An Invariant May Drive the Decision to Encroach at Unsignalized Intersections

Abstract

Summary: This paper introduces a novel approach to understanding when and where drivers make the Go / No Go decision (not) to turn left and encroach upon an approaching car that has the right-of-way in an unsignalized intersection. The source of data is approximately 2,400 hours of video recordings at two intersections near Göteborg, Sweden. Automated image processing software extracted the trajectories of the pairs of cars involved in more than 14,000 left turns across traffic at the first intersection and 2,400 at the second. We subdivided the data into four different left-turn scenarios -where the approaching car arrives from the opposite direction, from the lateral direction, from the intended direction (merging), and while making its own left turn. For each scenario, we found the distances between the turning car and the approaching car at the time when we can assume the decision (not) to turn is made and conducted logistic regressions to identify the distances associated with the 50/50 acceptance probabilities for the decision (not) to turn. We also calculated the resulting encroachment distances ('trailing buffers') for every decision to turn. We expected to find wide variability in these buffers. Instead, we observed separations that were virtually the same across scenarios at each intersection but differed across intersections. Tacit, intersection-dependent knowledge of this invariant may drive the decision of whether or not to turn and encroach. We discuss the implications this finding has for the design of in-vehicle active safety systems.

Probabilistic Risk Metrics for Navigating Occluded Intersections

“…During its maneuver, it must move from an origin lane to a new lane, passing through a junction, while avoiding incidents with other vehicles. We define an incident to include collisions, near-miss braking incidents, traffic conflicts [27], small gap spacing [15], or the threat of accident from other vehicles [13]. The risk to the ego-car at time t, denoted r t , is computed as the expected number of incidents that will occur if the ego-car enters the intersection at time t.…”

“…Successful ADAS deployment requires accurate assessment of the intersection topology and turning prediction. Increased fidelity in lane-level maps can improve turning predictions [14], and studies on human drivers have identified typical distances for turning decisions [15]. Predicting when to warn the driver serves as a key feature of ADAS systems.…”