Crash To Not Crash: Playing Video Games To Predict Vehicle Collisions

Kangwook Lee, Hoon Kim, Changho Suh

Abstract: Today's vehicle collision prediction algorithms are rule-based, and have not benefited from the recent developments in deep learning. This is because it is almost impossible to collect a large amount of collision data from the real world. To address this challenge, we collect a large accident data set using a popular video game named GTA V. Using this accident data set, we develop efficient prediction algorithms based on modern CNN architectures. The performances of our algorithms are compared with the simple rule-based algorithms. We observe that the best CNN-based algorithm among several variants achieves the prediction accuracy of 96.2% while the best rule-based one achieves the accuracy of 89.6%. We also show that our approaches can identify the source of danger when a collision is predicted. Moreover, our approaches are shown to learn how the wheel angles, vehicle orientations, and distances affect the collision probability.

TL;DR: We collect accident data from GTA V to build a collision prediction algorithm.

Keywords: Collision Prediction, Deep Learning, CNN