Application of Principal Component Analysis for Outlier Detection in Heterogeneous Traffic Data

Level-of-service (LOS) measures of two-lane highways exhibit incompatibility if the prevailing traffic is heterogeneous in character. Thus, such traffic warrants development of LOS criteria on the basis of compatible measures which capture its characteristics. The present paper has suggested the use of percent speed-reduction and percent slower vehicles, as the measures of performance, while defining LOS criteria. Defining such criteria is basically a classification problem and clustering could be applied as an effective technique for its solution. However, heterogeneity in the traffic mix results in the presence of significant proportion of outliers in the data set, which can distort the results and render into misleading or useless outcomes. The study considers principal component analysis to be an efficient technique in detecting outliers from the data set and accordingly applies it on the proposed LOS measures. An iterative process, adopted for removing outliers, indicates that significant proportion of outliers comprises of non-motorized traffic data; this accordingly ensures reliability of the data set. The study concluded the unfeasibility of LOS assessment of the entire traffic, considering both motorized and non-motorized modes, with respect to a common scale.