Estimation of lane data-based features by odometric vehicle data for driver state monitoring

Abstract

It is assumed that approximately one third of severe car accidents are related to drowsiness. Warning systems such as the Mercedes Benz Attention Assist try to tackle this problem by analyzing the driving style. Previous work investigated the estimation of measures (features) from lane data that correlate well with impaired driving. Unfortunately, these features require a lane-tracking camera, which is not available in many cars. Furthermore, the lane data signals are often affected from missing road markings, bad sight etc. Some lane-based features such as LANEDEV or ZIGZAGS do not require the absolute distance to the lane markings, but only depend on the lateral deviation within the lane. Our idea is to exploit odometric data (yaw rate and vehicle speed) to estimate this measure. The vehicle trajectory is a composition of the lurching between lane markings and the disturbing road curvature. Thus, we remove this curvature by a filter since its frequency is lower than the vehicle deviation. We compare the correlation between features based on lane data and odometric data as well as their relationship with sleepiness. An excerpt of the Attention Assist database with 294 drives and over 76 000 km is used. We show that some lane-based features can be approximated well. The zero-crossing rate (LATPOSZCR) performs even better than its lane-based pendant.