A method and system for predicting the direction of movement of a cluster of armored vehicles

Abstract

The application discloses an armored vehicle cluster motion direction prediction method and system, relates to the battlefield situation assessment field, and provides the armored vehicle cluster motion direction prediction method which establishes a trajectory prediction model of a single armored vehicle by collecting armored vehicle motion trajectory data, abstractly mines hidden rules from armored vehicle historical trajectories by using the ability of the LSTM in processing high-dimensional time sequence data, abstracts and divides the information of scattered armored vehicles by clustering the predicted trajectories, forms a subset cluster unit assumption in the relationship level, so as to reveal the mutual relationship between situation elements, determine the mutual cooperation function, and thus explain various behaviors of the land battlefield armored target, and provide necessary technical support for intention recognition according to the trajectory information.

Description

Technical Field

[0001] This invention relates to the field of battlefield situation assessment, and in particular to a method and system for predicting the movement direction of armored vehicle clusters. Background Technology

[0002] In the real environment of ground combat, the numerous operational targets, complex coordination relationships, frequent maneuvers, and rapidly changing battlefield situations pose significant challenges to comprehensive battlefield situation assessment. Armored vehicles, due to their superior mobility and maneuverability, as well as their powerful lethality, are increasingly widely used in the increasingly complex field of ground assault operations. The movement of enemy armored vehicles often carries strong military intentions, and the correlation between the target's trajectory and the battlefield environment embodies the commander's decision-making process, which integrates battlefield rules and personal factors. This can guide our forces in identifying their intentions and lay the foundation for seizing the initiative. Therefore, the rapid and accurate prediction of the future position of enemy vehicles, and the interpretation of various behaviors of armored targets on the land battlefield, are crucial for assessing operational intentions and estimating threats, and for achieving "transparency" in the fog of war.

[0003] Traditional trajectory prediction methods only consider the target's own movement trend and the influence of the environment on the target's movement, without considering the impact of the interaction characteristics between the two targets. In reality, in actual combat environments, the target's movement trajectory is mainly affected by the commander's intentions and the environment. Different commanders have different analyses and perceptions of the situation, which makes the influence of hidden intentions on the trajectory highly complex. That is, the interaction characteristics between vehicles in the cluster often hide the combat intentions of the decision-makers within the target cluster. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for predicting the movement direction of an armored vehicle cluster, which can accurately predict the movement trajectory of a single armored vehicle and thereby predict the overall movement direction of the armored vehicle sub-cluster.

[0005] To achieve the above objectives, this invention provides a method for predicting the movement direction of an armored vehicle cluster, wherein the armored vehicle cluster includes several armored vehicles; the method for predicting the movement direction of the armored vehicle cluster includes the following steps:

[0006] The battlefield simulation software is used to simulate several preset combat behaviors of a single armored vehicle, and the trajectory data of the armored vehicle at each moment of the simulation and the corresponding real trajectory are obtained as a simulation dataset; the trajectory data includes maneuver characteristics, environmental characteristics and vehicle-to-vehicle interaction characteristics.

[0007] The simulation dataset is used to train a single-vehicle trajectory prediction model, resulting in a simulated single-vehicle trajectory prediction model.

[0008] The trajectory data and corresponding real trajectories of each armored vehicle in the armored vehicle cluster at several historical moments are obtained as a historical dataset.

[0009] The parameters of the simulated vehicle trajectory prediction model are fine-tuned using the historical dataset to obtain the fine-tuned vehicle trajectory prediction model.

[0010] The trajectory data of any armored vehicle in the armored vehicle cluster at the current moment is input into the single-vehicle trajectory prediction model after parameter fine-tuning to obtain the predicted trajectory of the armored vehicle at several future moments.

[0011] The predicted trajectories of each armored vehicle in the armored vehicle cluster are clustered to obtain several armored vehicle sub-clusters; each armored vehicle sub-cluster includes several armored vehicles.

[0012] For any armored vehicle sub-group, the direction of movement of the armored vehicle sub-group is determined based on the predicted trajectories of each armored vehicle in the sub-group.

[0013] The motion direction vector sum of each armored vehicle sub-group in the target armored vehicle cluster is calculated to obtain the motion direction of the target armored vehicle cluster.

[0014] Optionally, the single-vehicle trajectory prediction model is a two-layer LSTM neural network.

[0015] Optionally, the maneuvering characteristics include latitude and longitude, speed, acceleration, vertical deflection angle, horizontal deflection angle, turning radius, and turning angle; the environmental characteristics include location coordinate information, terrain information, hydrological information, temperature, and visibility; and the vehicle-to-vehicle interaction characteristics include relative distance, relative altitude, and the angle between velocity vectors.

[0016] Optionally, before training the single-vehicle trajectory prediction model using the simulation dataset, the method for predicting the movement direction of an armored vehicle cluster further includes:

[0017] The trajectory data in the simulation dataset and the real trajectory are normalized.

[0018] The simulation dataset is divided into a simulation training set and a simulation test set using a staggered segmentation method.

[0019] The process of training a single-vehicle trajectory prediction model using the simulation dataset specifically includes:

[0020] The trajectory data from the simulation training set are input into the single-vehicle trajectory prediction model to obtain the simulation prediction trajectory corresponding to the trajectory data.

