Identification of safety-critical events using kinematic vehicle data and the discrete fourier transform

Recent technological advances have made it both feasible and practical to identify unsafe driving behaviors using second-by-second trajectory data. Presented in this paper is a unique approach to detecting safety-critical events using vehicles' longitudinal accelerations. A Discrete Fourier Transform is used in combination with K-means clustering to flag patterns in the vehicles' accelerations in time-series that are likely to be crashes or near-crashes. The algorithm was able to detect roughly 78% of crasvoid(0)hes and near-crashes (71 out of 91 validated events in the Naturalistic Driving Study data used), while generating about 1 false positive every 2.7h. In addition to presenting the promising results, an implementation strategy is discussed and further research topics that can improve this method are suggested in the paper.