Evaluation of Prediction Models in Trajectory Planning for Overtaking Maneuvers.

Sammanfattning: Developing autonomous overtaking could enhance safety by more accurate decision making of if, when and how to execute an overtaking and thereby decrease the collision risk caused by the human factor. Trajectory planning is a strategy to plan overtaking maneuvers, where a trajectory is defined as a path containing position, velocity and acceleration information along with time stamps. The purpose of this study is therefor to improve the decision-making for executing overtakes autonomously. There are several methods mentioned in the literature on how to solve path/trajectory problems. This study will use a state space sampling based motion planner, that optimizes the jerk between the start and end state. The jerk cost function is minimized by the use of quintic polynomial. This approach samples the end position, hence will a quartic polynomial be used instead. The motion planning algorithm will generate a collision free trajectory for the ego vehicle, to overtake an observed vehicle before the two lanes merges to one lane. The scope of this thesis is to evaluate how the three different prediction models CA, CV and IDM in conjunction with the state space sampling based motion planner affects the travel time. Additionally will three different road friction conditions be considered as constraints. The travel time is considered as the time from when the ego vehicle starts the overtaking to when it reaches the merging point. The results of this study showed that dry road condition with least strict constraints, allowed a faster travel time compared to wet road condition with stricter constraints. Same applies for the wet road condition with less strict constraints compared to icy road condition. The closer the two vehicles were to each other, the more challenging was the planning where different outcome scenarios occurred for the different prediction models. However, for the collision free trajectories, without performing an action plan, the results showed the exact same travel time and final trajectory, during the same weather condition. The reason is that among the trajectory set during the collision check, all of the trajectories will be valid or invalid for the different prediction models. If the complete trajectory set is valid, the same fastest trajectory is always selected, hence same travel time. If some of the trajectories were invalid and some valid, there would occur a difference in travel time. Investigations on the spread of trajectories were conducted, without any changes. A future work is to implement a more complex action plan, in that way could the results lead to different travel time.