[0021] For any trajectory data in the simulation training set, the variance between the simulated predicted trajectory corresponding to the trajectory data and the actual trajectory corresponding to the trajectory data is calculated to obtain the simulation prediction error of the trajectory data.

[0022] Based on the simulation prediction error of each trajectory data in the simulation training set, the parameters in the single-vehicle trajectory prediction model are corrected to obtain the single-vehicle trajectory prediction model after simulation training.

[0023] The accuracy of the single-vehicle trajectory prediction model after simulation training was tested using the simulation test set.

[0024] Optionally, the variance between the simulated predicted trajectory and the actual trajectory is calculated according to the following formula:

[0025]

[0026] Among them, y pre To simulate and predict the trajectory's coordinates in three dimensions at time i, y i Let be the coordinate data of the real trajectory in three dimensions at time i, and n be the time length of the simulated predicted trajectory and the real trajectory.

[0027] Optionally, before fine-tuning the parameters of the simulated single-vehicle trajectory prediction model using the historical dataset, the armored vehicle cluster motion direction prediction method further includes:

[0028] The trajectory data and the actual trajectory in the historical dataset are normalized.

[0029] The historical dataset is divided into a historical training set and a historical test set using a staggered segmentation method.

[0030] The step of fine-tuning the parameters of the simulated vehicle trajectory prediction model using the historical dataset specifically includes:

[0031] The trajectory data from the historical training set are input into the simulated vehicle trajectory prediction model to obtain the historical predicted trajectory corresponding to each trajectory data.

[0032] For any trajectory data in the historical training set, the variance between the historical predicted trajectory corresponding to the trajectory data and the actual trajectory corresponding to the trajectory data is calculated to obtain the historical prediction error of the trajectory data.

[0033] Based on the historical prediction error of each trajectory data in the historical training set, the parameters in the single-vehicle trajectory prediction model after simulation training are corrected to obtain the single-vehicle trajectory prediction model with fine-tuned parameters.

[0034] The accuracy of the single-vehicle trajectory prediction model after parameter fine-tuning was tested using the historical test set.

[0035] Optionally, the parameters in the single-vehicle trajectory prediction model after simulation training are corrected using the Adam optimization algorithm; the Adam optimization algorithm is shown in the following equation:

[0036]

[0037] Where momentum is the momentum term, β1 and β2 are hyperparameters, the subscript t represents time t, E is the error sum, and w t Let v be the parameters of the single-vehicle trajectory prediction model after simulation training at time t. t Let t be the velocity term at time t, and learning_rate be the learning rate.

[0038] Optionally, the step of clustering the predicted trajectories of each armored vehicle in the armored vehicle cluster to obtain several armored vehicle sub-clusters specifically includes:

[0039] The DBSCAN algorithm is used to scan the predicted trajectories of each armored vehicle in the armored vehicle cluster to determine at least one core trajectory.

[0040] For any of the core trajectories, calculate the similarity between the core trajectory and each of the predicted trajectories other than the core trajectory.

[0041] The armored vehicles corresponding to several predicted trajectories with similarity below a preset threshold, as well as the armored vehicles corresponding to the core trajectory, are considered as a sub-cluster of armored vehicles.

[0042] Optionally, the similarity between the core trajectory and each predicted trajectory other than the core trajectory is calculated according to the following formula:

[0043] H(A,B)=max(h(A,B),h(B,A))

[0044]

[0045]

[0046] Where H(A,B) is the similarity between the core trajectory A and the predicted trajectory B, h(A,B) is the one-way Hausdorff distance from the core trajectory A to the predicted trajectory B, and h(B,A) is the one-way Hausdorff distance from the predicted trajectory B to the core trajectory A. i Let b be a point on the core trajectory A. j Let |||| be a point on the predicted trajectory B, and |||| be the distance paradigm between the two points.

[0047] On the other hand, the present invention also provides an armored vehicle cluster motion direction prediction system, which, when run by a computer, executes the armored vehicle cluster motion direction prediction method described above.

[0048] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0049] This invention provides a method and system for predicting the movement direction of an armored vehicle cluster. The method includes: simulating several preset combat behaviors of individual armored vehicles using battlefield simulation software, acquiring trajectory data of the armored vehicles at each moment during the simulation and their corresponding real trajectories as a simulation dataset; training a single-vehicle trajectory prediction model using the simulation dataset to obtain a simulated-trained single-vehicle trajectory prediction model; acquiring trajectory data of each armored vehicle in the cluster at several historical moments and their corresponding real trajectories as a historical dataset; fine-tuning the parameters of the simulated-trained single-vehicle trajectory prediction model using the historical dataset to obtain a parameter-fine-tuned single-vehicle trajectory prediction model; inputting the current trajectory data of any armored vehicle in the cluster into the parameter-fine-tuned single-vehicle trajectory prediction model to obtain the predicted trajectory of the armored vehicle at several future moments; clustering the predicted trajectories of each armored vehicle in the cluster to obtain several armored vehicle sub-clusters; each armored vehicle sub-cluster includes several armored vehicles; and for any armored vehicle sub-cluster, determining the movement direction of the sub-cluster based on the predicted trajectories of each armored vehicle within the sub-cluster. The armored vehicle cluster movement direction prediction method provided by this invention establishes a trajectory prediction model for a single armored vehicle by collecting armored vehicle movement trajectory data. It utilizes the ability of LSTM to process high-dimensional time-series data to extract abstract implicit rules from the historical trajectories of armored vehicles. At the same time, by clustering the predicted trajectories, it abstracts and divides the scattered information of armored vehicles to form sub-cluster unit hypotheses at the relational level, so as to reveal the interrelationships between situational elements, determine the functions of mutual cooperation, and thus explain the various behaviors of armored targets on the land battlefield, providing the necessary technical support for intention recognition based on trajectory information. Attached Figure Description

