@inproceedings {pub3310,
	title = {Towards an On Demand Intersection Assistant: Initial User Acceptance and System Development },
	author = {Martin Ernst Heckmann AND Heiko Wersing AND Dennis Orth AND Dorothea Kolossa AND Nadja Sch{\"o}mig AND Christian Maag AND Mark Dunn},
	year = {2017},
	month = {September},
	abstract = {In recent years many new Advanced Driver Assistance Systems have been presented. These systems aim to support the driver in the driving task and to reduce her cognitive load. However, as these systems usually do not work flawlessly they also can lead to distraction and annoyance of the driver due to undesired warnings. In an attempt to overcome these limitations we recently developed the concept of {\textquotedblleft}Assistance on Demand{\textquotedblright}. This describes an advanced driver assistance system (ADAS) which supports a driver only if she asks for assistance. The two key elements of this concept are on one hand the control of the ADAS via speech and on the other hand the personalization of the system to the individual driver. The speech-based control allows the driver to flexibly formulate her requests for assistance while the situation develops. The personalization will help to adapt the interaction of the system to the driver{\textquoteright}s individual preferences and skills. Our application scenario is turning left at unsignalized urban intersections. After the driver has activated the system via a speech command it monitors the right side traffic and informs about suitable gaps to enter the intersection, just like a co-driver would do. A typical interaction with the system might look like this:
{\textbullet}	Driver: {\textquotedblleft}Please watch right.{\textquotedblright}
{\textbullet}	System: {\textquotedblleft}Okay.{\textquotedblright}
{\textbullet}	{\textellipsis}
{\textbullet}	System: {\textquotedblleft}Car approaching.{\textquotedblright}
{\textbullet}	{\textellipsis}
{\textbullet}	System: {\textquotedblleft}Gap after next car.{\textquotedblright}
In a first user study we investigated this concept in a driving simulator. We used the static driving simulator of the W{\"u}rzburg Institute for Traffic Sciences. Each of the 24 participants taking part in the study drove three rounds in the simulator. In the first round the driver was driving without system support (manual driving). The system type of the second and third round was randomized. The driver was either supported by our speech-based Assistance On Demand (AOD) system or a visual system which displayed coloured arrows in a virtual Head Up Display (HUD). A red arrow indicated that there is not suitable gap and an orange arrow indicated that there is a suitable gap after the next car.
Figure 1: Preferred drive of the 24 participants of the user study. Without assistance (manual), with speech-based Assistance On Demand (AOD) or visual via Head-Up Display (HUD).
Figure 1 displays which of these three drives the drivers preferred. The results show that most drivers prefer our speech-based system (14 out of 24). Only 3 drivers favored the HUD system. With our speech-based system drivers could focus visually on the part of the environment they currently considered the most relevant while still receiving input from the system via the unoccupied acoustic channel. The visual system on the other hand required them to divert their gaze to see the system response. We assume that this difference is at the heart of the clear preference for the AOD system. The remaining 7 drivers preferred to drive without any assistance.  
Another result of this study was that drivers expressed that the gaps suggested by the system did not perfectly match their own driving behavior. Some wanted larger gaps others felt that they could have used smaller gaps. From this we concluded that a personalization of the suggested gaps to the individual driver might help to further improve the system. To test how drivers differ in what they perceive as a suitable gap to make the left turn we have performed a second simulator study. In this study 9 participants were turning left in crossing traffic from both sides with varying gaps between the cars. The scenario was implemented using CarMaker and our own static low-fidelity driving simulator. Each participant drove 2 rounds with 16 intersections of identical road layout. We deploy a maximum likelihood method to estimate the smallest accepted gap of each driver, so called critical gap. Figure 2 shows the gaps the individual drivers took as well as the gap when the data of all drivers are taken together. The results reveal that there is, as postulated, a significant inter-individual difference in the critical gap between the drivers. This shows that there is indeed a very high potential to improve the system acceptance even further by adapting the gap suggestions to the individual driver{\textquoteright}s driving habits.
The next step is the integration of a running AOD system in our driving simulator. This will be the basis for the implementation of the system on a vehicle. To prepare the vehicle implementation we recorded different urban intersections with varying numbers of drive-throughs. For the recording we used our research platform vehicle. It is based on a XXXX model year Honda CR-V equipped with 3600 sensing from laser, radar and camera sensors. We are currently evaluating the results of these recordings. 

CONCLUSION 
 
Our speech-based Assistance On Demand (AOD) concept allows the driver to request for assistance whenever he deems it appropriate. We investigated the benefits of this approach in two different driving simulator studies. As scenario we have chosen the left turning from a subordinate road in dense urban traffic. In the first study we could show that drivers clearly prefer our proposed speech-based interaction to a visual assistance system or not having an assistance system at all. In a follow up study we investigated the left turning behavior of different drivers. The results showed that there is a large variation in what gaps the individual drivers take. This confirms our hypothesis that a personalization of the intersection assistant has a high potential to further improve usability and driver acceptance. Currently we are building a real-time demonstrator in the driving simulator and analyze the recordings we performed with our research vehicle. The next step will be the integration of the full system in the research vehicle.
},
	publisher = {Society of Automotive Engineers of Japan},
	booktitle = {FAST-zero17}
}