Naturalistic Driving Study (NDS) data are an important source for driver behavior data to design and evaluate autonomous vehicles and driver assistance systems. The number of serious crash events in NDS, however, is often small. As a result, surrogates such as “near crashes” or events identified using vehicle instrumentation are used with the assumption that they are relevant to real crash events. The objective of this study is to determine if NDS crash and near-crash data are indeed representative of crash events. To examine this issue, we focused on one subset of crash events, lane departure events where the vehicle drifts out of its lane. These are the events most likely to be mitigated by lane departure warning systems.

Four naturalistic datasets that covered the full range of events from lane departures during normal driving, to nearcrashes, to crashes were compared to data from a crash database. Our hypothesis is that the crash and near-crash NDS events will have the most similar vehicle kinematics compared to the crash database. Normal driving departure events were extracted from the Integrated Vehicle-Based Safety Systems (IVBSS) field operation test. Two departure event datasets from IVBSS were identified using the lane tracking cameras. The first dataset consisted of 12,760 cases of the vehicle departing and returning to its lane and the second consisted of 7,750 events where the equipped LDW systems were triggered. Thirty-two (32) near-crash lane departure events were analyzed from the 100-Car NDS. Finally, 49 curb strike events were analyzed from the SHRP-2 NDS. Data from lane departure crashes was extracted from the National Automotive Sampling System, Crashworthiness Data System (NASS/CDS). Event Data Recorders (EDRs) downloaded from 482 NASS/CDS crash investigations were analyzed.

There were important sampling differences between datasets. Younger drivers were overrepresented in the 100-Car near-crash and SHRP-2 curb strike events and crash data while the IVBSS participants were uniformly distributed over age and gender groups. The vehicle speeds from IVBSS were restricted to over 42 kph (25 mph), whereas the crash data had vehicles speed that contained both low and high speed events. The 100-Car near-crash and SHRP-2 curb strike departures had larger departure angles (2.6° and 14.1° median, respectively) and lateral excursion (0.63 m and 0.50 m median, respectively) compared to the IVBSS data (0.6° and 0.7° departure angle and 0.19 m and 0.10 m excursion for LDW and lane departure datasets, respectively). The differences in departure conditions may have also affected driver maneuvers after the departure. In 52% of crashes with EDRs there was a brake application in the last second before the crash compared with 38% of 100-Car near-crash and 33% of SHRP-2 curb strike events. The selection criteria for the IVBSS departures excluded almost all brake application, with only 4% of the IVBSS LDW events having brake application. Steering wheel input was also substantially larger in the 100-Car near-crashes (48°) compared to the IVBSS (4°-5°).

These results show that crash and near-crash events from NDS produce datasets that are most consistent with crash data compared to datasets generated using lane tracking information. If the research question involves replicating conditions relevant to departure crashes, such as in the design of test track experiments, crash and near-crash events should be used over less severe NDS departure events.