CN-122021245-A - Intention and track prediction method based on transducer-Glow model

CN122021245ACN 122021245 ACN122021245 ACN 122021245ACN-122021245-A

Abstract

The invention relates to an intention and track prediction method based on a transducer-Glow model, which comprises the steps of constructing a track prediction model based on a neural network, inputting observation information and all track information into a Temporal Transformer feature extraction layer, outputting global track information by a Temporal Transformer feature extraction layer, generating target intention by the Glow track distribution modeling layer according to the global track information extracted by the Temporal Transformer feature extraction layer, generating a future track by intention evolution, realizing reversible transformation between simple behaviors and complex behaviors, and predicting a target end point of an unmanned aerial vehicle by the Bitrap bidirectional track prediction layer through the historical track, carrying out forward prediction from a current position and reverse prediction from the end point, and finally outputting a final predicted track by combining a bidirectional prediction result.

Inventors

WEI MINGZHU
JIANG JUNZHE
LUO JIN
LIANG QING
WANG ZHENG
HUANG HONGNING
XIE XIAOMEI
XU LIMEI
LIU QIANG

Assignees

电子科技大学

Dates

Publication Date: 20260512
Application Date: 20251223

Claims (8)

1. The intention and track prediction method based on the transducer-Glow model is characterized by comprising the following steps of: constructing a track prediction model based on a neural network, wherein the neural network model comprises a Temporal Transformer feature extraction layer, a Glow track distribution modeling layer and a Bitrap bidirectional track prediction layer; Inputting Temporal Transformer the observation information and all track information into a feature extraction layer, and outputting a high-dimensional abstract representation of fusion information of a global track and an attitude angle by a Temporal Transformer feature extraction layer; The Glow track distribution modeling layer generates target intention according to the global track information extracted by the Temporal Transformer feature extraction layer, and then the intention evolves to obtain future tracks, so that reversible transformation between simple behaviors and complex behaviors is realized; the Bitrap bidirectional track prediction layer predicts a target end point of the unmanned aerial vehicle through observation track distribution information obtained by the Glow layer, then carries out forward prediction from the current position and reverse prediction from the end point, and finally combines a bidirectional prediction result to output a final prediction track.
2. The method for predicting intent and trajectory based on a transducer-Glow model of claim 1, wherein the Temporal Transformer layers include a data embedding module and a multi-headed attention mechanism module; the data embedding module is used for mapping the input low-dimensional parameter information to a high-dimensional space and calculating position codes and center codes; The multi-head attention mechanism module is used for learning different behavior representations from input information, overcomes the defect that single-head attention is excessively focused on the position of the single-head attention mechanism module, can improve the grabbing capacity of effective information, then combines the different behavior representations as knowledge, and captures the dependency relationships (such as short-distance dependency and long-distance dependency) of various ranges in a sequence.
3. The method for intent and trajectory prediction based on a transducer-Glow model of claim 1, wherein the Glow trajectory distribution modeling layer comprises PN layer for mode normalization; A reversible convolution layer located after the PN layer, the reversible convolution layer being used to ensure that reversible transformation of the flow model is generated, the bidirectional transformation from the motion distribution information to the Gaussian distribution can be realized; And the affine coupling layer is positioned behind the convolution layer, so that the affine coupling layer avoids the calculation of the jacobian matrix determinant under high dimensionality and reduces the calculation complexity of the model.
4. The fransformer-Glow model-based intent and trajectory prediction method of claim 1, wherein Bitrap bi-directional trajectory prediction layer includes a bi-directional trajectory prediction module having a bi-directional decoder of target estimates; the track prediction module comprises: a multi-layer perceptron (MLP) for predicting target track destination points; A gate control loop unit (GRU) for respectively extracting modeling from the forward and backward information; full connectivity layer (FC) for data mapping, information extraction, and fusion.
5. The method for predicting intent and trajectory based on a fransformer-Glow model as claimed in claim 1 or 4, wherein the method for predicting Bitrap bi-directional trajectory prediction layer comprises the steps of: Firstly, predicting a destination point of a target track through a multi-layer perceptron based on complex motion behaviors which are output by a stream model and are evolved from a simple line; then, forward track prediction is carried out, and a forward track from the track point at the last observed moment to the destination point is obtained; then, carrying out reverse track prediction to obtain a reverse track from the destination point to the observed track point at the last moment; Then, bi-directional prediction is performed taking both forward and backward trajectories into account, resulting in a bi-directionally predicted final trajectory.
6. The method for predicting intent and trajectory based on a transducer-Glow model of claim 1, wherein the observed information includes three-axis coordinates of a geocentric rectangular coordinate system of the target unmanned aerial vehicle, euler angles, azimuth angles and speeds of the target unmanned aerial vehicle.
7. The method for predicting intent and trajectory based on a transducer-Glow model of claim 6, wherein the euler angles include roll angle, pitch angle, yaw angle.
8. The method for predicting intent and trajectory based on a transducer-Glow model of claim 1, 6 or 7, wherein the total trajectory information includes all pose information within a period of unmanned aerial vehicle motion.

