Localization and perception for autonomous vehicles are often studied separately. However, they can be regroup on a dynamic map representing the environment of the vehicle. This dynamic map can be exchanged with other vehicles to be fused with their own dynamic maps to increase their accuracy. This paper presents a decentralized data fusion method for cooperative localization based on both Extended Kalman Filter and Covariance Intersection Filter. Different observation models of the relative measurements from the perception (Cartesian and polar relative poses, distances, bearings and relative yaws) are compared. The approach is tested on data for 10 vehicles generated from a real dataset and completed with a simulated perception.