Sensor Modelling with Recurrent Conditional GANs

Abstract

Autonomous vehicles rely on many sensors in order for the vehicles to perceive their surroundings. Consequently it is important with safety verification of the sensors which typically is done by collecting data in many different scenarios which is time consuming and expensive. For this reason, autonomous driving software companies are interested in virtual verification where the scenarios are simulated. In this thesis we have developed and used the Recurrent Conditional Generative Adversarial Network (RCGAN) in order to model the longitudinal error of sensors. The RCGAN is a modification of the original generative adversarial network (GAN) framework which makes use of recurrent neural networks and conditioning the networks on auxiliary information. These changes allows the model to learn and be able to generate realistic real-valued multi-dimensional time series.