Citation

Zaki MH, Sayed T. Transp. Res. C Emerg. Technol. 2013; 33: 50-73.

Copyright

(Copyright © 2013, Elsevier Publishing)

DOI

10.1016/j.trc.2013.04.007

PMID

unavailable

Abstract

New urban planning concepts are being redefined to emphasize active modes of transportation such as walking and cycling. In this environment, a disparate mix of road users shares the same road. In order to address road-safety and level of service requirements for the different types of road-users, traffic information needs to be collected with high degree accuracy. The reliability and accuracy of the data collected can significantly affect the quality of analysis. In recent years, automated video analysis was established as a robust tool for data collection. Road-user trajectories obtained through automated computer vision are rich in information. They hold features that reveal the structure of the traffic scene and provide clues to the movement characteristics of the road-users. The objective of this paper is to present and evaluate a road-user classification procedure. The classification relies on the motion pattern attributes associated with the trajectories of each road-user type, namely, vehicles, pedestrians and cyclists. A novel approach for features selection is proposed where singular spectrum analysis (SSA) identifies the basic harmonics (speed variation patterns) characterizing the movement behaviors. Constrained based classification and spectral clustering are then applied on the selected features to categorize the road-users. Several case studies are used for the evaluation of the proposed classification. The case studies use real world data sets collected at a roundabout and a conventional four-legged intersection in Greater Vancouver, British Columbia. Very promising results were demonstrated and evaluated through several performance measures. The main benefit of this research is to apply classification as a first step in the activity and behavior recognition of road-users in traffic scenes. The proposed approach is useful as a complement to vision based classification, where road-users classes can be difficult due to partial occlusion or when the extraction of physical characteristics (descriptive features) of road-users does not provide enough details to reveal the correct type of a road-user.