Human Behaviour-Based Trajectory Planning for Autonomous Overtaking Maneuver

Abstract

Autonomous driving is a contemporary research area that bears great potential to improve safety and environmental sustainability in the automotive domain. Trajectory planning is an important layer in the hierarchical structure of such autonomous vehicle systems, and has been a topic of extensive research in recent years. This layer is responsible for generating a sequence of continuous states that guides the vehicle through its environment safely, comfortably and following traffic rules. A majority of the currently used trajectory planning algorithms depend on a predetermined reference trajectory generated by GPS or high-resolution vector maps, which often are either unreliable, expensive or simply unavailable. The aim of this thesis is to develop a computationally robust trajectory planning algorithm for overtaking maneuver, that uses naturalistic driving data to generate a human like trajectory to navigate the vehicle through a dynamic environment. A Euler spiral and Markov Decision Process (MDP) based framework is developed that is used to solve the trajectory generation and selection problem. The generated reference trajectory is then input to a local feedback controller to actuate the steering of vehicle. The developed algorithm is implemented on a simulation platform to assess parameters such as safety, passenger comfort and maneuver execution time. The robustness of the algorithm is evaluated by simulating the overtaking maneuver for different velocity profiles of the target vehicle. An unsupervised learning based intent recognition model is proposed to identify and predict lane change intentions of surrounding vehicles, with the aim of further improving the safety of the autonomous vehicle in dynamic environments. The model is evaluated for accuracy against a simple Pseudo-Ground Truth (PGT) model, to validate the promptness of lane change prediction. The overtaking maneuver simulation results demonstrated that the developed algorithm is able to successfully execute the maneuver in all of the test scenarios, maintaining sufficient clearances to the target vehicle and road boundaries while respecting passenger comfort limits. The trained intent recognition model is found to be prompt in issuing lane change warnings as a rear end collision is avoided in ten out of eleven test cases.