Traffic Flow Crystallization Method for Trajectory Approximation and Lane Change Inference

Abstract

Whereas on many motorways, traffic operations are permanently monitored, and long historical logs of such data exist, they are not directly usable for lane change studies, as they only register local passages and speeds. This study proposes a novel method to transform discrete vehicle passage records of individual vehicle data (IVD) into approximations of vehicle trajectories and inference of lane change maneuvers (LCMs), such that large-scale LCM dataset can be retrieved from existing infrastructures where IVD is recorded at sufficiently close spacings (~600 meters). The method’s core is a probabilistic re-identification of individual vehicles in successive, lane-specific loop detectors. Dubbed Traffic Flow Crystallization (TFC), the methodology enhances traffic monitoring by providing vast and diverse LCM datasets. It consists of two key re-identification (ReID) modules: a lane-restricted module that matches vehicles strictly within the same lane and a non-lane-restricted module that recursively identifies lane-changing vehicles using boundary conditions imposed by previously matched vehicles. This recursive process resembles crystal growth, inspiring the method’s name. The ReID methodology is based on a weighted likelihood function consisting of Bayesian probability estimators that integrate three similarity measures: vehicle length, passage time, and passage speed. A lane-change feasibility filter ensures that re-identified vehicles satisfy plausible spatiotemporal constraints. The final module resolves inconsistencies and infers LCMs. The proposed method is trained and validated using CCTV footage, where visually-identified vehicles serve as ground truth. Validation results demonstrate a vehicle ReID success rate exceeding 96% and an inferred LCM rate with only a 2% underestimation compared to ground truth.