A Risk-Avoiding Disengagement Frequency Model for Assessing Driving Ability of Autonomous Vehicles in Road Testing

Abstract:

To fill the gap that the traditional autopilot disengagements （ADs） frequency model has underestimated the driving ability of autonomous vehicles in road testing， this paper first proposed the concepts of risk-avoiding disengagements and non-risk-avoiding disengagements and then determined the disengagement temporal threshold by using the nonparametric rank sum test. Three characteristic values including average speed difference （ASD）， instantaneous speed difference （ISD）， and short-term speed difference （SSD） were considered as the indicators to quantify the differences of the behavior of vehicles before and after disengagements. Thus， the K-means clustering algorithm in unsupervised learning was used to distinguish the two ADs types. The risk-avoiding disengagement frequency （RADF） model for assessing the driving ability of autonomous vehicles in road testing was finally created. Based on the empirical data covering scenarios of urban road and expressway in Shanghai， the RADF model is verified. The results show that based on the RADF model the driving ability of autonomous vehicles in road testing has averagely been increased by 4.8 and 7.3 times respectively in the urban road scenario and expressway scenario， compared with the traditional ADs frequency model.

Key words:transportation;risk-avoiding disengagement frequency model;unsupervised learning;autonomous vehicle road testing;driving ability