Adversarial Autoencoder for trajectory generation and maneuver classification

Abstract

For the development of self-driving cars, it is essential to perceive the environment as accurately as possible and to interpret the movement of the surrounding vehicles. It makes sense to draw conclusions based on the past trajectories of these vehicles, whether it is maneuver detection or, in more complex cases, maneuver or trajectory prediction. Trajectories are time series data, so it is obvious to deploy recurrent neural networks for their analysis, or 1-dimensional convolutional networks that can capture temporal patterns. The concept is presented that trajectories starting from the origin can be compressed efficiently so that the reconstructed trajectory is quite similar to the original, and the latent space code obtained by compression can be used for maneuver detection. Using a variational autoencoder, assuming a normal distribution, the latent spatial distribution can be approximated. However, in this article, the goal was to test this concept with adversarial training, so the so-called adversarial autoencoder is trained. It has been shown that this method is suitable for twelve-fold compression of trajectories, and the latent code is suitable for maneuver detection. This proves that the encoder has learned useful features about the distribution that generates trajectories.