Vehicle rollover detection using recurrent neural networks

Rollover accidents have a higher fatality rate than other types of accidents. Therefore, rollover prevention systems are of great importance for driver safety. The implementation of rollover prevention systems requires an estimation of the rollover risk. To assess that risk, different rollover indices have been introduced. A difficulty is the dependence of these indices on unknown parameters, e.g., center of gravity and current load of the vehicle. One solution is to implement an algorithm for the estimation of the required parameters that can be online measured. In this work however, we investigate the use of recurrent neural networks for the estimation of the rollover index. Their ability to work on sequential data is promising for a data based estimation without the need of an additional estimation algorithm. We implement and test different recurrent neural network architectures and compare the results with the achievable performance of a static neural network. The results are validated in simulation in the industry standard software CarSim.