Abstract:
|
This project aims to identify similarities between trips with risky driver behaviors based on time-series sensor data. Our study uses the SHRP2 naturalistic driving study (the largest of its kind in the US). Study participants were monitored while driving using cameras, sensors and other devices. We recorded a multivariate time series of sensor data for each trip and coded manually to determine whether the driver was engaged in one of the 55 risky behaviors or driving regularly. We perform a pairwise comparison of all trips using dynamic time warping. We find the pairs of trips that are most similar and use this information to deduce which driver behaviors produce time series sensor data with similar patterns. We select the top 500 pairs and show using a chi-squared test that the distribution of these pairs is significantly different from the distribution of all pairs. We find that some behaviors, like speeding or drowsiness, present like regular driving. Other behaviors, like stop sign violations or failing to signal, present like distracted driving. Using this information to consolidate behaviors will produce a more accurate predictive model for risky driver behaviors.
|