Abstract:
Conditional automation is the next step towards the fully automated vehicle. Under
prespecified conditions an automated driving function can take-over the driving task
and the responsibility for the vehicle, thus enabling the driver to perform secondary
tasks. However, performing secondary tasks and the resulting reduced attention towards
the road may lead to critical situations in take-over situations. In such situations, the
automated driving function reaches its limits, forcing the driver to take-over responsibility
and the control of the vehicle again. Thus, the driver represents the fallback level for
the conditionally automated system. At this point the question arises as to how it can
be ensured that the driver can take-over adequately and timely without restricting the
automated driving system or the new freedom of the driver.
To answer this question, this work proposes a novel prototype for an advanced driver
assistance system which is able to automatically classify the driver’s take-over readiness
for keeping the driver ”in-the-loop”. The results show the feasibility of such a
classification of the take-over readiness even in the highly dynamic vehicle environment
using a machine learning approach. It was verified that far more than half of the drivers
performing a low-quality take-over would have been warned shortly before the actual
take-over, whereas nearly 90% of the drivers performing a high-quality take-over would
not have been interrupted by the driver assistance system during a driving simulator study.
The classification of the take-over readiness of the driver is performed by means of machine learning algorithms. The underlying features for this classification are mainly based on the head and eye movement behavior of the driver. It is shown how the secondary tasks currently being performed as well as the glances on the road can be derived from these measured signals. Therefore, novel, online-capable approaches for driver-activity recognition and Eyes-on-Road detection are introduced, evaluated, and compared to each other based on both data of a simulator and real-driving study. These novel approaches are able to deal with multiple challenges of current state-of-the-art methods such as: i) only a coarse separation of driver activities possible, ii) necessity for costly and time-consuming calibrations, and iii) no adaption to conditionally automated driving scenarios.