[0050] 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 introduced 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.

[0051] Figure 1 This is a flowchart of a method for predicting the movement direction of an armored vehicle cluster provided in Embodiment 1 of the present invention;

[0052] Figure 2 This is a flowchart illustrating step S6 in the method provided in Embodiment 1 of the present invention.

[0053] Figure 3 This is a schematic diagram illustrating battlefield scenario simulation in the method provided in Embodiment 1 of the present invention;

[0054] Figure 4 This is a schematic diagram of the trajectory data obtained in the method provided in Embodiment 1 of the present invention;

[0055] Figure 5 A schematic diagram of the structure of an armored vehicle cluster motion direction prediction system provided in Embodiment 2 of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] Trajectory prediction is essentially the prediction of time-series data, which is highly nonlinear and time-varying. Currently, there are two main approaches to trajectory prediction: model-based methods and data-driven methods.

[0058] In modeling methods, T. Baklacioglu et al. proposed a transport aircraft climb and descent trajectory prediction model based on a genetic algorithm; Tang Xinmin et al. established a time-varying hybrid system model based on aircraft dynamics to predict aircraft trajectories, achieving accurate and rapid trajectory prediction; Qiao Shaojie et al. proposed a dynamic trajectory prediction algorithm based on Kalman filtering, achieving high prediction accuracy. These methods require high model accuracy and offer high prediction accuracy when the flight trajectory is relatively simple and stable, but their prediction accuracy is relatively low for complex models.

[0059] In data-driven approaches, Yang Rennong et al. established an aircraft flight performance model based on a NARX feedback neural network, achieving good prediction results and convergence speed. Tan Wei et al. used a genetic algorithm to optimize the weights and structural parameters of the neural network, establishing a BP neural network maneuvering target prediction model with high prediction accuracy. Wang Jianchen et al. improved the particle swarm optimization algorithm and the gradient method to optimize the Elman network weights, predicting the landing trajectory of UAVs, with good model convergence performance. The advantage of this approach is that it does not require building an aircraft motion model or considering motion details; it simply starts from the data perspective, using machine learning algorithms to uncover the underlying aircraft motion patterns. However, the model input is simple, the "learned patterns" are not deep, and it cannot perform long-term stable predictions.

[0060] The above findings reveal that traditional trajectory prediction methods only consider the target's own motion trend and the influence of the environment on the target's movement, without taking into account the impact of the interaction characteristics between the two targets. At the same time, due to the special structure of armored vehicles, the turning process is always accompanied by the slippage of the high-speed side track and the slippage of the low-speed side track. Compared with wheeled vehicles, the slip steering characteristics of armored vehicles make trajectory prediction more difficult.

[0061] Therefore, in order to avoid the problems existing in the prior art, the purpose of this invention is to provide a method and system for predicting the movement direction of armored vehicle clusters. By establishing a single-vehicle trajectory prediction model and training the model and adjusting the parameters through simulation data and historical data, the obtained single-vehicle trajectory prediction model can accurately predict the movement trajectory of a single armored vehicle. Furthermore, by using a clustering algorithm, multiple armored vehicles are clustered into multiple armored vehicle sub-clusters, and the overall movement direction of the armored vehicle sub-clusters is accurately predicted, thereby inferring the command intention of the armored vehicle cluster.

[0062] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0063] Example 1:

[0064] This embodiment provides a method for predicting the movement direction of an armored vehicle cluster, wherein the armored vehicle cluster includes several armored vehicles; such as Figure 1 As shown, the method for predicting the movement direction of an armored vehicle cluster includes the following steps:

[0065] S1. Construct a simulation dataset; use battlefield simulation software to simulate several preset combat behaviors of a single armored vehicle, and obtain the trajectory data of the armored vehicle at each moment of the simulation and the corresponding real trajectory as the simulation dataset; the trajectory data includes maneuver characteristics, environmental characteristics and vehicle-to-vehicle interaction characteristics.

[0066] S2. Preliminary training of the prediction model; using the simulation dataset to train the single-vehicle trajectory prediction model, and obtain the single-vehicle trajectory prediction model after simulation training.

[0067] S3. Obtain historical dataset; Obtain trajectory data and corresponding real trajectories of each armored vehicle in the armored vehicle cluster at several historical moments, as historical dataset.

[0068] S4. Fine-tuning the prediction model parameters: Using the historical dataset, fine-tun the parameters of the simulated vehicle trajectory prediction model to obtain the fine-tuned vehicle trajectory prediction model.

