Characterization of Lane Keeping and Departure Scenarios Using Large-Scale Naturalistic Driving Data

<div>Roadway departures remain a major cause of crashes, injuries, and fatalities on U.S. roads. Technologies such as lane keeping assist (LKA) and lane centering assist (LCA) can help mitigate these crashes, but their development involves extensive characterization of the parameter space in which they operate. Lane and road departures (LDs/RDs) and lane changes (LCs) must be systematically described and quantified to distinguish kinematic features, identify contributing factors, and benchmark system influence on lateral control.</div> <div>This study developed a unified pipeline to mine over 36 million miles of naturalistic driving study (NDS) data collected from more than 3800 participants. The pipeline integrates various types of signals to detect roadway boundary crossings, classify LKA-relevant scenarios, and extract roadway, driver, environmental, and assistance-related parameters. Lane keeping epochs with and without LKA were also extracted to quantify system influence on lateral control.</div> <div>In the NDS analysis, crashes include both object contact events and RDs, defined as non-premeditated departures from the intended travel surface involving at least one tire. Analysis of pre-identified crashes in the NDS showed that unintentional RDs accounted for 5.67%, unintentional LDs for 1.76%, and intentional LCs for 1.55%, corresponding to lower-bound rates of 2.7, 0.8, and 0.7 crashes per million vehicle miles traveled. RD crashes were predominantly right-sided, LD crashes left-sided, and both were overrepresented on curves and under adverse conditions. Loss of control preceded 22% of RD crashes and 69% of LD crashes.</div> <div>Beyond crashes and near-crashes (CNCs), the algorithm identified approximately 3 million LCs and 0.3 million LDs/RDs. LCs typically involved larger crossing angles that decreased with speed, while departures clustered within 0°–2°. Compared with CNCs, these occurred at higher speeds and smaller angles. LKA consistently reduced lateral variability without biasing the mean offset.</div>