative UAV falling into each horizontal grid and each vertical height interval in the future prediction window, respectively, and obtain the spatiotemporal position probability distribution prediction of the non-cooperative UAV:

[0033] Classify the current motion state Mapped to state embedding vectors:

[0034]

[0035] in, This represents the state embedding mapping function. The Transformer-based trajectory prediction model encodes temporal features from the motion sequence and state embedding to obtain the predicted features:

[0036]

[0037] Within the future prediction time window, let the set of prediction step sizes be:

[0038]

[0039] in, The number of future prediction moments. The pre-defined horizontal prediction area around the target is divided into... The preset vertical prediction range is divided into several horizontal grids. A vertical height interval. For any future prediction time... The trajectory prediction model outputs unnormalized prediction values ​​for the horizontal grid based on the motion sequence within a short observation window and the current motion state category. Unnormalized predicted values ​​for vertical height range And by normalization, the probabilities of non-cooperative drones falling into each horizontal grid and each vertical height interval at each predicted time in the future are obtained:

[0040]

[0041]

[0042] In the formula, For non-cooperative drones, predicting future moments Falling into The probability of a horizontal grid. For non-cooperative drones, predicting future moments Falling into The probability of a vertical height interval and These represent the predicted values ​​output by the trajectory prediction model for the corresponding horizontal grid and vertical height intervals, respectively. and These represent the unnormalized output values ​​corresponding to each horizontal grid and each vertical height interval traversed during the normalization process. and These represent the number of grid cells in the two horizontal directions of the preset horizontal prediction region. Let be the number of vertical height intervals. The above probabilities satisfy:

[0043]

[0044]

[0045] The spatiotemporal location probability distribution prediction of non-cooperative UAVs is constructed from the horizontal grid probability distribution and the vertical height interval probability distribution corresponding to each prediction time within the future prediction time window.

[0046] S4. Based on the known or planned flight path of the target UAV and the spatiotemporal position probability distribution prediction of non-cooperative UAVs, calculate the horizontal and vertical gap loss probabilities within the future prediction time window to determine the comprehensive gap loss risk, thereby determining whether there is a risk of gap loss or conflict between the target UAV and surrounding non-cooperative UAVs, and completing the conflict detection between the target UAV and surrounding non-cooperative UAVs:

[0047] Based on the target UAV's known or planned flight path, determine the target UAV's predicted future time. Location:

[0048]

[0049] in, , and These are the horizontal position coordinates and altitude of the target UAV at the predicted future time, respectively. Let the preset horizontal safety interval be... The preset vertical safety interval is Then, the horizontal risk area is determined centered on the future location of the target drone:

[0050]

[0051] And determine the vertical risk range:

[0052]

[0053] Based on the probability that non-cooperative drones will fall into each horizontal grid at a future predicted time, the probabilities of the horizontal grids that intersect with the horizontal risk area are accumulated to obtain the horizontal interval loss probability:

[0054]

[0055] In the formula, For the first One horizontal grid, This is a horizontal risk area. This represents the probability that a non-cooperative drone will enter the horizontal safe distance range of the target drone at a predicted future time. Based on the probability of the non-cooperative drone falling into each vertical altitude range, the probabilities of the vertical altitude ranges that intersect with the vertical risk ranges are summed to obtain the vertical distance loss probability:

[0056]

[0057] In the formula, For the first A vertical height range, This represents the vertical risk range. This represents the probability that a non-cooperative drone will enter the target drone's vertical safe distance range at a predicted future time.

[0058] A joint determination is made based on the horizontal and vertical interval loss probabilities. If, within the future prediction time window, there exists at least one prediction moment where the horizontal interval loss probability exceeds a preset horizontal risk threshold:

[0059]

[0060] Furthermore, the probability of vertical interval loss is greater than the preset vertical risk threshold:

[0061]

[0062] This determines that the target drone and surrounding non-cooperative drones face a comprehensive risk of gap loss, and further determines that there is a risk of gap loss or conflict between them within a future predicted time window. To preset the horizontal risk threshold, A preset vertical risk threshold is set. For multiple surrounding non-cooperative drones, the above probability prediction and interval loss probability calculation are performed respectively; when at least one surrounding non-cooperative drone meets the joint determination condition, it is determined that there is an interval loss or conflict risk between the target drone and surrounding non-cooperative drones, and the conflict detection is completed.