[0069] S5. Use a prediction model to predict the trajectory; input the trajectory data of any armored vehicle in the armored vehicle cluster at the current moment into the single-vehicle trajectory prediction model after parameter fine-tuning, and obtain the predicted trajectory of the armored vehicle at several future moments.

[0070] S6. Cluster the armored vehicles to determine sub-clusters; cluster the predicted trajectories of each armored vehicle in the armored vehicle cluster to obtain several armored vehicle sub-clusters; each armored vehicle sub-cluster includes several armored vehicles. For example... Figure 2 As shown, step S6 specifically includes the following steps:

[0071] S61. Use the DBSCAN algorithm to scan the predicted trajectories of each armored vehicle in the armored vehicle cluster to determine at least one core trajectory.

[0072] S62. For any one of the core trajectories, calculate the similarity between the core trajectory and each predicted trajectory other than the core trajectory. In this embodiment, the similarity between the core trajectory and each predicted trajectory other than the core trajectory is calculated according to the following formula:

[0073] H(A,B)=max(h(A,B),h(B,A))

[0074]

[0075]

[0076] Where H(A,B) is the similarity between the core trajectory A and the predicted trajectory B, h(A,B) is the one-way Hausdorff distance from the core trajectory A to the predicted trajectory B, and h(B,A) is the one-way Hausdorff distance from the predicted trajectory B to the core trajectory A. i Let b be a point on the core trajectory A. j Let |||| be a point on the predicted trajectory B, and |||| be the distance paradigm between the two points.

[0077] S63. The armored vehicles corresponding to several predicted trajectories with similarity below a preset threshold, as well as the armored vehicles corresponding to the core trajectory, are considered as a sub-cluster of armored vehicles.

[0078] S7. Determine the movement direction of the sub-cluster; for any armored vehicle sub-cluster, determine the movement direction of the armored vehicle sub-cluster based on the predicted trajectory of each armored vehicle in the sub-cluster.

[0079] Given the continuous and temporal characteristics of armored vehicle trajectory data, a Long Short Memory (LSTM) neural network, which excels at processing time-series data, is employed as the prediction model for armored vehicle trajectories. This model can selectively memorize information from previous moments and combine it with current moment information to make judgments, reducing the computational load for armored vehicle trajectory prediction and enabling rapid inference of the armored vehicle's trajectory route over a future period. Therefore, in this specific embodiment, the single-vehicle trajectory prediction model is a two-layer LSTM neural network.

[0080] S8. Determine the overall movement direction of the cluster; calculate the sum of the movement direction vectors of each armored vehicle sub-cluster in the target armored vehicle cluster to obtain the movement direction of the target armored vehicle cluster.

[0081] In this embodiment, the vehicle's maneuvering characteristics, environmental characteristics, and vehicle-to-vehicle interaction characteristics are defined. The maneuvering characteristics include latitude and longitude, speed, acceleration, vertical deflection angle, horizontal deflection angle, turning radius, and turning angle. The environmental characteristics include location coordinate information, terrain information, hydrological information, temperature, and visibility. The vehicle-to-vehicle interaction characteristics are the interactive features between vehicles, including the relative distance, relative height, and velocity vector angle between any two vehicles.

[0082] The collected simulation datasets and historical datasets often cannot be used directly; they need to be preprocessed before they can be used to train the model. Specifically, in this embodiment, before step S2, the method for predicting the movement direction of the armored vehicle cluster further includes:

[0083] The trajectory data in the simulation dataset and the real trajectory are normalized.

[0084] The simulation dataset is divided into a simulation training set and a simulation test set using a staggered segmentation method; specifically, in this embodiment, the simulation training set and the simulation test set are divided in a 7:3 ratio.

[0085] After preprocessing the data in the simulation dataset and dividing it into training and testing sets, step S2 specifically includes the following steps:

[0086] S21. Input the trajectory data of the simulation training set into the single-vehicle trajectory prediction model to obtain the simulation prediction trajectory corresponding to the trajectory data.

[0087] S22. For any trajectory data in the simulation training set, calculate the variance between the simulated predicted trajectory corresponding to the trajectory data and the real trajectory corresponding to the trajectory data to obtain the simulation prediction error of the trajectory data.

[0088] S23. Correct the parameters in the single-vehicle trajectory prediction model based on the simulation prediction error of each trajectory data in the simulation training set to obtain the single-vehicle trajectory prediction model after simulation training.

[0089] S24. Use the simulation test set to test the accuracy of the single-vehicle trajectory prediction model after simulation training.

[0090] Correspondingly, prior to step S4, the method for predicting the movement direction of the armored vehicle cluster further includes:

[0091] The trajectory data and the actual trajectory in the historical dataset are normalized.

[0092] The historical dataset is divided into a historical training set and a historical test set using a staggered segmentation method; specifically, in this embodiment, the historical training set and the historical test set are divided in a 7:3 ratio.

[0093] After preprocessing the data in the historical dataset and dividing it into training and test sets, step S4 specifically includes the following steps:

[0094] S41. Input the trajectory data in the historical training set into the single-vehicle trajectory prediction model after simulation training to obtain the historical predicted trajectory corresponding to each trajectory data.