Description

Intention and track prediction method based on transducer-Glow model Technical Field The invention relates to the technical field of unmanned aerial vehicle end-to-end intention and track prediction, in particular to an intention and track prediction method based on a transducer-Glow model. Background In the military field, unmanned aerial vehicles are widely used for tasks such as reconnaissance, surveillance, information collection, striking, and the like. The unmanned aerial vehicle is accurately predicted, the next moment and even the tracks of a plurality of moments of the unmanned aerial vehicle are predicted, the unmanned aerial vehicle on the own side and an operator can be assisted to better predict the subsequent state of the unmanned aerial vehicle when executing tasks, the subsequent position and intention of an enemy aircraft can be also better predicted in the countermeasure, and the like, and the unmanned aerial vehicle on the own side has great significance for executing various tasks and preventing the enemy unmanned aerial vehicle. In the field of unmanned aerial vehicle trajectory prediction, conventional methods such as kalman filtering (KALMAN FILTER), bayesian networks (Bayesian networks), markov Models (Markov Models), and dynamics Models based on physical modeling have been widely used for state estimation and trajectory estimation. The method generally depends on explicit system modeling and statistical assumption, has stronger interpretability and instantaneity, and is suitable for scenes with smaller noise and relatively simple environment structure. However, the conventional method is not attractive in view of the high nonlinearity of unmanned aerial vehicle motion, the diversity of flight strategies and the complex environmental interaction factors in practical applications. The modeling method has the main limitations that modeling capability is limited, complex nonlinear dynamics cannot be effectively processed, priori assumptions are relied on, performance is highly dependent on accurate modeling of a noise model and a system model, multi-mode tracks are difficult to express, prediction results are often single tracks, behavior uncertainty is ignored, and expansibility is weak, so that collaborative prediction requirements in a multi-unmanned-plane system or a complex three-dimensional space are difficult to adapt. In contrast, in recent years, neural network-based learning methods, particularly cyclic neural networks (RNNs), long-short-term memory networks (LSTM), and Transformer structures, exhibit significant advantages in unmanned aerial vehicle trajectory prediction. The method has the characteristics that potential modes can be directly learned from historical track data without explicit modeling system dynamics, strong nonlinear modeling capability is achieved, adaptability is high, multi-mode track distribution can be captured, various possible tracks are generated through probability modeling or generating models, multi-source information (such as sensor data, map information and wind field models) is easy to fuse, and prediction robustness is improved. In terms of how to accurately extract feature information in original data in the case of application of a neural network to track prediction, the accuracy of final prediction is determined, and many track prediction methods still use a cyclic neural network to extract features at present, which cannot be processed in parallel, have low time efficiency, and have obvious forgetting problems with the increase of time length. The method adopts a transducer architecture to extract high-order information in the original track, has quick parallel processing, and effectively reduces the forgetting problem due to the existence of a self-attention mechanism. Meanwhile, the current track prediction for the aerial target is mostly based on the data, and the track is generated directly, so that the method loses the high-order statistical information contained in the original track data. Disclosure of Invention The invention provides an intention and track prediction method based on a transducer-Glow model for solving the technical problems. The invention is realized by the following technical scheme: According to the intention and track prediction method based on the transducer-Glow model, a bidirectional track prediction module Bitrap model is used as a prediction core, a target end point of the unmanned aerial vehicle is predicted through a historical track, forward prediction is performed from a current position, reverse prediction is performed from the end point, a final prediction track is output finally by combining a bidirectional prediction result, a time dependent relation of flight parameter data of the historical target unmanned aerial vehicle is extracted by combining a Temporal Transformer architecture, and track distribution of the unmanned aerial vehicle based on historical observation data is fitted by combining a