Transform-Glow model-based intention and trajectory prediction method
By using the Transformer-Glow model and the Bitrap bidirectional trajectory prediction module, the shortcomings of traditional methods in handling nonlinearity and multimodal conditions in UAV trajectory prediction are addressed, achieving high-precision and interpretable UAV trajectory prediction that is suitable for complex environments and multi-UAV systems.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-23
- Publication Date
- 2026-05-12
AI Technical Summary
Traditional UAV trajectory prediction methods are insufficient when dealing with complex nonlinear dynamics, multimodal trajectories, and collaborative prediction of multi-UAV systems. Furthermore, existing neural network methods have limitations in feature extraction efficiency and the forgetting problem.
The Transformer-Glow model is adopted, combined with the Bitrap bidirectional trajectory prediction module and the Temporal Transformer architecture. It learns latent patterns through historical trajectory data, generates UAV trajectory distribution, and uses a generative flow network for bidirectional mapping to achieve high-precision prediction.
It improves the accuracy and interpretability of trajectory prediction, overcomes the problem of error accumulation, and can adapt to the collaborative prediction needs of complex environments and multi-UAV systems.
Smart Images
Figure CN122021245A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of end-to-end intent and trajectory prediction technology for unmanned aerial vehicles (UAVs), and particularly to an intent and trajectory prediction method based on the Transformer-Glow model. Background Technology
[0002] In the military field, drones are widely used for reconnaissance, surveillance, intelligence gathering, and strike missions. Accurately predicting the intentions of drones and their trajectories at the next or even multiple moments in time can help friendly drones and operators better anticipate the drone's subsequent status during missions. Similarly, in combat, it can better predict the subsequent positions and intentions of enemy aircraft. This is of great significance for friendly drones to perform various missions and for defending against enemy drones.
[0003] In the field of UAV trajectory prediction, traditional methods such as Kalman filtering, Bayesian networks, Markov models, and physics-based dynamic models have been widely used for state estimation and trajectory inference. These methods typically rely on explicit system modeling and statistical assumptions, offering strong interpretability and real-time performance, and are suitable for scenarios with low noise and relatively simple environmental structures. However, traditional methods fall short in the face of the highly nonlinear motion of UAVs, the diversity of flight strategies, and complex environmental interactions in real-world applications. Their main limitations are: limited modeling capabilities (unable to effectively handle complex nonlinear dynamics); reliance on prior assumptions (performance highly dependent on accurate modeling of noise and system models); difficulty in representing multimodal trajectories (predictions often result in a single trajectory, ignoring behavioral uncertainties); and weak scalability (difficult to adapt to collaborative prediction needs in multi-UAV systems or complex three-dimensional spaces).
[0004] In contrast, neural network-based learning methods, especially recurrent neural networks (RNNs), long short-term memory networks (LSTMs), and Transformer architectures, have demonstrated significant advantages in UAV trajectory prediction in recent years. These methods possess the following characteristics: they do not require explicit modeling of system dynamics and can directly learn latent patterns from historical trajectory data; they have strong nonlinear modeling capabilities and are more adaptable; they can capture multimodal trajectory distributions and generate multiple possible trajectories through probabilistic modeling or generative models; and they easily integrate multi-source information (such as sensor data, map information, and wind field models) to improve prediction robustness.
[0005] In the application of neural networks to trajectory prediction, the accuracy of the final prediction hinges on the precise extraction of feature information from the original data. Currently, many trajectory prediction methods still use recurrent neural networks (RNNs) for feature extraction, which are inefficient in parallel processing and suffer from significant forgetting issues as time progresses. This method employs a transformer architecture to extract higher-order information from the original trajectory, enabling fast parallel processing. Furthermore, the self-attention mechanism effectively mitigates the forgetting problem.
[0006] Meanwhile, most current trajectory predictions for aerial targets are based on the data itself, directly generating trajectories from one to another. This method loses the high-order statistical information contained in the original trajectory data. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention provides an intent and trajectory prediction method based on the Transformer-Glow model.