[0095] S42. For any trajectory data in the historical training set, calculate the variance between the historical predicted trajectory corresponding to the trajectory data and the actual trajectory corresponding to the trajectory data to obtain the historical prediction error of the trajectory data.

[0096] S43. Based on the historical prediction error of each trajectory data in the historical training set, correct the parameters in the single-vehicle trajectory prediction model after simulation training to obtain the single-vehicle trajectory prediction model with fine-tuned parameters.

[0097] S44. Use the historical test set to test the accuracy of the single-vehicle trajectory prediction model after parameter fine-tuning.

[0098] Specifically, in this embodiment, the variance between the simulated predicted trajectory and the actual trajectory is calculated according to the following formula, and the variance between the simulated predicted trajectory and the actual trajectory is used as the loss for forward propagation:

[0099]

[0100] Among them, y pre To simulate and predict the trajectory's coordinates in three dimensions at time i, y i Let be the coordinate data of the real trajectory in three dimensions at time i, and n be the time length of the simulated predicted trajectory and the real trajectory.

[0101] The Adam optimization algorithm adaptively adjusts the learning step size, resulting in faster model convergence, lower resource requirements, and better learning performance. Therefore, in this embodiment, the Adam optimization algorithm is used to correct the parameters in the simulated vehicle trajectory prediction model; the Adam optimization algorithm is shown in the following equation:

[0102]

[0103] Where momentum is the momentum term, β1 and β2 are hyperparameters, typically set to β1 = 0.9 and β2 = 0.999, the subscript t represents time t, E is the sum of errors, which is obtained from the loss calculated by the aforementioned formula, w t Let v be the parameters of the single-vehicle trajectory prediction model after simulation training at time t. t Let t be the velocity term at time t, and learning_rate be the learning rate.

[0104] The following example illustrates the method for predicting the movement direction of armored vehicle clusters provided in this embodiment, including the following steps:

[0105] A1. Using the military battlefield scenario simulation software VMS, simulate various combat behaviors of a single armored vehicle during combat and collect simulation data, including the basic trajectory data of the armored vehicle.

[0106] In this example, the combat behavior designed using the battlefield scenario simulation software VMS is as follows: To more intuitively observe the changes in the trajectory of the armored vehicle formation, such as... Figure 3 As shown in the figure, the basic elements are explained as follows: squares represent armored vehicles of Party A, triangles represent armored vehicles of Party B, arrows represent the velocity vector direction of the corresponding armored vehicles, ellipses represent lakes, and ellipsoids represent hills.

[0107] In this example, it is assumed that the detection range of both sides is limited (i.e., the location of targets outside the detection range is unknown). In the first selected frame, the armored vehicles of both sides are assembled at their respective bases and ready to act according to the superior strategy. In the second selected frame, side B initiates the advance order first, while side A remains in place due to insufficient detection range. In the third selected frame, side B continues to advance in different directions (A, B, and C move counterclockwise around the lake, while D and E continue to advance on the lower half of the lake with different heading angles). In the fourth selected frame, side A discovers the position of side B D and sends side A a to intercept side B D. side B D immediately retreats, while side B A and side B B begin to climb the hill, and side B C and side B E continue to advance. In the selected 4th frame, Party A stops at the mountaintop, while Party B and Party C continue towards the summit. Party A continues to pursue Party D, who accelerates to escape. At this moment, Party B spots Party E's position and intercepts them. The simulation depicts a series of combat missions for both sides, with both constantly shifting their positions, ultimately reaching the layout shown in the selected 6th frame.

[0108] A2. After performing differential standardization or normalization and segmentation on the collected simulation data, the processed simulation data is used to initially train the single-vehicle trajectory prediction model based on a two-layer long short-term memory neural network, resulting in a single-vehicle trajectory prediction model trained on the simulation data. In this embodiment, the specific training process is as follows:

[0109] A2.1. Normalize the simulation data, then segment it using a staggered segmentation method. Divide the training and test sets into a 7:3 ratio; the normalization process is performed using the following formula:

[0110] x*=(xx min ) / (x max -x min )

[0111] Where x* represents the normalized data, x represents the data to be processed, min represents the minimum value, and max represents the maximum value.

[0112] A2.2. Train the Long Short-Term Memory (LSTM) neural network using the training set. The activation function is Sigmoid, the loss function is specified as mean_squared_error, the optimizer is Adam, and the training epochs are 400.

[0113] A2.3. Save the Long Short-Term Memory (LSTM) neural network trained on the simulation data, test it using the test set, and then test the LSM network again using cross-validation. The entire training set is divided into K groups using K-fold cross-validation. Each time, one group is selected as the validation set, and the other K-1 groups are used as the training set. After K-fold cross-validation, each set of parameters will receive K scores. The average of these K scores is taken as the final score for that set of parameters.

