Machine Learning for Predicting the Impact Point of a Low Speed Vehicle Crash

Using time series in-car data, this research focuses on predicting the point of impact of a low speed crash by developing an automatized machine learning approach for time series applications. After an initial data exploration, we discuss the extraction of features from time series and different ways to select the most relevant features. From 3,176 extracted features 9 are selected and used for a classification with a decision tree. To optimize the hyper-parameters of the decision tree algorithm, a randomized search with 50,000 iterations is conducted. The modeling results are graphically presented and discussed. With a final prediction accuracy of 89% (cross-validated 76%), the optimized decision tree offers great potential for utilization in vehicle insurance processing for automatized settlement of low-speed crash damages.