Deep Learning-Based Risk Classification for Gap Acceptance Decisions at Unsignalized Intersections Toward Proactive V2X Safety System

Abstract

Unsignalized intersections account for a disproportionate share of traffic fatalities, with gap acceptance misjudgments as a primary crash mechanism. Existing collision warning systems react after conflict has materialized, relying on threshold-based metrics such as time-to-collision that offer little predictive capacity before a crossing is committed. This study presents a machine learning pipeline to classify accepted gap acceptance events as safe or safety-critical using pre-decision trajectory data alone, supporting proactive V2X warning system design. Crossing events were extracted from two geometrically distinct scenarios in the INTERACTION naturalistic driving dataset -- a T-junction (EP0) and an all-way stop (MA) -- and labeled using Post-Encroachment Time (PET <= 1.0 s) as the risk criterion, yielding 1,023 events. A 30-frame pre-entry observation window capturing ego kinematics, conflict vehicle kinematics, and pairwise interaction features served as model input. Five classifiers were trained on MA and evaluated on EP0 to assess cross-scenario generalization. The LSTM achieved the best overall performance (AUC = 0.647, SC recall = 0.862), with conflict vehicle geometry identified as the dominant risk predictor and temporal importance rising toward the entry point. A pronounced domain shift between scenarios established that intersection geometry critically shapes the risk signal, motivating scenario-aware training strategies for real-world deployment.