One of the critical challenges in a self-driving system addressed by the perception module is the tracking of actors in the vicinity of an autonomous vehicle. This is a difficult and complex task, that entails inferring higher-order states of the actors (such as velocity, acceleration, heading, and so on) as well as their tracking through time. The task is particularly challenging in the case of vehicles that have a non-rigid shape, such as articulated vehicles where two or more rigid parts are connected by hitch points. In addition, if we consider a highway case where a large number of such vehicles can be present, the task becomes even more demanding as all surrounding actors need to be tracked very efficiently in real-time. While there exist accurate vehicle kinematic models that account for articulation, applying them directly to tracking of a large number of articulated vehicles can be difficult due to the complexity and high latency cost of such models when used in the update step of the tracker. We focus on this problem and propose an approach that allows for the accurate tracking of articulated vehicles in a very efficient manner. In particular, we achieve this by tracking each rigid part independently and applying tracker updates using both actual observations as well as articulated phantom observations. These phantom observations are computed by considering the articulated kinematic model and the state of the neighboring connected part, which allows for efficient and kinematically consistent update steps. We evaluated the proposed approach on large-scale real-world data collected on highways in Texas, and compared it to a number of state-of-the-art baselines. The results strongly indicate the benefits of using phantom updates for accurate and efficient tracking of articulated vehicles.