Adaptive physics-informed trajectory reconstruction exploiting driver behavior and car dynamics.

As more and more trajectory data become available, their analysis creates unprecedented opportunities for traffic flow investigations. However, observed physical quantities like speed or acceleration are often measured having unrealistic values. Furthermore, observation devices have different hardware and software specifications leading to heterogeneity in noise levels and limiting the efficiency of trajectory reconstruction methods. Typical strategies prune, smooth, or locally modify vehicle trajectories to infer physically plausible quantities. The filtering strength is usually heuristic. Once the physical quantities reach plausible values, additional improvement is impossible without ground truth data. This paper proposes an adaptive physics-informed trajectory reconstruction framework that iteratively detects the optimal filtering magnitude, minimizing local acceleration variance under stable conditions and ensuring compatibility with feasible vehicle acceleration dynamics and common driver behavior characteristics. Assessment is performed using both synthetic and real-world data. Results show a significant reduction in the speed error and invariability of the framework to different data acquisition devices. The last contribution enables the objective comparison between drivers with different sensing equipment.