[0114] A3. Real-time acquisition and storage of trajectory data for each armored vehicle at historical moments; that is, collecting characteristic information of armored vehicles through sensors. In this example, this includes extracting vehicle maneuvering characteristics, environmental characteristics, and vehicle-to-vehicle interaction characteristics. In this example, maneuvering characteristics include latitude and longitude, speed, acceleration, vertical yaw angle, horizontal yaw angle, turning radius, and turning angle, etc. Environmental characteristics include position coordinates, terrain information, hydrological information, temperature, and visibility, etc. Vehicle-to-vehicle interaction characteristics are the interactive features between vehicles, including the relative distance, relative height, and velocity vector angle between any two vehicles, etc. Figure 4 As shown.

[0115] A4. After performing differential standardization preprocessing on the real-time acquired armored vehicle trajectory feature data, the historical data is segmented. The segmented historical data is then used to retrain the long short-term memory neural network trained on the simulation data in real time. The parameters of the long short-term memory neural network are adjusted to obtain the retrained long short-term memory neural network.

[0116] In the process of training a two-layer LSTM neural network using simulated data and then retraining the simulated LSTM neural network in real time using segmented historical data, the variance between the predicted and actual results is used as the loss for forward propagation, as detailed below:

[0117]

[0118] Among them, y pre For the predicted coordinates of the armored vehicle at time i, y i For the actual coordinate data of the armored vehicle at time i, use y pre and y i The variance is used to correct the parameters in the two-layer LSTM neural network, so that the two-layer LSTM neural network converges.

[0119] The LSTM network training algorithm employs the Adam optimization algorithm. The Adam optimization algorithm adaptively adjusts the learning step size, resulting in faster model convergence, lower resource requirements, and better learning performance. The weight update formula is as follows:

[0120]

[0121] Where momentum is the momentum term; v is the velocity term; t represents time t; E is the sum of errors obtained after each training iteration; w t Here are the parameters of the model at time t, β1 and β2 are hyperparameters, typically set to β1 = 0.9 and β2 = 0.999, and learning_rate is the learning rate.

[0122] A5. Continue to collect, differentially standardize or normalize, and segment the trajectory data of the armored vehicles. Then, use the retrained long short-term memory neural network to make predictions and output the prediction results to obtain the trajectory coordinates of the armored vehicles in the future.

[0123] In this example, the trajectory data of the armored vehicle collected in step A1 and the trajectory data of the armored vehicle collected in step A5 both include 17-dimensional data, namely 14-dimensional data composed of the aforementioned maneuver features, environmental features and vehicle-to-vehicle interaction features, plus the current 3-dimensional coordinates; the 17-dimensional trajectory data is input into a two-layer LSTM neural network, which outputs multiple sets of 3-dimensional predicted coordinate data.

[0124] A6. Extract the most accurate predicted trajectory data of each armored vehicle from the database. Scan the entire trajectory segment dataset using the DBSCAN algorithm to find a core trajectory segment, and then traverse and query the Eps neighborhood of a sample point to obtain a set of clusters.

[0125] Before step A6, it is necessary to determine the global parameters of the DBSCAN algorithm, namely the scan radius Eps and the minimum number of contained points MinPts. The parameters (Eps, MinPts) are used to describe the density of the sample distribution in the neighborhood. Eps describes the neighborhood distance threshold of a sample, and MinPts describes the threshold of the number of samples in the neighborhood of a sample at a distance of Eps.

[0126] Secondly, in step A6, a DBSCAN cluster can have one or more core objects. If there is only one core object, then all other non-core object samples in the cluster are in the Eps neighborhood of this core object. If there are multiple core objects, then any core object in the Eps neighborhood of the cluster must have another core object; otherwise, the two core objects cannot be density-reachable. The set of all samples in the Eps neighborhood of these core objects constitutes a DBSCAN cluster.

[0127] In the specific implementation, the entire trajectory segment dataset is scanned to find a core trajectory segment. If there is only one core object, it is directly selected as the core trajectory segment. If there are multiple core objects, one of them is arbitrarily selected as the core trajectory segment, and this core trajectory segment is then expanded. When traversing and querying the Eps neighborhood of a sample point, points within the neighborhood are temporarily excluded, and the query focuses on points within the Eps neighborhood of the core node that are not marked, thus avoiding duplicate node queries and reducing the number of neighborhood queries and query time.

[0128] In step A6, Hausdorff distance is used to measure trajectory similarity and quantify the distance between trajectories. Given two trajectories A and B, each consisting of several ordered trajectory points, the Hausdorff distance between trajectory A and trajectory B is:

[0129] H(A,B)=max(h(A,B),h(B,A))

[0130]

[0131]

[0132] In the formula: H(A,B) is the bidirectional Hausdorff distance between A and B; h(A,B) is the one-way Hausdorff distance from trajectory A to trajectory B; h(B,A) is the one-way Hausdorff distance from trajectory B to trajectory A; |||| is the distance paradigm between two points.

[0133] In this example, the process of obtaining the set of clusters using the DBSCAN algorithm is as follows:

[0134] A6.1 Input trajectory dataset D, scan radius Eps, and minimum number of points MinPts.

[0135] A6.2 Detect objects p in the database that have not yet been checked.

[0136] A6.3 Determine whether the number of objects contained in p is not less than MinPts. If it is not less, proceed to the next step; if it is less, mark p as noise.

[0137] A6.4. Establish a new cluster C1 and add all points in it to the candidate set N.

