Development Of Safe Autonomous Driving Systems Under Perception And Motion Uncertainties

The autonomous driving(AD)technology has huge potentials in dramatically improving transportation efficiency and convenience as well as urban traffic conditions,leading to various positive economic and social impacts.Recently,it has attracted widespread attentions and is creating a new industry.Above all,the safety guarantee of AD is among the top issues to be addressed.In real-world AD applications,uncertainties in perception,transmission,and execution will reduce vehicular safety.Besides,uncertainties of individual modules will be propagated into the whole AD system,resulting in various safety issues.This thesis aims to develop a set of novel methods for robust motion planning despite perception and motion uncertainties.The novelty of this work is twofold:(1)developing a Bayesian deep neural network model which can generate Gaussian models of two kinds of perception uncertainties: data and model uncertainties;(2)developing a perception uncertainty rapid random tree(PU-RRT)motion planning algorithm based on Gaussian uncertainty models of both obstacles and ego-vehicle.The developed algorithms have been verified using the CARLA simulation platform with various testing scenarios.The experimental results show that the developed motion planning scheme can cope with the impact of both perception and motion uncertainties,and improve the safety of autonomous vehicles while ensuring operating efficiency.