3.4 Context-driven Camera Switching
In what follows, we present the design of our context-driven
camera switching.
Switching Algorithm. We provide the intuition that underpins
the design of our camera switching mechanism; a formal description
is provided in Algorithm 1.
Primarily, the switching between the front and rear cameras is
regulated by two time intervals (i.e., Tf
e , and Tb
e ). Each interval is
associated with one of the two cameras. Once the time interval for
a particular camera expires CarSafe switches to the other camera.
The pair of time intervals are set by an event arrival prediction process,
which seeks to dynamically adjust the time allocated to each
camera based on the current driver, road or car conditions. For example,
if the system determines that the driver is likely to enter a
dangerously drowsy state more time is allocated to the front camera
to monitor this situation closely. Figure 8 presents an example of
the common case operation of the camera switching mechanism.
Scheduled camera switching based on a pair of timers is supplemented
with temporary pre-emption of the switching plan driven
by discrete events – we refer to this behavior as the pre-emption of
the switching plan. For example, if the driver starts to drive close
to the lane markers then a planned switch to the front camera can
be delayed to allow further monitoring of the lane marker situation.
Event Arrival Prediction. We use two separate predictors that
target two event types: (1) front-camera events, which focus on
driver states; and (2) rear-camera events, which focus on car events
and road conditions.
Front-Camera Events. The driver state event with the shortest
time-scale is driver drowsiness (i.e., blink duration). We find for
this reason the prediction of other driver state events are not needed
and instead we focus on predicting driver drowsiness only. We
again use PERCLOS to quantify drowsiness; the likelihood that a
driver is drowsy increases inline with increases in PERCLOS. As a
result, the time allocated to the rear camera is proportional to the
most recent PERCLOS. More formally,