Development of Machine-Learning Models for Autonomous Vehicle Decisions on Weaving Sections of Freeway Ramps

Abstract

This research aims to develop human data-driven automated lane-change models for freeway weaving sections using computational methods that assist drivers taking an exit ramp or entering a freeway. A naturalistic driving dataset with 108 adult drivers served as the data source to observed drivers’ lane change maneuvers over 53 freeway weaving sections in the southeastern Michigan area. With the Cox proportional hazards model, we could identify at least 83% of weaving initiation time and provided at least 81% of accuracy for the models. The models were further evaluated based on computer simulations, which showed that collisions with the other vehicle in the target lane might occur if the ego vehicle drove with a same speed as that vehicle. Also, the vehicle would possibly decide not to engage a lane change before reaching the end of the weaving section if the driving speed was greater than 70 mph and the other vehicle with 55 mph or higher. As successfully implementing the models to an autonomous driving platform at Mcity, no physical traffic could be applied since the models did not provide a complete collision-free environment. Therefore, an augmented reality environment was adopted, for which the autonomous vehicle interacted with a ghost car simulated by ROS signals and no ‘virtual’ collision was observed in the demonstrations. Further improvement for the models is needed, including the variety of the weaving scenarios from the data for model development and the consideration of speed adjustment for the autonomous vehicle before entering the weaving sections.