Real-Time Crash Severity Estimation with Machine Learning and 2D Mass-Spring-Damper Model

Making the right decisions milliseconds before an unavoidable car accident is a challenging, nearly impossible task for a human driver. While it is difficult enough to choose a proper maneuver for oneself in the limited time, most of all, it remains unknown how the opponent will react. Thus, an optimal decision can only exist as a probabilistic quantity. One way to estimate the latter is to evaluate a large number of driver actions. With the number of notable combinations easily reaching the mark of thousands of crash constellations, even the most powerful PCs take minutes, hours or even days to finish simulations such as the highly accurate Finite Element Method (FEM). Too long for a decision making only milliseconds ahead of the collision. In this paper, a real-time capable approach with two parallel paths is proposed. Path A consists of a machine learning component that is trained with FEM-data, enabling the prediction of even complex crash severity measures in less than one millisecond. Path B contains a 2D mass-spring-damper model for estimating the crash forces and accelerations and thus, acting as a fall-back and plausibilization layer for the machine learning. A FEM-database of car-to-car collisions is used to train the machine learning model and tune the parameters of the 2D mass-spring-damper model. Together, both paths provide two diverse perspectives of the same crash scenario, leading to an accurate and reliable prediction result in realtime.