Outlier Detection in Traffic Data Based on the Dirichlet Process Mixture Model

Traffic data collections are exceedingly useful for road network management. Such collections are typically massive and are full of errors, noise and abnormal traffic behaviour. These abnormalities are regarded as outliers because they are inconsistent with the rest of the data. Hence, the problem of outlier detection (OD) is non-trivial. This paper presents a novel method for detecting outliers in large-scale traffic data by modelling the information as a Dirichlet process mixture model (DPMM). In essence, input traffic signals are truncated and mapped to a covariance signal descriptor, and the vector dimension is then further reduced by principal component analysis. This modified signal vector is then modelled by a DPMM. Traffic signals generally share heavy spatial-temporal similarities within signals or among various categories of traffic signals, and previous OD methods have proved incapable of properly discerning these similarities or differences. The contribution of this study is to represent real-world traffic data by a robust DPMM-based method and to perform an unsupervised OD to achieve a detection rate of 96.67% in a ten-fold cross validation.