[0008] This invention is achieved through the following technical solution:
[0009] This intent and trajectory prediction method, based on the Transformer-Glow model, uses the Bitrap model, a bidirectional trajectory prediction module, as its core. It first predicts the UAV's target destination using historical trajectories, then performs forward prediction from the current position and backward prediction from the destination. Finally, it combines the bidirectional prediction results to output the final predicted trajectory. Furthermore, it utilizes the TemporalTransformer architecture to extract the temporal dependencies of historical UAV flight parameter data and incorporates the Glow generative flow model to fit the UAV's trajectory distribution based on historical observation data. Ultimately, only a Gaussian distribution parameter and the currently observed historical data are input; the Glow layer models this as the distribution data of the entire trajectory, which is then input into the prediction module for prediction. The model exhibits strong interpretability and can effectively solve the current UAV trajectory prediction problem.
[0010] This application uses a Transformer layer to extract features from the input flight trajectory, attitude angles, and other information, obtaining a high-dimensional abstract representation of this information. A generative flow model, Glow, is used to model the action distribution of this high-dimensional abstract representation, and based on this distribution modeling, flight trajectory prediction is performed. This application introduces the Glow network for trajectory prediction. The Glow network was originally applied in image generation, its principle being to decompose the original image into Gaussian noise step by step through noise reduction. Similarly, the flight intent of a drone can be viewed as a complex distribution composed of steps, making the two very similar. Therefore, this application introduces the Glow network to generate the target intent from the original trajectory, and then evolves the intent to obtain the future trajectory. The intent contains the feature information of the original data. Furthermore, due to the network's reversibility, the predicted information can also be interpreted as an evolutionary process of the intent.
[0011] Compared with the prior art, this application has at least the following beneficial effects:
[0012] 1. This application differs from traditional prediction methods that directly predict trajectory points from features. It is the first to propose using a generative flow network to model the trajectory distribution of UAVs and using the trajectory distribution to guide trajectory prediction, thereby improving the accuracy of trajectory prediction.
[0013] 2. Compared with deep learning methods, this application has a certain degree of interpretability. At the same time, due to the use of generative flow network, it can achieve bidirectional mapping, which can not only complete trajectory prediction, but also retrieve historical trajectory information through trajectory inversion. This overcomes the shortcomings of continuous error accumulation in traditional unidirectional prediction and can achieve high-precision trajectory prediction results. Attached Figure Description
[0014] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0015] Figure 1 This is a schematic diagram of the trace prediction network model in the embodiment;
[0016] Figure 2 This is a schematic diagram of the data embedding module in the embodiment;
[0017] Figure 3 The diagram below is a schematic diagram of the time feature extraction module in the embodiment, where (a) is a schematic diagram of the self-attention layer and (b) is a structural diagram of the multi-head attention mechanism module;
[0018] Figure 4 This is a schematic diagram of the Glow trajectory distribution modeling layer in the embodiment;
[0019] Figure 5 This is a schematic diagram of the Bitrap bidirectional trajectory prediction layer in the embodiment;
[0020] Figure 6 The example diagram is an ADE / TC diagram.
[0021] Figure 7 The following is a diagram showing the prediction results of simple maneuvers in the embodiment, where (a) is climbing, (b) is turning left, (c) is diving, and (d) is turning right.
[0022] Figure 8 The following are the prediction results of complex maneuvers in the embodiments, where (a) is a right turn and dive after climbing, (b) is a left turn and dive after climbing, (c) is a vertical dive after climbing, and (d) is a turn and climb after diving. Detailed Implementation
[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 some embodiments of the present invention, but not all embodiments.
[0024] It should be noted that, unless otherwise specified, the embodiments and features described in this invention can be combined with each other. It should also be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments; similar or identical parts between embodiments can be referred to interchangeably.
[0025] The preferred embodiment of this invention discloses an intent and trajectory prediction method based on the Transformer-Glow model, which includes the following steps:
[0026] Step 1: Construct a trajectory prediction model based on neural networks
[0027] like Figure 1 As shown, in some embodiments, the neural network model includes a Temporal Transformer feature extraction layer, a Glow trajectory distribution modeling layer, and a Bitrap bidirectional trajectory prediction layer.
