Traffic Outlier Detection by Density-Based Bounded Local Outlier Factors

Abstract

Outlier detection (OD) is widely used in many fields, such as finance, information and medicine, in cleaning up datasets and keeping the useful information. In a traffic system, it alerts the transport department and drivers with abnormal traffic situations such as congestion and traffic accident. This paper presents a density-based bounded LOF (BLOF) method for large-scale traffic video data in Hong Kong. A dimension reduction by principal component analysis (PCA) was accomplished on the spatial-temporal traffic signals. Previously, a density-based local outlier factor (LOF) method on a two-dimensional (2D) PCA-proceeded spatial plane was performed. In this paper, a three-dimensional (3D) PCA-proceeded spatial space for the classical density-based OD is firstly compared with the results from the 2D counterpart. In our experiments, the classical density-based LOF OD has been applied to the 3D PCA-proceeded data domain, which is new in literature, and compared to the previous 2D domain. The average DSRs has increased about 2% in the PM sessions: 91% (2D) and 93% (3D). Also, comparing the classical density-based LOF and the new BLOF OD methods, the average DSRs in the supervised approach has increased from 94% (LOF) to 96% (BLOF) for the AM sessions and from 93% (LOF) to 95% (BLOF) for the PM sessions.