A new method for the objective assessment of ADAS based on multivariate time series classification

Autoren

Uwe Moser, Nick Harmening, BMW Group;
Dieter Schramm, Universität Duisburg-Essen

Jahr

2019

Zusammenfassung

Advanced Driver Assistance Systems (ADAS) such as Adaptive Cruise Control (ACC) are appearing increasingly more often in modern premium cars and have great potential for increasing both comfort and safety. These gains are only appreciated by the customer, when the driving behavior of the system is perceived as positive, e.g., comfortable. Therefore, the different calibrations of an ADAS must be assessed during the development process to ensure positive perceived comfort. However, the general practice is that the new calibrations of an ADAS are assessed subjectively by experts in test drives. Although this system basically works, it has several disadvantages. The perception and consequently the assessment vary from person to person, are dependent on the drivers’ physical and mental conditions (e.g. tiredness and concentration) and also vary with the duration of the experiment (e.g. adaption). As a consequence, it is highly desirable to increase the objectivity in the assessment in order to assess driving comfort based on objectively measurable variables, which are determined in an experiment with a representative group of customers. To fulfill this vision, the subjective perception must be described by objectively measurable variables, which is nontrivial due to the complexity of the human perception process, the variety of driving scenarios and the complexity of the influences on the driver. A method to determine the most influential variables in ADAS perceptions is presented. Based on these variables, it is possible to build models that can predict the subjective perceptions of different ADAS calibrations in order to compare them in an objective way.
The contribution of this paper is to present two Multivariate Dynamic Time Warping-based classifiers for the objective assessment of ADAS, namely, a k-Nearest Neighbors classifier and a kernel density-based classifier. Both classifiers can handle the characteristics of time series, have low user involvement, deal with multivariate relations and lead to a transparent classification.

Mitglieder-Login