Cooperative positioning based on Impulse Radio - Ultra WideBand (IR-UWB) is known to provide a centimeter-level accuracy when the positions of anchor nodes are perfectly known. In Vehicular Ad-Hoc Networks (VANETs), vehicles acting as 'virtual anchor' nodes are highly mobile with imperfect position estimates delivered by the Global Navigation Satellite System (GNSS). The large difference between measurement noises of GNSS positioning and IR-UWB Vehicle-to-Vehicle (V2V) ranging creates a bias in the localization filter, which is cooperatively propagated to other to other vehicles, and therefore significantly attenuates the benefits of IR-UWB for Cooperative Localization (CLoc). This paper compensates this drawback by a novel 2-step CLoc fusion framework. It first selects the 'virtual anchor' nodes with the lowest GNSS position uncertainty to mitigate and stop the propagation of the biases. Once all biases have been reduced, it refines localization accuracy through exhaustive fusion. This strategy increases the probability to reach a 40 cm accuracy from 25% (conventional IR-UWB) to 95%, and even a 20 cm accuracy from 5% to 40%. © 2016 IEEE.