Definition of Renewal Cycle

To describe the complete process of attention allocation more clearly, this study expanded the concept of the scan path to analyze attention allocation using renewal cycles. A renewal cycle represents the process of shifting vision from a reference point, transiting to other points, then shifting back to the reference point. Identifying a renewal cycle requires determining its beginning and ending points. Moving forward is a major activity of driving; thus, this study regarded the forward area as the focal point at which drivers look naturally and comfortably. The forward area is also the point to which drivers eventually return their attention after shifting it away (Crundall et al. 2006). Therefore, using “forward” as the initial reference point, this study defined a renewal cycle as the driver directing his or her attention forward, transiting to other focal points, and then returning the gaze again to the forward area.

This approach not only distinguishes on- and off-road glances but also represents a complete chain process of the driver shifting attention from one point to another and transiting their vision back to the front. Using the renewal cycle as the basic component of attention allocation facilitates in-depth exploration of drivers’ visual transition characteristics among focal points, especially the transition among non-forward focal points. In addition to the extracted paths transiting toward or from the forward area, this method enables the inclusion of additional serial focal points as a pattern to reflect the entire process of drivers allocating their attention during certain tasks or events. Analyzing renewal cycles can help clarify the interactions between forward and non-forward glances. For instance, drivers employing different strategies of attention allocation by varying the durations of forward and non-forward glances in one renewal cycle may indicate their various ways of searching information.

4.2 100-car Naturalistic Driving Data

This study used the 100-car naturalistic glance data collected by the Virginia Tech Transportation Institute (VTTI) (Neale et al. 2002, Dingus et al. 2006, Klauer et al. 2006). It has been applied in several studies regarding the exploration of drivers behavior and attention allocation (Bagdadi 2013, Dozza 2013, Wu and Jovanis 2013).

The dataset contained 241 drivers in United States driving 100 instrumented cars, on which the sensors, data processors, eye trackers and GPS were installed. There was no experimenters presented during the data collection period. No special instructions were given, except of asking them to drive as they usually did. Detail information regarding the sensors and the data processing could be found in (Dingus et al. 2006).

In total, approximately 2,000,000 vehicle-miles of driving and 43,000 hours of data were collected by VTTI. Two datasets among these data were released online:

event database and baseline database. The baseline database contained only 6 s of glance data in each record, which is insufficient for this analysis. Therefore, this study adopted the event database, which contains 68 crash and 760 near-crash incidents (VTTI 2012). In each incidents, drivers’ visual glances and related attributes for the 30 s before crash or near-crash incidents were recorded. The 30-s duration was divided into two parts according to the precipitating events that were determined as causing the crash or near-crash incidents. Data collected after the precipitating events were related to emergency evasion and crash prevention. Such actions do not represent typical driver behavior. By contrast, data collected before precipitating events could be assumed to contain the time period that drivers were driving without being consciously affected by dangers and should be similar to the sample drivers’

habitual behavior. Therefore, the data before the precipitating event (on average 25 s) were applied for the analyses.

However, the drivers in the 100-car event database ultimately experienced crashes or near-crashes. The results derived in this study only represent the common patterns of a limited sample of drivers’ behavior, which might include potentially risky behaviors. Moreover, the data were collected in United State, where the driving environment, complexity of traffic flow and driving culture are different from ones in Taiwan. Therefore, the pattern observed in this study may not be able to explain the driving behavior in Taiwan. Table 4-1 shows the attributes of the 100-car event database used in this study. Four types of attributes were used: roadway and traffic, driving tasks, environment and eye-glance data.

Table 4-1 Attributes of 100-car event database

Attributes Category

Roadway and Traffic

‧ Relation to junction Non-junction, Intersections (Intersection, Intersection ‐ related, Driveway, alley access), Other (Entrance/exit ramp, Rail grade crossing, Interchange Area, Parking lot)

‧ Traffic density Level‐of‐service A (less than 12 pc/mi/ln), B (12~20 pc/mi/ln), C (20~30 pc/mi/ln), D (30~42 pc/mi/ln), E (42~67 pc/mi/ln)

Driving tasks

‧ Pre-event maneuver Driving straight (Going straight in constant speed, accelerating, but with unintentional "drifting" within lane or across lanes, Decelerating in traffic lane, Starting in traffic lane or Stopped in traffic lane), Lane Change (Passing or overtaking another vehicle, Changing lanes or Merging), Turning left and Turning right

‧ Distraction (time series data) Cognitive, cell phone, in-vehicle devices, external clutter, activity

‧ Turning signal (time series data) Recorded when turning signal (left, right and both) were on.

‧ Driving speed (time series data) mph Environment

‧ Time of day Day time (including dawn and dusk), Night time with light

‧ Weather Clear, Poor (cloud, rainy, mist, snow)

Eye-glance data

‧ Focal point (time series data) Forward, Left forward, Right forward, Rearview mirror, Left window, Left mirror, Right window, Right mirror, In-vehicle distractions (Instrument Clutter, Center stack, Interior Object, Passenger, Cell Phone)

‧ Duration of glancing at forward and other focal points

Continuous variable

The drivers’ glance locations, including Forward (F), Left Forward (LF), Right Forward (RF), Left Window (LW), Left Mirror (LM), Right Window (RW), Right Mirror (RM), Rear-view Mirror (ReM), and In-vehicle Distractions (InvD), were recorded every 0.1 s. The period of continual glances to the same focal point is considered the glance duration. Among these focal points, InvD refers to all glances inside the vehicles, including the center stack, interior objects, cell phone, passengers, and instrument cluster. Each focal point received varying degrees of attention from different drivers. To simplify the analysis, this study first analyzed only the areas where drivers glance, i.e., the InvD. Detailed characteristic differences among multiple locations or objects inside the vehicles were not considered. Moreover, this study excluded the first and final glances of each event in the glance data since these two glances may not be complete ones. Any events with a glanced area recorded as

“eyes closed” or “no video” were also excluded.

4.3 Attention Allocation Analysis from a Renewal Cycle Approach