Analyzing Inter-Vehicle Collision Predictions During Emergency Braking with Automated Vehicles.

Automated Vehicles (AVs) require sensing and perception to integrate data from multiple sources, such as cameras, lidars, and radars, to operate safely and efficiently. Collaborative sensing through wireless vehicular communications can enhance this process. However, failures in sensors and communication systems may require the vehicle to perform a safe stop or emergency braking when encountering hazards. By identifying the conditions for being able to perform emergency braking without collisions, better automation models that also consider communications need to be developed. Hence, we propose to employ Machine Learning (ML) to predict inter-vehicle collisions during emergency braking by utilizing a comprehensive dataset that has been prepared through rigorous simulations. Using simulations and data-driven modeling has several advantages over physics-based models in this case, as it, e.g., enables us to provide a dataset with varying vehicle kinematic parameters, traffic density, network load, vehicle automation controller parameters, and more. To further establish the conditions for inter-vehicle collisions, we analyze the predictions made through interpretable ML models and rank the features that contribute to collisions. We also extract human-interpretable rules that can establish the conditions leading to collisions between AVs during emergency braking. Finally, we plot the decision boundaries between different input features to separate the collision and non-collision classes and demonstrate the safe region of emergency braking.