Generation of Pedestrian Scenarios for Autonomous Vehicle Testing

Abstract

As fully autonomous driving is introduced on our roads, the safety of vulnerable road users is of the greatest importance. The variety of real-world data required for testing autonomous vehicles is limited and, furthermore, available data does not present a fair representation of different scenarios and rare events. A safety threshold with regards to pedestrians and vulnerable road users must be reached before autonomous vehicles are deployed. In recent years, pedestrian deaths have been on the rise in the developed world, indicating that more needs to be done to protect vulnerable road users. There currently exists no reliable way to generate pedestrian scenarios when crossing the road, and this thesis aims to change that. There has been some research presented on human pose estimation and generating a human pose, but not to the extent of generating crossing scenarios. This thesis presents a method for generating pedestrian scenarios based on a Generative Adversarial Network named the Ped-Cross GAN. The Ped-Cross GAN can generate crossing sequences of pedestrians in the form of human pose sequences. The Ped-Cross GAN is trained with the novel Pedestrian Scenario dataset. The novel Pedestrian Scenario dataset, derived from existing datasets, contains data regarding a pedestrian’s movement, behaviour, and other characteristics, enabling for training on more valuable pedestrian scenarios. Examples of its use are demonstrated through several scenarios trained on Ped-Cross GAN. The Ped-Cross GAN is able to generate new pedestrian crossing scenarios that are of the same distribution as those contained in the Pedestrian Scenario dataset. The validated results show that the Ped-Cross GAN generates samples with an error of just 0.48% when classifying samples as crossing from the left or crossing from the right. This will allow for adequate testing, and a greater test coverage when testing the performance of autonomous vehicles in pedestrian crossing scenarios. Moreover, the Ped-Cross GAN and Pedestrian Scenario dataset show that it is possible to use available data to generate scenarios that are scarcely observed in the wild. Ultimately, this will lead to fewer pedestrian casualties on our roads.

Date of Award	Nov 2021
Original language	English
Awarding Institution	Coventry University
Sponsors	HORIBA MIRA Ltd.
Supervisor	Vasile Palade (Supervisor), Stratis Kanarachos (Supervisor), Alireza Daneshkhah (Supervisor) & Madeline Cheah (Supervisor)