Automatic Imputation of Missing Highway Traffic Volume Data

Abstract

Automatic data imputation is often needed in cyber-physical systems to enhance the quality of incomplete datasets produced by sensors. Existing methods require significant modeling and analysis efforts at each sensor location which hinders their applicability in large-scale systems. This paper presents an automatic approach to selecting and estimating autoregressive integrated moving average models for traffic volume imputation. We study real-life data collected from around 1030 sensors distributed along major highways in Toronto, Canada. We study the characteristics of missing data in order to provide measures for the quality of data collected. The proposed method estimates missing traffic volume data for any sensor from its own observed values. Results show that the proposed procedure can estimate short and mid-sized gaps (less than one week) with an accuracy that ranges between 7 and 25%.