[0063] The beneficial effects of this invention are as follows: by identifying motion states through trajectory clustering and convolutional neural networks, the adaptability of trajectory prediction to motion patterns is improved; by using horizontal grids and vertical interval probabilities to characterize future position uncertainties and calculating interval loss risk accordingly, the reliability of collision detection under incomplete perception conditions is improved. Attached Figure Description

[0064] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort, wherein:

[0065] Figure 1 This is a flowchart of a UAV conflict detection method under incomplete perception state according to an embodiment of the present invention;

[0066] Figure 2 This is a schematic diagram of the distribution of the number of motion state category samples and typical trajectory segments according to an embodiment of the present invention;

[0067] Figure 3 This is a schematic diagram of a convolutional neural network trajectory state recognition model according to an embodiment of the present invention;

[0068] Figure 4 This is a three-dimensional hierarchical visualization of the horizontal and vertical probability distributions of future prediction times according to an embodiment of the present invention. Detailed Implementation

[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be noted that the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0070] In the description of this invention, the terms "comprising" and "including" should be interpreted broadly, that is, encompassing the content without excluding elements not explicitly listed. Furthermore, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0071] This invention provides a method for UAV conflict detection under incomplete perception conditions, such as... Figure 1 As shown, it includes the following steps:

[0072] (a) Construct large-scale trajectory segments based on the UAV historical trajectory database, extract UAV behavior features from the historical trajectory segments through spatiotemporal trajectory clustering, and construct a set of typical motion state categories;

[0073] (b) Obtain the motion sequence of surrounding non-cooperative UAVs within a short observation window, construct the current trajectory segment, and input the current trajectory segment into a pre-trained convolutional neural network trajectory state recognition model to obtain the current motion state category of the non-cooperative UAVs;

[0074] (c) Input the motion sequence and current motion state category within the short observation time window into the trajectory prediction model based on Transformer, and predict the probability of the non-cooperative UAV falling into each horizontal grid and the probability of falling into each vertical height interval in the future prediction time window, respectively, to obtain the spatiotemporal position probability distribution prediction of the non-cooperative UAV.

[0075] (d) Based on the known or planned flight path of the target UAV and the spatiotemporal position probability distribution prediction of non-cooperative UAVs, calculate the horizontal interval loss probability and vertical interval loss probability within the future prediction time window respectively, determine the comprehensive interval loss risk, and determine whether there is an interval loss or conflict risk between the target UAV and surrounding non-cooperative UAVs, and complete the conflict detection between the target UAV and surrounding non-cooperative UAVs.

[0076] In one specific embodiment of the present invention, such as Figure 2 As shown, the method includes:

[0077] The historical trajectories in the drone's historical trajectory database are resampled according to a preset sampling interval. In one embodiment, the sampling interval is 0.1 seconds, and a sliding time window trajectory segment is constructed using three consecutive trajectory points. Let the first... Each trajectory segment is represented as:

[0078]

[0079] in, Indicates the first The first trajectory segment Each trajectory point includes its horizontal position coordinates, height, velocity, and direction of motion.

[0080] Based on the three-dimensional position changes and time intervals between adjacent trajectory points, the velocity components of each motion unit in three spatial directions are calculated, and the corresponding cosines of motion direction changes are also calculated. In one embodiment, each motion unit extracts six behavioral attributes, including velocity components in three spatial directions and cosines of motion direction changes; two adjacent motion units are concatenated in chronological order to form a 12-dimensional behavioral feature vector.

