Probabilistic collision risk assessment via lane-based motion prediction of surrounding vehicles

Abstract

Since the vehicle safety standard has been globally tightened to reduce traffic accidents, research towards the autonomous driving has been actively done over the last few years. However, the capability of coping with dangerous situations is still insufficient for improving the reliability of autonomous driving vehicles. To solve this problem, system should detect future hazardous driving situations and have enough time before initiating an evasive action to avoid accidents. Thus, a reliable risk assessment algorithm is a key enabling technology which allows to predict how collision risk will evolve in the future. Therefore, in this thesis, we propose the risk assessment algorithm that can accurately estimate the collision probability of each local path candidate via lane-based probabilistic motion prediction of surrounding vehicles. In contrast with the way of using directly the Gaussian distribution for the error propagation along the predicted trajectories, we estimate the future motion uncertainty in the level of lane that which lane the vehicle is likely to proceed in near future. The advantage of this method is that the driving road geometry such as the lane number and the width can be inherently considered in aspect of modeling the motion uncertainty. Here, this algorithm is called as the probabilistic target lane detection and can compute probabilities representing where the surrounding vehicle is likely to move into a certain lane in the same framework of Interacting Multiple Model (IMM) approach except for the filtering step. This is composed of multiple lane models corresponding to the lateral offset from the road centerline and the lateral velocity-dependent transition probability matrix. And then, the collision probability of each local path candidates is computed by incorporating both the time-to-collision between a pair of trajectories and model probability distribution of lanes. Time-to-collision is calculated by checking whether the rectangular shape of two vehicles overlap or not at each discrete time step by applying the separating axis theorem. Also, this quantity is converted into the collision risk value by using an exponential function with a constant rate of risk regarding time. Finally, collision probabilities obtained above are plotted on a trajectory plane that is composed of the each set of the tangential acceleration and the final lateral offset of local path candidates. This collision risk map is able to provide the intuitive risk monitoring as well as be utilized to determine a control strategy for collision avoidance. The model validation are conducted by comparing the model probabilities obtained from the probabilistic target lane detection with maneuver probabilities derived from the NGSIM database. Also, the effectiveness of the proposed risk assessment results is verified in rear-end and cut-in driving scenarios on the curved highway with multiple surrounding vehicles.