Traffic congestion monitoring using an improved kNN strategy

Abstract

A systematic approach for monitoring road traffic congestion is developed to improve safety and traffic management. To achieve this purpose, an improved observer merging the benefits of a piecewise switched linear traffic (PWSL) modeling approach and Kalman filter (KF) is introduced. The PWSL-KF observer is utilized as a virtual sensor to emulate the traffic evolution in free-flow mode. In the proposed approach, residuals from the PWSL-KF model are used as the input to k-nearest neighbors (kNN) schemes for congestion detection. Here, kNN-based Shewhart and exponential smoothing schemes are designed for discovering the traffic congestions. The proposed detectors merge the desirable properties of kNN to appropriately separating normal from abnormal features and the capability of the monitoring schemes to better identify traffic congestions. In addition, kernel density estimation has been utilized to set nonparametric control limits of the proposed detectors and compared them with their parametric counterparts. Tests on traffic measurements from the four-lane State Route 60 in California freeways show the effectiveness of the PWSL-KF-based kNN methods in supervising traffic congestions.