Compression of Vehicle Trajectories with a Variational Autoencoder

Abstract

The perception and prediction of the surrounding vehicles’ trajectories play a significant role in designing safe and optimal control strategies for connected and automated vehicles. The compression of trajectory data and the drivers’ strategic behavior’s classification is essential to communicate in vehicular ad-hoc networks (VANETs). This paper presents a Variational Autoencoder (VAE) solution to solve the compression problem, and as an added benefit, it also provides classification information. The input is the time series of vehicle positions along actual real-world trajectories obtained from a dataset containing highway measurements, which also serves as the target. During training, the autoencoder learns to compress and decompress this data and produces a small, few element context vector that can represent vehicle behavior in a probabilistic manner. The experiments show how the size of this context vector affects the performance of the method. The method is compared to other approaches, namely, Bidirectional LSTM Autoencoder and Sparse Convolutional Autoencoder. According to the results, the Sparse Autoencoder fails to converge to the target for the specific tasks. The Bidirectional LSTM Autoencoder could provide the same performance as the VAE, though only with double context vector length, proving that the compression capability of the VAE is better. The Support Vector Machine method is used to prove that the context vector can be used for maneuver classification for lane changing behavior. The utilization of this method, considering neighboring vehicles, can be extended for maneuver prediction using a wider, more complex network structure.