[0081]

[0082] To measure the motion similarity between different trajectory segments, the correlation coefficient between the corresponding behavioral feature vectors of any two trajectory segments is calculated:

[0083]

[0084] In the formula, For the first The trajectory segment and the first The correlation coefficient between trajectory segments and The first The and the first The th behavioral feature vector 1 eigenvalue, and The first The and the first The mean of each behavioral feature vector. In this embodiment, .

[0085] A trajectory segment similarity matrix is ​​constructed based on the correlation coefficient, and candidate cluster centers are determined based on the similarity matrix. Each trajectory segment is then assigned to the category corresponding to the candidate cluster center with the highest correlation coefficient. In one embodiment, the number of cluster categories K is set to 35, resulting in a set of typical UAV motion state categories.

[0086]

[0087] In one specific embodiment of the present invention, such as Figure 3 As shown, the method includes:

[0088] Historical trajectory segments and their corresponding typical motion state categories are used as training samples, and each typical motion state category is used as a supervision label to classify and train the convolutional neural network, thus obtaining a convolutional neural network trajectory state recognition model.

[0089] In one embodiment, the input to the convolutional neural network is a feature matrix of a trajectory segment consisting of three consecutive trajectory points. Each trajectory point includes horizontal position coordinates, height, velocity, and direction of motion, thereby forming... The trajectory segment feature matrix is ​​obtained. After being input into the convolutional neural network, this matrix is ​​processed through convolutional layers, nonlinear activation layers, fully connected layers, and a classification output layer to output the probability that the current trajectory segment belongs to each typical motion state category.

[0090] For the currently detected non-cooperative drones in the surrounding area, let the short-time observation window length be... At the current moment The corresponding motion sequence is represented as:

[0091]

[0092] in, Indicates time The observed state. The motion sequence. Organize the current trajectory segments in chronological order and input them into the trained convolutional neural network trajectory state recognition model to obtain the current trajectory segment belonging to the [number missing]th [item missing]. The probability of each typical motion state category:

[0093]

[0094] In the formula, The current trajectory segment belongs to the first... The probability of a typical motion state category. This is the feature matrix corresponding to the current trajectory segment. This is the set of classification parameters for the output layer of a convolutional neural network. Features extracted by a convolutional neural network This represents the total number of typical motion state categories.

[0095] The most probable typical motion state category is determined as the current motion state category of the non-cooperative drone:

[0096]

[0097] In one specific embodiment of the present invention, such as Figure 4 As shown, the method includes:

[0098] The motion sequence within a short observation window and the current motion state category are input into a Transformer-based trajectory prediction model. The current motion state category is mapped to a state embedding vector through an embedding layer, and this vector, along with the motion sequence within the short observation window, is input into the Transformer model to extract short-term motion change features and state condition features of the non-cooperative UAV.

[0099] In one embodiment, the input length of the Transformer-based trajectory prediction model is 20 historical sampling points. When the sampling interval is 0.1s, the 20 historical sampling points correspond to a short-term observation window of 2s. The future prediction window of the trajectory prediction model includes 30 prediction steps, which, when the sampling interval is 0.1s, corresponds to a prediction range of the next 3s.

[0100] The preset horizontal prediction area surrounding the current reference position of the non-cooperative drone is divided into multiple horizontal grids, and the preset vertical prediction range is divided into multiple vertical height intervals. In one embodiment, with the current reference position of the non-cooperative drone as the center, a horizontal range of ±300m is divided into... The resolution is divided into A horizontal grid; the vertical range of ±150m is divided into 150 vertical height intervals at 2m intervals.

[0101] For any future prediction time The trajectory prediction model outputs unnormalized prediction values ​​for the horizontal grid and the vertical height interval, respectively, and obtains the probability of a non-cooperative drone landing in each horizontal grid by normalization. And the probability of falling into each vertical height range. The spatiotemporal location probability distribution prediction of non-cooperative UAVs is constructed from the horizontal grid probability distribution and the vertical altitude interval probability distribution corresponding to each prediction time within the future prediction time window.