[0138] A6.5. For each point q in the candidate set N, if a point is a core node and its neighborhood does not overlap with that of object p, or if there is no core node in the overlapping area, then mark it with a new cluster C2, and then return to step A6.2 to continue judging the remaining points q; otherwise, proceed to the next step.

[0139] The process continues until all points q in the candidate set N have been evaluated, and finally the set of clusters is output.

[0140] A7. After identifying the armored vehicles within a cluster, a representative trajectory of the cluster describes the overall motion of that cluster. It can be considered a motion model of the cluster. This example extracts quantitative information about the motion within the cluster so that the combat commander can understand the motion in the trajectory. The representative trajectory of the cluster must meet the following two requirements:

[0141] 1. The overall direction of the cluster's representative trajectory needs to reflect the overall direction of movement of the cluster. That is, the direction of the cluster's representative trajectory is the vector sum of the trajectory directions of each unit within the cluster; that is, the total direction of movement of the cluster is obtained by summing the motion direction vectors of each armored vehicle in a sub-cluster.

[0142] 2. The local direction of the cluster representative trajectory needs to reflect the changes in the internal characteristics of the cluster. That is, the inflection point of the local direction of the cluster representative trajectory, i.e. the point of change, must be an internal feature point of the cluster trajectory.

[0143] Regarding point 1, firstly, assume there are k vehicles in a cluster. The representative trajectory of the cluster can be obtained using the scan line method: Select all vehicle trajectories within any cluster; count the displacement vectors of each vehicle in the cluster; sum the vectors to obtain the overall displacement vector direction of the cluster; establish a new coordinate system with the overall displacement vector of the cluster as the horizontal axis and the direction perpendicular to the horizontal axis as the vertical axis; translate the vertical axis, stopping when encountering a vehicle, and obtain several sets of scan lines in sequence, setting a threshold k; if the number of points where the scan line passes through a vehicle and the vehicle trajectory is greater than k, then a key point is considered to exist; the coordinates of the key point are calculated as the average value of the points passed through by the scan line; connect the key points in sequence to obtain the overall motion trajectory of the cluster, which satisfies requirement 1.

[0144] For point 2, set the hyperparameter k / 2, and draw the intersection line between the start point and end point of each vehicle and the perpendicular line of the overall direction. If the intersection line passes through k / 2 key points (key points include the position of the start point and end point of each vehicle and the line segment in between), then it is considered to satisfy requirement 2.

[0145] This example illustrates the process of the armored vehicle cluster movement direction prediction method provided in this embodiment. By collecting armored vehicle trajectory data, a trajectory prediction model for a single armored vehicle is established. LSTM's ability to process high-dimensional time-series data is utilized to extract abstract implicit rules from the historical trajectories of armored vehicles. Simultaneously, by clustering the predicted trajectories, scattered information about armored vehicles is abstracted and divided to form sub-cluster unit hypotheses at the relational level. This reveals the interrelationships between situational elements, determines their cooperative functions, and thus explains the various behaviors of armored targets on the land battlefield, providing necessary technical support for intention recognition based on trajectory information.

[0146] Example 2:

[0147] Furthermore, the method of Embodiment 1 of the present invention can also be used by means of Figure 5 The architecture of the armored vehicle cluster movement direction prediction system shown is implemented as follows. Figure 5 As shown, the armored vehicle cluster motion direction prediction system may include a simulation module, a simulation dataset construction module, a prediction model training module, a historical dataset construction module, a prediction model fine-tuning module, a trajectory prediction module, an armored vehicle clustering module, and a motion direction determination module. Some modules may also have sub-units to implement their functions; for example, the armored vehicle clustering module includes a core trajectory determination unit, a trajectory similarity evaluation unit, and a sub-cluster clustering unit. Of course, Figure 5 The architecture shown is merely exemplary; it can be omitted as needed when implementing different functionalities. Figure 5 One or at least two components of the system shown.