[0028] Step 2, Feature extraction based on Temporal Transformer
[0029] The input information for the Temporal Transformer feature extraction layer includes observation information and complete trajectory information. Observation information includes all information about the target UAV, including three-axis coordinates (x, y, z) in a geocentric Cartesian coordinate system, Euler angles (including roll, pitch, and yaw), the UAV's azimuth, and velocity—a total of eight dimensions. Complete trajectory information includes all pose information within the UAV's motion cycle. Complete trajectory information only contains trajectory coordinates.
[0030] To facilitate the analysis of subsequent results, the position of the UAV needs to be converted into meters. Therefore, coordinate transformation is first performed to convert the UAV's latitude, longitude and altitude coordinates into three-axis coordinates (x, y, z) in the geocentric coordinate system.
[0031] This represents all flight parameter information of the UAV at multiple moments, where T is a time parameter representing the number of input time steps;
[0032] This represents the flight parameters of the target UAV at time i, including the three axes of the geocentric rectangular coordinate system (x, y, z), Euler angles (roll, pitch, yaw), the target UAV's azimuth and velocity as input target state parameters, and all trajectory information; Understanding 1: Flight parameters include the three axes of the geocentric rectangular coordinate system (including x, y, z), Euler angles (including roll, pitch, yaw), the target UAV's azimuth and velocity. Flight parameters and all trajectory information The target state parameter is used as input.
[0033] This represents the flight trajectory data of the entire sample drone, i.e., all trajectory information.
[0034] , representing the trajectory coordinates of the UAV at time i.
[0035] In some embodiments, T is 60 and n is 100.
[0036] The Temporal Transformer feature extraction layer mainly includes a data embedding module and a multi-head attention mechanism module.
[0037] like Figure 2As shown, the data embedding module begins with a fully connected layer, whose purpose is to map the low-dimensional parameter information of the input to a high-dimensional space to facilitate subsequent feature extraction. Since the sub-attention mechanism does not contain temporal information, positional encoding is required so that the network can understand the time point at which each input vector is located. Simultaneously, to ensure the network fully recognizes the importance ranking of different time points, this embodiment introduces the idea of center encoding, using the out-degree of a certain time point (the number of subsequent time points) as the encoded value of that node, which is then mapped through a fully connected neural network and added to the embedded data.
[0038] The multi-head attention mechanism module adopts the classic transformer multi-head attention mechanism architecture. To better align with the characteristics of real-world time series data (i.e., earlier time nodes are unaware of later time nodes), a lower triangular matrix mask is introduced. This ensures that during the attention mechanism computation, attention is only focused on the current time node and data prior to it. A schematic diagram of the multi-head attention mechanism module is shown below. Figure 3 As shown in Figure (3), path a represents single-head attention, while path b represents multi-head attention.
[0039] The multi-head attention mechanism module receives the data feature vector H output by the first sub-module of the Temporal Transformer feature extraction layer. Each linear layer corresponds to a weight matrix W, and each feature vector is mapped to a query vector Q, a key vector K, and a value vector V, as shown in the following equation:
[0040]
[0041] The specific calculation formula for the multi-head attention mechanism module can be expressed as:
[0042]
[0043] In the above formula, Let Q be the dimension size. For matrix dot product, Let represent the weight parameter matrix of the multi-head attention linear layer. Qi, Ki, and Vi are the query vector Q, key vector K, and value vector V of the i-th attention head, respectively. Concat represents the numerical concatenation operation, which concatenates all values in the same dimension. SoftMax is the activation function selected for this layer, and mask is the time mask, which ensures that at any given moment during the attention calculation process, only data before the current moment can be considered.
[0044] Through the Temporal Transformer layer, we can obtain the embedded representation HI of historical observation information and the embedded representation TI of all trajectory information.
[0045] Step 3: Modeling the invertible transformation of intent distribution based on flight features
[0046] In the Glow trajectory distribution modeling layer, the method in this embodiment simulates the degradation and evolution process between complex and simple aircraft behaviors. This process is reversible and is called the mapping between simple and complex behaviors. This process is represented as:
[0047]
[0048] in, This represents a set of reversible transformations that simulate degradation and evolution processes, while It indicates intermediate movement behavior.
