Generative Adversarial Network for Car Following Trajectory Generation and Anomaly Detection

Abstract

Vehicle trajectory generation and anomaly detection are critical in the sensing and prediction module for automated driving. However, developing models that capture realistic trajectory data distribution and detect anomalous driving behaviors could be challenging. This paper proposes ‘TrajGAN,’ an unsupervised approach based on the Generative Adversarial Network (GAN) to exploit vehicle trajectory data for generation and anomaly detection. TrajGAN consists of an encoder-decoder Long Short-Term Memory (LSTM)based generator and an LSTM-multilayer perceptron (MLP) based discriminator. Trained with the Next Generation Simulation (NGSIM) dataset, TrajGAN can generate realistic trajectories with a similar distribution of training data and identify a manifold of anomalous trajectories based on an anomaly scoring scheme. Applied to new trajectory data, the model scores trajectory sections at the microscopic level, indicating their fit into the learned distribution. Simulated experiments validate that the approach can effectively reproduce artificial trajectories and identify anomalous driving behaviors.

Keywords: Automated Driving, GAN, LSTM, Anomaly Detection