Multi-stage temporal inference and adaptive windowing for automatic lane-change onset labeling from trajectory data

Abstract

Accurate identification of lane-change onset from vehicle trajectories is essential for driving behavior analysis and trajectory-based modeling. Existing signal-based studies commonly localize onset using fixed retrospective windows on pre-selected single-event samples, leaving limited evidence on how window specification affects localization performance in full trajectories with heterogeneous maneuver durations, traffic states, and multiple lane-change events. This study proposes a decision zero-crossing (DZC)-anchored adaptive event-level window construction framework for retrospective onset localization from completed trajectories. For each event, target-direction zero-crossing candidates are extracted from historical lateral motion, and a Hidden Markov Model infers the DZC as an internal temporal anchor. An event-specific analysis window is then constructed around this anchor, within which onset localization uses a common wavelet-based localization rule. The framework is evaluated using UAV-collected freeway work-zone trajectories and examined on highD and CQSkyEyeX. Results show that onset localization is sensitive to temporal window construction. Under the same localization rule, fixed-window and Global strategies exhibit unstable errors across traffic states and trajectory structures, whereas the DZC-based strategy yields more concentrated error distributions and smaller upper-tail errors. External validation shows that this advantage extends to freeway basic-section datasets. Comparisons with alternative baselines indicate that adaptive windowing benefits depend on robust anchor selection. A downstream sensitivity experiment shows that onset-based sample construction affects lane-change prediction performance and that prediction metrics should be interpreted together with onset localization accuracy. These findings support adaptive event-level temporal support construction aligned with lane-change stage structure as a reliable basis for retrospective onset localization in freeway trajectory data.