[0102] To visually demonstrate the spatial relationship between the horizontal grid probability distribution and the vertical height interval probability distribution, every five consecutive 2m vertical height intervals are aggregated into a 10m drawing height layer. Under the approximate condition that the horizontal and vertical positions are independent, the probability of each 3D spatial unit is reconstructed according to the following formula:

[0103]

[0104] In the formula, For predicting future moments The reconstruction probability of a three-dimensional spatial unit For the first The sum of the probabilities of the original vertical height intervals contained in each 10m drawing height layer.

[0105] It should be noted that the preset horizontal prediction area range, horizontal grid resolution, preset vertical prediction range, and vertical height interval resolution mentioned above are values ​​for a specific embodiment. In other embodiments, these values ​​can be adjusted according to the target UAV detection range, the movement speed of non-cooperative UAVs, the prediction time window length, and the collision detection accuracy requirements. Height layer aggregation and display range filtering are only used for... Figure 4 The three-dimensional layered visualization does not change the original probability distribution output by the trajectory prediction model; the subsequent horizontal interval loss probability and vertical interval loss probability are calculated based on the complete horizontal grid probability distribution and the original 2m vertical height interval probability distribution, respectively.

[0106] In one specific embodiment of the present invention, the collision detection process based on horizontal and vertical intervals includes:

[0107] Based on the target UAV's known or planned flight path, determine the target UAV's predicted future time. The location of the target UAV is determined by centering on its horizontal position at that moment and determining the horizontal risk zone according to a preset horizontal safety interval; and by centering on its altitude at that moment and determining the vertical risk zone according to a preset vertical safety interval.

[0108] Based on the probability that non-cooperative drones will fall into each horizontal grid at a future predicted time, the probabilities of the horizontal grids that intersect with the horizontal risk area are accumulated to obtain the horizontal interval loss probability:

[0109]

[0110] in, For the first One horizontal grid, This is a horizontal risk area.

[0111] Based on the probability of non-cooperative drones falling into each vertical altitude range, the probabilities of vertical altitude ranges that intersect with the vertical risk ranges are summed to obtain the vertical interval loss probability:

[0112]

[0113] in, For the first A vertical height range This represents the vertical risk range.

[0114] If at least one prediction moment exists within the future prediction time window, such that the probability of horizontal interval loss is greater than a preset horizontal risk threshold and the probability of vertical interval loss is greater than a preset vertical risk threshold, it is determined that the target UAV and the surrounding non-cooperative UAVs have a comprehensive interval loss risk, and it is further determined that the two have an interval loss or conflict risk.

[0115] For multiple non-cooperative drones in the vicinity, the above-mentioned current motion state recognition, spatiotemporal position probability distribution prediction and interval loss risk determination processes are executed respectively; when at least one non-cooperative drone in the vicinity meets the risk determination conditions, it is determined that there is an interval loss or conflict risk between the target drone and the non-cooperative drones in the vicinity, and the conflict detection between the target drone and the non-cooperative drones in the vicinity is completed.

[0116] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for detecting UAV conflicts under incomplete perception conditions, characterized in that, Includes the following steps: (a) Construct a large-scale historical trajectory fragment based on the UAV historical trajectory database, extract UAV behavior features from the historical trajectory fragment through spatiotemporal trajectory clustering, and construct a set of typical motion state categories; (b) Obtain the motion sequence of surrounding non-cooperative UAVs within a short observation window, construct the current trajectory segment, and input the current trajectory segment into a pre-trained convolutional neural network trajectory state recognition model to obtain the current motion state category of the non-cooperative UAVs; (c) Input the motion sequence and current motion state category within the short observation time window into the trajectory prediction model based on Transformer, and predict the probability of the non-cooperative UAV falling into each horizontal grid and the probability of falling into each vertical height interval in the future prediction time window, respectively, to obtain the spatiotemporal position probability distribution prediction of the non-cooperative UAV. (d) Based on the known or planned flight path of the target UAV and the spatiotemporal position probability distribution prediction of non-cooperative UAVs, calculate the horizontal interval loss probability and vertical interval loss probability within the future prediction time window respectively, determine the comprehensive interval loss risk, and determine whether there is an interval loss or conflict risk between the target UAV and surrounding non-cooperative UAVs, and complete the conflict detection between the target UAV and surrounding non-cooperative UAVs.

