Generating Realistic Aircraft Trajectories Using Generative Adversarial Networks

The evaluation of target tracking algorithms is highly dependent on the available data. Ideally recordings from various situations of interest, ranging from low manoeuvring targets up to closely spaced and highly manoeuvring targets, should be evaluated. Obtaining such data sets is difficult, costly and sometimes could not be done for security reasons. In this paper we show the initial result of a data driven approach for target kinematic behaviour modelling. A method based on GAN is presented and trained on more than 33 000 aircrafts. The results are discussed, highlighting the contributions and drawbacks of the current implementation. The presented method could serve as an alternative to model based trajectory generation with the advantages of maintaining the temporal manoeuvre correlations. As proposed, the methodology could also be used for data anonymization which would allow to build data driven models even from restricted data.