[0049] TI is an abstract representation of the global trajectory information extracted by the Temporal Transformer feature extraction layer, representing the distribution of complex motion. The forward process of the above equation gradually degenerates complex motion behavior (TI) into simpler behavior, until it degenerates into a normal distribution (the simplest maneuver behavior can be regarded as satisfying a normal distribution). The reverse process represents the evolution from simple behavior to complex motion behavior. This is consistent with real-world scenarios; complex motion behavior (TI) is not achieved overnight, but is based on simple motion behavior. (Such as each step forward or each maneuver), combined with various aspects such as underlying intentions. Therefore, Defined as N represents a multivariate standard normal distribution, 0 is the mean vector, and I is the covariance matrix of a multidimensional normal distribution.
[0050] The generative flow model of the Glow trajectory distribution modeling layer, Glow, essentially acts as a mapping between the simple behavior distribution z and the real complex maneuver behavior x, that is:
[0051]
[0052] In the above formula, P is the probability density function, Glow represents the mapping relationship from z to x, and 𝜭 refers to the learnable parameters of the network. By continuously updating 𝜭, the fitted data distribution continuously approaches the real data distribution.
[0053] The logarithmic probability of complex motion behavior can then be expressed as:
[0054]
[0055] In the above formula, the first term It is the log-likelihood of the standard normal distribution, the second term is a scalar value of a Jacobian matrix, which reflects the transformation between actions. Therefore, in order to better map x and z, the loss function in this embodiment is defined as minimizing the negative log-likelihood function:
[0056]
[0057] In the above formula, min represents the minimization operation. This represents the data distribution of the i-th fitted data. Log-likelihood, Here, N represents the learnable parameters of the network, and N is the number of samples taken during execution mode standardization.
[0058] The flow model, as a reversible transformation of learning motion behavior from complex to simple, for the forward process, uses TI and HI as inputs to generate the flow. After multiple steps, the complex motion behavior degenerates into a simple behavior that can be represented by a simple distribution. For the reverse process, once the generating flow model is trained based on all trajectory information and historical observation information, only the normally distributed simple behavior and historical observation information need to be input to accurately fit the current complex maneuver data distribution. A schematic diagram of this process is shown in Figure 4.
[0059] Figure 4 In this context, PN represents mode normalization, which is essentially an affine transformation. Its forward and inverse calculation processes are as follows:
[0060]
[0061] Figure 4 middle, Both b are learnable network parameters, x is the input of PN, and y is the output of PN. Through this affine transformation, we can ensure that it is bidirectionally invertible and that the data has zero mean and unit variance after normalization, which promotes model convergence.
[0062] Next, to ensure the bidirectional invertibility of the entire flow model, an invertible convolutional layer is applied after the PN layer and before the affine coupling layer. The weights of the convolution are initialized with random rotation matrices to ensure invertibility. This transformation can be expressed as:
[0063]
[0064] In the above formula, W represents the parameter matrix of the invertible convolutional layer, y is the output after processing by the PN layer, and f conv This indicates a reversible transformation, and `det` is used to calculate the determinant.
[0065] Calculating the determinants of high-dimensional Jacobian matrices and large matrices is extremely complex. Therefore, this embodiment introduces an affine coupling layer after the PN layer, and constructs the parameter matrix s of the mapping transformation through an MLP and a linear layer.f , through s f Perform matrix multiplication with the input data and continuously update s f The numerical value bypasses the complex calculation of the Jacobian matrix, ultimately achieving the goal of data distribution transformation. It can effectively calculate the forward function, the inverse function, and the logarithmic determinant; the calculation process is as follows:
[0066]
[0067] In the above formula, The `split()` function represents the input of the affine coupling layer. Divide it in half along the channel dimension to obtain and The `concat()` operation performs a concatenation operation. `HI` represents the historical observation information embedded by the `TemporalTransformer` module. f And t are the parameters of the learnable MLP and the linear layer. Because this input It remains invariant during affine transformation operations, so MLPs and linear layers can undergo arbitrary transformations. Here, only... The variable change term alters the dimension of the mapping in the affine coupling layer and is added to the loss function. The logarithmic determinant and final likelihood function of the affine coupling layer are calculated as follows:
[0068]
[0069] In the above formula, s represents the Jacobian matrix corresponding to the affine transformation. g ,W,s f These represent the learnable parameters of the PN layer, the invertible convolutional layer, and the affine transformation layer in this module, respectively. The original data distribution representing the fit. It represents the standard normal distribution.
