Robustness of vehicle identification via trajectory dynamics to noisy measurements and malicious attacks

The possibility of using infrastructure-based sensors to identify individual vehicles, and then actuate transportation control infrastructure in response to their individual dynamics will enable to next generation of traffic infrastructure control. However, any such system to identify individual vehicles in the flow will be prone to faulty data or worse, cyberattacks where a malicious actor intentionally injects faulty data. With this context in mind, we investigate the resilience of a recently proposed deep learning based approach to identify individual vehicles in the traffic flow. We conduct numerical experiments where increasing amounts of noise is injected into time series trajectory data and conclude that while the proposed classification method is accurate at identifying individual vehicles when there is no noise, classification accuracy deteriorates quickly when noise is injected.

I. Introduction

Large-scale vehicle trajectory extraction demonstrations, such as the recently constructed I-24 MOTION testbed in Nashville, TN [1] have shown that ubiquitous trajectory extraction at the corridor level may soon be possible. System-atic trajectory extraction will enable the next generation of traffic control, where the dynamics of individual constituents of the flow are taken into account in real-time to optimize control [2]. One example of such control is the actuation of flow-level traffic control devices (i.e., variable speed limit or ramp metering control) based on the properties of the individual vehicles in the flow. This could be based on the proportion or characteristics of automated or partially automated vehicles with driver assist capabilities such as adaptive cruise control (ACC), or based on the driving dynamics of individual drivers to consider characteristics such as aggressively.