| dc.description.abstract | Driver assistance systems are paving the way for automated driving. Until fully autonomous driving will be available and in wide-spread use, assistance systems, such as lane-keeping assistance, can help prevent accidents by supporting a human driver. However, if a lane-keeping assistance system strongly restricts the driver’s autonomy, the driver may overly trust the system (Inagaki & Itoh, 2013) and distract herself more frequently (Llaneras et al., 2013). It is advantageous to base the intensity of assistance on the attentiveness of the driver (see Pohl et al., 2007 and Blaschke et al., 2009). Thereby, the driver is kept in the loop when attentive but is supported during periods of distraction. Driver distraction is a serious issue; 14% of crashes in the USA were affected by distracted driving in 2018 (NHTSA, 2020). Taking into account the driver’s distraction for the activation of assistance technology could help prevent such accidents.
We design an agent as a lane-keeping assistant that shares control of the vehicle with the driver. The driver’s distraction is estimated online, allowing the agent to assist a distracted driver while keeping an attentive driver in control. For the estimation of the driver’s distraction, the agent relies solely on driving performance measures, such as the driver’s steering movements and sensory information about the vehicle’s position. To account for uncertainty about the driver’s distraction and the exact position of the vehicle, the problem is modeled as a Partially observable Markov decision process (POMDP). To the best of our knowledge, this is the first study using only commonly available driving performance metrics instead of sophisticated driver monitoring systems to estimate the driver’s distraction with a POMDP.
We apply the Partially observable Monte-Carlo Planning (POMCP) algorithm (Silver & Veness, 2010) to solve the POMDP online. The algorithm performs Monte-Carlo tree search, sampling possible future scenarios to form a strategy. Our experiments confirm that the driver’s overall lane-keeping performance is significantly enhanced. Our approach has potential. However, there are obstacles that have to be overcome for the method to be viable in practice. First, the solver is not efficient enough; planning takes too much time. Second, we use simple hand-crafted driver models for our experiments. The driver model can and should be replaced by a more sophisticated and realistic model. Third, our method relies on the discretization of the action and observation spaces. A car’s sensory information and steering actions are naturally continuous. We discuss these limitations in detail and provide suggestions for future improvements. | |
