Towards Multimodal Driver’s Stress Detection - Digital Signal Processing for In-Vehicle Systems and Safety - page 5

Fig. 1.2 Data collection: residential (left) and business (right) routes segmented according to
assigned tasks
The UTDrive corpus includes data from the above-mentioned sensor channels
(13 separate data streams: two video, six audio, one GPS, one optical distance, one
CAN-Bus, two pressure sensors on gas/brake). The corpus is organized to have a
balance in gender (37 males, 40 females), age (18-65), and different experience
level (novice-expert) in driving. In order to examine the effect of distraction
and secondary common tasks on these driver groups, a close-to-naturalistic data
collection protocol is used.
The routes taken during data collection are given in Fig. 1.2 , comprising a
mixture of secondary, service, and main roads in residential (left-hand side map)
and business (right-hand side map) districts in Richardson, TX. Each driver
participating in the study is required to drive these two routes at least twice in
each session to obtain a baseline and a distracted version of the same route.
A session includes a mixture of several secondary tasks as listed in Table 1.1 ,taking
place in road segments depicted in Fig. 1.2 . According to this protocol, a participant
performs 12 runs of data, with six being baselines for that day and that route, the other
half featuring several distraction conditions. Each session is separated at least by
2 weeks in order to prevent driver complacency with the route and vehicle. Almost
60% of the data in the corpus have a full session profile from drivers. The remaining
part contains incomplete sessions and data portions due to the consent of the partici-
pant not to continue data collection or several sensor failures. The secondary driver
tasks are low to medium level of cognitive load while driving.
In this study, cell phone dialog parts including interaction speech with automated
portals Tell-Me (information system) and American Airlines (reservation system)
are utilized and analyzed using driver's speech and CAN-Bus signals. The cell phone
conversation takes place in route segment two which includes lane keeping and
lane curvature negotiation tasks while the driver is engaged in cell phone dialog.
In order to segment the data in terms of driving event and task timelines and find
overlapping portions, two different transcription protocols are applied. First, using the
audio and video, a task transcription is performed, having 13 labels to annotate the
segments of the data in terms of where the driver and passenger talk and where other
types of distractions occur. The second is called “event transcription” and performed