Comparison of vehicle re-identification models for trucks based on axle spacing measurements

Abstract

ABSTRACTIn previous research, it has been demonstrated that there is enough variation within the truck population in terms of axle spacings and vehicle lengths, which enable anonymous vehicle re-identification between two measurement stations (e.g., two weigh-in-motion (WIM) sites). Matching trucks between two sites can support various applications, such as calibration of WIM equipment and estimation of travel times and origin-destination flows. In this paper, several modeling approaches to solve the re-identification problem are explored including Naïve Bayes (NB), Bayesian Models (BM) fitted by mixture models, and the formulation of the re-identification problem as a mathematical assignment problem. In addition, the influence of selecting a similarity measure is evaluated through numerical experiments conducted on real-world data from six pairs of upstream–downstream WIM stations. The results demonstrate that solving the re-identification problem with BMs fit by mixture distributions outperforms solving with NB models, while both are outperformed by the mathematical assignment formulation of the same problem, especially when vehicle-pairs exceeding a high threshold of similarity are matched. In addition, expressing the similarity between measurements from two stations as a percentage difference is found to be relatively more advantageous. For the presented pairs of WIM stations, up to 90% matching accuracy can be achieved when the best combination of re-identification method and similarity measure are implemented, and only those vehicle-pairs exceeding a high threshold of similarity are matched.AbbreviationsAAAssignment AlgorithmBMBayesian ModelDDownstreamGMMGaussian Mixture ModelNBNaïve Bayes MethodpdfProbability Density FunctionUUpstreamVRIVehicle Re-Identification