[0148] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0149] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for predicting the movement direction of an armored vehicle cluster, wherein the armored vehicle cluster comprises several armored vehicles; characterized in that, The method for predicting the movement direction of armored vehicle clusters includes: Battlefield simulation software was used to simulate several pre-defined combat behaviors of a single armored vehicle, and the trajectory data of the armored vehicle at each moment during the simulation and the corresponding real trajectory were obtained as a simulation dataset. The trajectory data includes maneuver characteristics, environmental characteristics, and vehicle-to-vehicle interaction characteristics. The maneuver characteristics include latitude and longitude, speed, acceleration, vertical deflection angle, horizontal deflection angle, turning radius, and turning angle. The environmental characteristics include position coordinate information, terrain information, hydrological information, temperature, and visibility. The vehicle-to-vehicle interaction characteristics include relative distance, relative altitude, and velocity vector angle. A single-vehicle trajectory prediction model is trained using the simulation dataset to obtain a simulated single-vehicle trajectory prediction model; the single-vehicle trajectory prediction model is a two-layer LSTM neural network. The trajectory data and corresponding real trajectories of each armored vehicle in the armored vehicle cluster at several historical moments are obtained as a historical dataset. The parameters of the simulated vehicle trajectory prediction model are fine-tuned using the historical dataset to obtain the fine-tuned vehicle trajectory prediction model. The trajectory data of any armored vehicle in the armored vehicle cluster at the current moment is input into the single-vehicle trajectory prediction model after parameter fine-tuning to obtain the predicted trajectory of the armored vehicle at several future moments. Clustering the predicted trajectories of each armored vehicle in the armored vehicle cluster yields several armored vehicle sub-clusters; each armored vehicle sub-cluster includes several armored vehicles; the specific steps of clustering the predicted trajectories of each armored vehicle in the armored vehicle cluster to obtain several armored vehicle sub-clusters include: The DBSCAN algorithm is used to scan the predicted trajectories of each armored vehicle in the armored vehicle cluster to determine at least one core trajectory. For any of the core trajectories, the similarity between the core trajectory and each predicted trajectory other than the core trajectory is calculated based on the Hausdorff distance; The armored vehicles corresponding to several predicted trajectories with similarity below a preset threshold, as well as the armored vehicles corresponding to the core trajectory, are considered as a sub-cluster of armored vehicles. The formula for calculating the Hausdorff distance is: in, H ( A , B (with) as the core trajectory A With predicted trajectory B Similarity between them h ( A , B ) for the core trajectory A To the predicted trajectory B One-way Hausdorf distance, h ( B , A ) for predicting the trajectory B To the core trajectory A One-way Hausdorf distance, a i core trajectory A The point above, b j For predicting trajectories B Let || be a point on the graph, and || be the distance normal form between the two points. For any armored vehicle sub-group, the movement direction of the armored vehicle sub-group is determined based on the predicted trajectory of each armored vehicle in the sub-group. The motion direction vector of each armored vehicle sub-group in the target armored vehicle group is calculated to obtain the motion direction of the target armored vehicle group.

2. The method for predicting the movement direction of an armored vehicle cluster according to claim 1, characterized in that, Before training the single-vehicle trajectory prediction model using the simulation dataset, the method for predicting the movement direction of an armored vehicle cluster also includes: The trajectory data and real trajectories in the simulation dataset are normalized. The simulation dataset is divided into a simulation training set and a simulation test set using a staggered segmentation method. The process of training a single-vehicle trajectory prediction model using the simulation dataset specifically includes: The trajectory data of the simulation training set are input into the single-vehicle trajectory prediction model to obtain the simulation prediction trajectory corresponding to the trajectory data. For any trajectory data in the simulation training set, calculate the variance between the simulated predicted trajectory corresponding to the trajectory data and the actual trajectory corresponding to the trajectory data to obtain the simulation prediction error of the trajectory data. Based on the simulation prediction error of each trajectory data in the simulation training set, the parameters in the single-vehicle trajectory prediction model are corrected to obtain the single-vehicle trajectory prediction model after simulation training. The accuracy of the single-vehicle trajectory prediction model after simulation training was tested using the simulation test set.

3. The method for predicting the movement direction of an armored vehicle cluster according to claim 2, characterized in that, The variance between the simulated predicted trajectory and the actual trajectory is calculated using the following formula: in, y pre To simulate and predict trajectories in i Coordinate data in three dimensions at any given time. y i For the true trajectory in i Coordinate data in three dimensions at any given time. n The time lengths for simulating the predicted trajectory and the actual trajectory.

4. The method for predicting the movement direction of an armored vehicle cluster according to claim 1, characterized in that, Before fine-tuning the parameters of the simulated single-vehicle trajectory prediction model using the historical dataset, the method for predicting the movement direction of an armored vehicle cluster further includes: The trajectory data and actual trajectories in the historical dataset are normalized. The historical dataset is divided into a historical training set and a historical test set using a staggered segmentation method. The step of fine-tuning the parameters of the simulated vehicle trajectory prediction model using the historical dataset specifically includes: The trajectory data in the historical training set are input into the simulated and trained single-vehicle trajectory prediction model to obtain the historical predicted trajectory corresponding to each trajectory data. For any trajectory data in the historical training set, calculate the variance between the historical predicted trajectory corresponding to the trajectory data and the actual trajectory corresponding to the trajectory data to obtain the historical prediction error of the trajectory data. Based on the historical prediction error of each trajectory data in the historical training set, the parameters in the single-vehicle trajectory prediction model after simulation training are corrected to obtain the single-vehicle trajectory prediction model after parameter fine-tuning. The accuracy of the single-vehicle trajectory prediction model after parameter fine-tuning was tested using the historical test set.

5. The method for predicting the movement direction of an armored vehicle cluster according to claim 4, characterized in that, The Adam optimization algorithm is used to correct the parameters in the simulated vehicle trajectory prediction model after training; the Adam optimization algorithm is shown in the following equation: in, momentum For momentum, β 1 and β 2 is a hyperparameter. t Subscript represents t time, E For the error sum, w t for t The parameters of the single-vehicle trajectory prediction model after time-lapse simulation training. v t for t The velocity term at time, learning_rate This is the learning rate.

6. A system for predicting the movement direction of an armored vehicle cluster, characterized in that, When the armored vehicle cluster motion direction prediction system is run by a computer, it executes the armored vehicle cluster motion direction prediction method as described in any one of claims 1-5.

Images