[0070] The above implements the reversible transformation design of the flow model. The simple behavior derived from the standard normal distribution can be directly transformed into a complex motion behavior trajectory distribution through the inverse process, given historical observation information. Then obtain Input the decoder to predict future trajectories.
[0071] Step 4: Target trajectory prediction based on Bitrap
[0072] In the Bitrap bidirectional trajectory prediction layer, this embodiment designs a bidirectional trajectory prediction module, such as... Figure 5As shown, this module is a bidirectional decoder with target estimation. To mitigate the error accumulation problem caused by RNNs in existing trajectory prediction, this embodiment adopts the decoder proposed by Bitrap, which is designed in a bidirectional manner. It predicts and monitors future trajectories both forward and backward to enhance representation learning during bidirectional prediction.
[0073] First, the target point of the trajectory is predicted using a multilayer perceptron (MLP), that is:
[0074]
[0075] In the above formula, It is a complex motion behavior representation that evolves from simple behavior by the reverse process of trajectory distribution modeling layer, where n represents the time node and is the last moment to be predicted.
[0076] Then, forward trajectory prediction is performed:
[0077]
[0078] In the above formula, The output of the fully connected layer (FC) at time t. This represents the output of the gated recurrent unit (GRU) at time t, which receives fh and fi from the previous time step as input. , , concat indicates a concatenation operation. This represents the final trajectory point predicted by forward propagation.
[0079] Next, reverse trajectory prediction is performed, as shown below:
[0080]
[0081] In the above formula, Represents the reverse time point. Indicates t b The output of the time-gated loop unit receives b from the next time step. i and b h As input, from time n-1 to time T+1, , concat indicates a concatenation operation. This represents the final trajectory point predicted in reverse.
[0082] Finally, consider the bidirectional prediction of both forward and backward trajectories:
[0083]
[0084] Here we obtain the loss functions for forward, backward, and bidirectional trajectory prediction:
[0085]
[0086] In the above formula, and These represent the actual target point and the predicted target point, respectively. Let be the actual trajectory point at time t. Let t be the forward predicted trajectory point. Let t be the reverse-predicted trajectory point. Let K be the bidirectional predicted trajectory point at time t, where t = T+1, T+2, ..., n, and K represents the final number of predicted trajectories. and The coefficients are used to balance different losses.
[0087] Combining the log-likelihood loss generated by the flow model to produce motion behavior, the final loss function is obtained as follows:
[0088]
[0089] During the inference phase, simple behaviors are sampled from the standard normal distribution to generate diverse trajectories by reversing the flow model under environmental interaction conditions.
[0090] This application employs a novel network architecture to achieve trajectory and intent prediction, expanding the application scope of neural networks in trajectory prediction. Model training and validation were conducted using data from an aircraft simulation platform, divided into 1379 training samples and 279 validation samples. Ultimately, the average trajectory prediction accuracy on the validation samples reached within 4 meters, basically meeting the current operational needs of unmanned aerial vehicles (UAVs).
[0091] Evaluation metrics: (1) ADE (Average Displacement Error): the average Euclidean distance between the predicted 40 frames and the input 40 frames.
[0092] (2) FDE (Final Displacement Error): Only considers the Euclidean distance between the last points.
[0093] (3) Time series trajectory complexity (TC): Q is defined as the distance between two adjacent points in the time series trajectory, and its formula is as follows:
[0094]
[0095] The time series trajectory complexity is calculated as follows:
[0096]
[0097] The larger the TC, the more complex the flight trajectory of the sample.
[0098] The following comparison of this embodiment with the traditional Bitrap in terms of the above metrics is shown in Table 1:
[0099] Table 1: ADE and FDE Indicators
[0100]
[0101] The improved method in this patent outperforms the classic Bitrap method in all FDE metrics.
[0102] Figure 6 The diagram shows the ADE / TC of the method in this embodiment. It can be seen that even with the increase in trajectory complexity, the ADE remains at a low level, indicating that the method of the present invention can be applied to complex aerial game scenarios.
