Target Trajectory Prediction And Intention Understanding In Adversarial Environment

In adversarial environment,because the interaction between agents is complex,it is very important to predict the trajectory and understand the intention of hostile targets.However,nowadays the results of target intention understanding and trajectory prediction methods in adversarial environment are generally poor.Based on the ICRA DJI RoboMaster Artificial Intelligence Challenge,this paper mainly studies the trajectory prediction and intention understanding of targets in the adversarial environment.The main achievements of this paper are as follows:Firstly,the kinematics model of wheeled robot is established,the coordinate system and transformation relationship are given,the descriptions of the trajectory prediction and intention understanding are given.Secondly,aiming at the complex interaction in the adversarial environment,the encoder module and decoder module for processing time series data are given.In order to reduce the computational complexity,a simplified discriminator network is given.In order to extract the interactive information between agents,a downsampling module between generator encoder and decoder is designed.Aiming at the difficulty of convergence in training process,the method of adjusting parameters and structure is given.Combining each module,a trajectory prediction method based on Generative Adversarial Network is proposed,which learns the implicit interaction between robots through the back propagation of network.Thirdly,a data compression and feature extraction network based on AutoEncoder is given.Aiming at the problem of difficult acquisition of target intention,a deep embedding clustering network is proposed by using the encoder part of feature extraction network,which can realize automatic clustering.According to the result of intention clustering,an intention classification network based on Long Short-Term Memory module is given.Finally,a simulation platform is built and experiments are designed to verify the trajectory prediction method and intention understanding method.In order to quantitatively analyze the accuracy of trajectory prediction method,two indexes are given to measure the accuracy of trajectory prediction.According to the needs of trajectory prediction and intention understanding methods,three training data acquisition methods are given.Three scenario-specific interaction models are designed to verify the accuracy of trajectory prediction method and the expressive ability of intention.Four kinds of agent interaction modes in the environment of simulation platform are given,and the advantages and disadvantages of trajectory prediction method compared with other methods are evaluated.Aiming at the problem that the result of intention classification is difficult to describe,the result of intention classification is mapped to the competition field.The actual physical significance of intention classification results is analyzed.