In application fields such as data mining or uncertainty quantification very large, high dimensional datasets are rather common. While there are several different approaches to kernel density estimation, they all of them share the disadvantage of being very time consuming, when the sample set is really large. In this thesis, we want to overcome this issue by sub-sampling the dataset and choosing an adaptive kernel density estimator, which is as easy and time-saving, while still being as accurate as possible. Apart from being directly compared to another proved efficient density estimation technique, the Rosenblatt transformation is adducted to serve as an application example.