A physics-informed uncertainty quantification framework for deep learning-based real-time lane-change intention prediction

Lane-change intention prediction represents a significant challenge in autonomous driving, particularly within mixed traffic environments where autonomous vehicles (AV) must interact and share road space with human-driven vehicles. While deep learning has advanced lane-change intention prediction, the absence of real-time uncertainty quantification limits operational robustness. We propose a physics-informed uncertainty quantification framework that integrates seamlessly with prediction models, enabling reliable and interpretable lane-change assessment for autonomous vehicles. We present a framework that advances lane-change intention prediction by overcoming two long-standing challenges: bias from class imbalance in naturalistic data and dependence on precise maneuver labels. It does so through empirical probability distribution maps and a Scenario-Aware Dynamic Distribution Integration (SADI) method that adaptively captures temporal driving patterns while embedding vehicle physics. This physics-informed approach aligns uncertainty estimates with plausible behavior, yielding reliable and interpretable predictions. Tests on the highD dataset show that, unlike traditional uncertainty quantification methods such as dropout or ensemble methods, our framework delivers real-time reliability scores that remain robust across diverse and evolving traffic scenarios. To validate generalizability across diverse traffic conditions, we further evaluate the framework on the AD4CHE dataset, demonstrating zero-shot transferability without hyperparameter retuning. By embedding vehicle physics into uncertainty estimation, our framework transforms lane-change prediction from brittle heuristics into reliable, real-time assessments that adapt to complex traffic dynamics.