[0103] The following section selects several common maneuvers to visually represent the prediction results of this method, such as... Figure 7 and Figure 8 As shown, Figure 7 This indicates the prediction results for relatively simple maneuvers, such as horizontal climb, horizontal left turn, horizontal right turn, and horizontal descent. Figure 8 This indicates the prediction results for more complex maneuvers, such as climbing followed by tail descent, descent followed by a turn and climb. Figure 7 and Figure 8 This demonstrates that, regardless of whether it is a relatively smooth simple maneuver or a violent and complex maneuver, this embodiment can model the motion distribution well and make relatively accurate trajectory predictions based on it.
[0104] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An intent and trajectory prediction method based on the Transformer-Glow model, characterized in that... This includes the following steps: A trajectory prediction model based on a neural network is constructed. The neural network model includes a Temporal Transformer feature extraction layer, a Glow trajectory distribution modeling layer, and a Bitrap bidirectional trajectory prediction layer. The observation information and all trajectory information are input into the Temporal Transformer feature extraction layer, and the Temporal Transformer feature extraction layer outputs a high-dimensional abstract representation of the fusion information of global trajectory and attitude angle; The Glow trajectory distribution modeling layer generates target intent based on the global trajectory information extracted by the Temporal Transformer feature extraction layer, and then the future trajectory is obtained from the intent, realizing the reversible transformation between simple and complex behaviors; The Bitrap bidirectional trajectory prediction layer first uses the observation trajectory distribution information obtained from the Glow layer to predict the target endpoint of the UAV, then performs forward prediction from the current position and backward prediction from the endpoint, and finally combines the bidirectional prediction results to output the final predicted trajectory.
2. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1, characterized in that, The Temporal Transformer layer includes a data embedding module and a multi-head attention mechanism module; The data embedding module is used to map the input low-dimensional parameter information to a high-dimensional space and to calculate the position code and center code. The multi-head attention mechanism module is used to learn different behavioral representations from the input information, which makes up for the defect of single-head attention being overly focused on its own position. It can improve the ability to capture effective information, and then combine different behavioral representations as knowledge to capture dependencies of various ranges within the sequence (e.g., short-distance dependencies and long-distance dependencies).
3. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1, characterized in that, The Glow trajectory distribution modeling layer includes PN layer used for pattern normalization; The reversible convolutional layer located after the PN layer is used to ensure the reversible transformation of the generated flow model, enabling a bidirectional transformation from action distribution information to Gaussian distribution. The affine coupling layer, located after the convolutional layer, avoids the calculation of the determinant of the Jacobian matrix in high dimensions, thus reducing the computational complexity of the model.
4. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1, characterized in that, The Bitrap bidirectional trajectory prediction layer includes a bidirectional trajectory prediction module, which has a bidirectional decoder for target estimation. The trajectory prediction module includes: Multilayer perceptron (MLP) for predicting target trajectory destination points; Gated recurrent unit (GRU) used for modeling the extraction of forward and backward information respectively. A fully connected layer (FC) used for data mapping, information extraction, and fusion.
5. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1 or 4, characterized in that, The prediction method of the Bitrap bidirectional trajectory prediction layer includes the following steps: First, based on the complex motion behavior that evolves from simple behavior output by the flow model, the target trajectory and destination point are predicted by a multilayer perceptron. Then, forward trajectory prediction is performed to obtain the forward trajectory from the observed trajectory point at the last moment to the destination point; Next, reverse trajectory prediction is performed to obtain the reverse trajectory from the target point to the observed trajectory point at the last moment. Then, a bidirectional prediction is performed, considering both the forward and backward trajectories simultaneously, to obtain the final bidirectional predicted trajectory.
6. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1, characterized in that, The observation information includes the geocentric Cartesian coordinates of the target UAV, Euler angles, azimuth angle, and velocity of the target UAV.
7. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 6, characterized in that, The Euler angles include roll angle, pitch angle, and yaw angle.
8. The intention and trajectory prediction method based on the Transformer-Glow model according to claim 1, 6, or 7, characterized in that, The complete trajectory information includes all pose information within the UAV's motion cycle.