In the foreseeable future, CAVs and human drivers will share the roads. Therefore, the main challenge for CAVs is to figure out how to interact and even cooperate with human drivers for safer and smoother driving. We assume no direct intention exchanges between CAVs and human drivers, unlike what could have been possible with DSRC/CV2X communication on all vehicles. Without V2X, the CAVs will rely on observed motions (positions and velocities) as the input, and a behavior prediction algorithm to predict human drivers’ future movements statistically. With such prediction, the CAV can plan its future trajectory over a horizon. Moreover, the behaviors of CAVs, if not designed to have nice roadmanship, may cause confusion and misjudgment from human drivers. Therefore, if an adequate amount of data is collected and the behavior of other human-driven vehicles can be accurately predicted, it will not only help the CAV to drive more safely but also how to drive with good roadmanship.
To better understand and anticipate human drivers’ behaviors, we need to consider the traffic scenario jointly observed by all the CAVs, when more than one CAV is in a relevant scenario, e.g., in a roundabout or near an intersection. To achieve more reasonable predictions, different from the pure rule-based or data-driven methods, we propose to combine historical trajectories, road network information, interactions among neighboring vehicles, and machine learning approach to predict other vehicles’ behaviors. The prediction should capture human drivers’ natural cooperative driving behaviors established by traffic rules and social norms. As a result, we shall adopt the natural cooperative behaviors of human drivers to design safe and intuitive interactions between CAVs and human drivers. Ideally, the CAVs will just “merge into the local driving culture” and behave like a normal human driver.
The project has three main research tasks:
1) Develop a trajectory prediction algorithm.
2) Develop a cooperative motion planning algorithm for CAVs to improve traffic flow and safety.
3) Validate the proposed trajectory prediction and cooperative motion planning algorithm on the Mcity MKZ platform.
-Learning-based motion prediction algorithm using historical trajectories, roadway information, and vehicle interactions. Graph neural networks will be designed to predict the future trajectory intelligently. The proposed algorithm shall have the ability to capture the vehicle interactions in a wide variety of road topologies.
-A cooperative driving algorithm leveraging the motion prediction algorithm outputs- resulting in safer and smoother driving. The algorithm can be used for CAVs and for information/advice for human driven vehicles.
-Demonstration of the proposed motion prediction and cooperative driving algorithms using the Mcity MKZ platform.