Interaction-Field Residual Learning for Calibrated Lane-Change Intent under Mixed Traffic

Abstract

The prediction of lane-change intent in mixed traffic environments, characterized by the coexistence of human-driven vehicles and autonomous vehicles, presents a formidable challenge due to the stochastic nature of human driving behaviors and complex vehicular interactions. Conventional trajectory prediction and intent recognition models often struggle to capture the nuanced, continuous spatial-temporal dynamics of such environments, frequently resulting in overconfident or poorly calibrated predictions. To address these critical shortcomings, this paper proposes a novel framework termed Interaction-Field Residual Learning for Calibrated Lane-Change Intent. The framework introduces a continuous interaction-field representation that models traffic participants as sources of physical potentials, thereby capturing both repulsive and attractive social forces in a unified topographical space. Built upon this representation, a residual learning architecture is deployed to separate the deterministic vehicle dynamics from the highly stochastic, socially driven residual movements. Furthermore, a post-hoc calibration module is integrated to ensure that the probabilistic outputs accurately reflect the true likelihood of lane-change maneuvers, reducing critical false-positive rates in safety-critical scenarios. Extensive evaluations on large-scale naturalistic driving datasets demonstrate that the proposed method significantly outperforms existing baseline models in both predictive accuracy and uncertainty calibration metrics. The findings offer profound implications for the design of safe, socially aware navigation systems in transitional mixed-traffic periods.