Adaptive behaviour selection for autonomous vehicle through naturalistic speed planning

Abstract

As autonomous technologies in ground vehicle application begin to mature, there is a greater acceptance that they can eventually exhaust human involvement in the driving activity. There is however still a long way to go before such maturity is seen in autonomous ground vehicles. One of the critical limitations of the existing technology is the inability to navigate complex dynamic traffic scenarios such as non-signalised roundabouts safely, efficiently and while maintaining passenger drive comfort. The navigation at roundabouts has often been considered as either a problem of collision avoidance alone or the problem of efficient driving (reducing congestion). We argue that for any autonomous planning solution to be accepted for replacing the human driver, it has to consider all the three objectives of safety, efficiency and comfort. With human drivers driving these complex and dynamic scenarios for a long time, learning from the human driving has become a promising area of research. In this work, we learn human driver's longitudinal behaviours for driving at a non-signalised roundabout. This knowledge is then used to generate longitudinal behaviour candidate profiles that give the autonomous vehicle different behaviour choices in a dynamic environment. A decision-making algorithm is then employed to tactically select the optimal behaviour candidate based on the existing scenario dynamics. There are two important contributions in this paper, firstly the adaptive longitudinal behaviour candidate generation algorithm and secondly the tactical, risk aware, multi-objective decision-making algorithm. We describe their implementation and compare the autonomous vehicle performance against human driving.