In visual simultaneous localization and mapping (SLAM) field, especially for feature based stereo-SLAM, data association is one of the most important and time-consuming sub-tasks. In this paper, we investigate the roles of different measured features during the data association process and present a new hybrid feature parametrization approach for stereo SLAM, which only selects a subset of the matched features that contributes most and treats nearby and distant features separately with different parametrization. We formulate a pipeline to filter, store and track the features which saves time for further state estimation. For different types of features on manifold and Euclidean space we apply corresponding designed maximum likelihood estimator with quadratic constraints and thus get a near-optimal estimation. Experimental results on EuRoC dataset and real tests show that our proposed algorithm leads to accurate state estimation with big progress in consistency.