Short-Term Collision Estimation by Stochastic Predictions in Multi-Agent Intersection Traffic

Abstract

Multi-agent modeling is suitable to simulate complex interaction dynamics of microscopic urban road traffic. Valuable motion predictions can systematically be generated and exchanged among the participants (agents) to study and quantity benefits of advanced V2X-communication, for example. However, such predictions are inherently uncertain which needs to be considered for traffic safety. This work proposes a stochastic motion prediction and evaluation approach suitable for multi-agent-based simulation and control. Dynamic occupancy probability grid maps are constructed, and their interpretation clearly shows the uncertainty generated by unknown road user intentions or traffic interactions. By formulating joint occupancy probability maps, a quantification of near-accident risk becomes possible which seems to be a promising tool to examine safety aspects in 'non-critical' traffic situations. The studies are based on published naturalistic driving measurement data, and both data-based as well as model-based predictions are discussed.