A Predictive Perception Model and Control Strategy for Collision-Free Autonomous Driving

Abstract

A key issue in autonomous driving is the problem of decision logic, particularly, as it pertains to mixed traffic involving autonomous and human-driven vehicles. With this in mind, we develop an approach for modeling the interaction between autonomous and human-driven (or perhaps other autonomous) vehicles via game theory and assess the time evolution of potential collision paths via a reachability analysis of the solution set of a simple hybrid dynamical system that captures the kinematics of the lane-change process. We further utilize a control strategy based on model predictive control to develop a safety-assured motion profile on the part of the subject vehicle. This model is subsequently evaluated in multiple simulation scenarios as a function of a programmable safety assurance parameter. We should emphasize that the proposed approach is proactive in its outlook, as it considers a temporal horizon spanning several seconds and assumes a certain behavioral consistency on the part of the local traffic participants. As such it must be augmented with an appropriate reactive collision mitigation scheme that responds to immediate threats caused by abrupt behaviors that are not predictable via algorithmic schemes. Nevertheless, the scenarios studied here appear to produce intuitively reasonable behaviors that seem consistent with ordinary driving and highlight the prospects of this approach being implemented in autonomous driving strategies.