2. The method according to claim 1, characterized in that, In step (a), the three-dimensional velocity components and cosines of motion direction changes of each trajectory segment are calculated to construct the corresponding behavioral feature vectors, and spatiotemporal trajectory clustering is performed based on the correlation coefficient between any two behavioral feature vectors: In the formula, For the first The trajectory segment and the first The correlation coefficient between trajectory segments For the first The first one in the behavioral feature vector corresponding to the trajectory segment 1 eigenvalue, For the first The th behavioral feature vector 1 eigenvalue, and The first The and the first The mean of each behavioral feature vector. The feature dimension of the behavior feature vector; Clustering of each trajectory segment based on the correlation coefficient forms a set of typical motion state categories for UAVs.

3. The method according to claim 1, characterized in that, In step (b), the historical trajectory segments and their corresponding typical motion state categories from step (a) are used as training data to pre-train a convolutional neural network, resulting in a convolutional neural network trajectory state recognition model. The motion sequences of surrounding non-cooperative UAVs within a short observation window are acquired, the current trajectory segment is constructed, and converted into a trajectory segment feature matrix. This matrix is ​​then input into the convolutional neural network trajectory state recognition model to obtain the probability that the current trajectory segment belongs to each typical motion state category. In the formula, The current trajectory segment belongs to the first... Typical motion state categories The probability, This is the feature matrix corresponding to the current trajectory segment. This is the set of classification parameters for the output layer of a convolutional neural network. Trajectory features extracted for convolutional neural network models. The total number of typical motion state categories; the typical motion state category with the highest probability is determined as the current motion state category of the non-cooperative drone.

4. The method according to claim 1, characterized in that, In step (c), the prediction of the spatiotemporal position probability distribution of the non-cooperative UAV includes: dividing the preset horizontal prediction area and the preset vertical prediction range into multiple horizontal grids and multiple vertical height intervals, respectively; the trajectory prediction model based on Transformer outputs the probability of the non-cooperative UAV falling into each horizontal grid and the probability of falling into each vertical height interval at each prediction time in the future, according to the motion sequence within the short-term observation window and the current motion state category. In the formula, For non-cooperative drones, predicting future moments Falling into The probability of a horizontal grid. For non-cooperative drones, predicting future moments Falling into The probability of a vertical height interval and These represent the predicted values ​​output by the trajectory prediction model for the corresponding horizontal grid and vertical height intervals, respectively. and These represent the unnormalized output values ​​corresponding to each horizontal grid and each vertical height interval traversed during the normalization process. and These represent the number of grid cells in the two horizontal directions of the preset horizontal prediction region. The number of vertical height intervals; the spatiotemporal position probability distribution of non-cooperative UAVs is composed of the horizontal grid probability distribution and the vertical height interval probability distribution corresponding to each future prediction time.

5. The method according to claim 1, characterized in that, In step (d), the determination of interval loss risk includes: based on the known or planned position of the target UAV at the future prediction time and the preset horizontal and vertical safety intervals, determining the horizontal risk area and vertical risk interval corresponding to the prediction time; and summing the horizontal grid probabilities that intersect with the horizontal risk area and the vertical height interval probabilities that intersect with the vertical risk interval to obtain the horizontal interval loss probability and the vertical interval loss probability. In the formula, Predicting the future The horizontal interval loss probability, Predicting the future The probability of vertical spacing loss, For the first One horizontal grid, For the first A vertical height range This is a horizontal risk area. The vertical risk range is defined as follows: when there is at least one prediction moment within the future prediction time window, such that the probability of horizontal interval loss is greater than the preset horizontal risk threshold and the probability of vertical interval loss is greater than the preset vertical risk threshold, it is determined that there is a risk of interval loss or conflict between the target UAV and surrounding non-cooperative UAVs.

Images