The only thing I can think of off-hand is a median split.

Iman,

In principle you can use any clustering algorithm in one dimension( for example K-means or Patrick's suggestion).

The real problems that you must consider though are the following:

1) Does the single feature describe completely the phenomena that you want to capture?

2) Is the selected feature enough to discriminate the classes?

3) Is the data that you have enough to describe the underlying distribution that you are looking for?