Koopman operator approach for instability detection and mitigation in signalized traffic
E. Ling, L. Ratliff, S. Coogan
IEEE International Conference on Intelligent Transportation Systems, 2018
Abstract
The objective of this paper is to demonstrate an application of the Koopman Operator approach for automated detection of unstable behavior in traffic dynamics. We propose an algorithm that searches for long sequences of unstable eigenvalues in the learned dynamics as an anomaly feature, and demonstrate feasibility of the approach on a case study day with abnormally high queue lengths due to an accident. Secondly, we present a method for modeling traffic dynamics with signal phases included as exogenous input. Given an unusually congested condition, the analysis is consistent in that longer green times for the affected leg result in less severe congestion and vice versa for shorter green times. This shows promise for anticipating queue behavior under modified traffic signal phase sequences. We encapsulate these ideas within the larger picture of building improved adaptive traffic control systems.