Dynamic Nonrecurrent Congestion Event Detection and Tracking Method With DBSCAN on Speed Watersheds

Abstract

Nonrecurrent congestion (NRC) events caused by incidents bring unexpected delays and affect normal traffic operations. Imprecise NRC event detection methods can trigger false alarms and repetitive incident alerts for the same congestion event. The speed watershed from the historical profile based on DBSCAN can provide a reference for identifying NRC. This paper proposes a DBSCAN-based dynamic NRC tracking (DyNRTrac) algorithm to detect and track NRC events. By comparing real-time spatial–temporal patterns of the speed contour diagram against the historical speed contour diagram along a corridor, this method effectively distinguishes NRC events from regular traffic patterns. The proposed algorithm applies the Rauch–Tung–Striebel smoother for speed noise reduction and establishes a historical congestion profile for each recurrent congestion event within a corridor by each day of the week and season. A new event-profile–based 3D speed volume comparison method is proposed to detect NRC events that do not significantly overlap with recurrent congestions in the historical profile. Finally, a bilevel congestion confirmation process is introduced for NRC persistency checking and filtering. The proposed algorithm was evaluated by using field travel time data and with the New Jersey Department of Transportation OpenReach incident database. Overall, the proposed model shows up to 88.3% detection rate for NRC that can match the incident in the database, and it shows superior detection rates on NRC events at a similar false alarm rate level when compared with three prior models over the same datasets. Furthermore, a detailed spatial–temporal map analysis is provided to show the capability of the proposed method in distinguishing NRC and RC and identifying nonaccidental NRC events, providing its potential for traffic operation management systems to assist traffic operators to be alerted about NRC events.