We present a method to determine the polarization of body waves from three-component, high-frequency data and examples of its application. The method is based on the principal component approach. One advantage of this approach is that the polarization state can be determined for small time windows compared with the predominant period of the wave. This is particularly useful for identifying converted waves within the crust. The stability of the result is analyzed with synthetic cases by adding simultaneous arrivals from waves and random noise. The method works well with both synthetic and local data in the detection of the polarization of the wave by separating arrivals from different directions. From the local data, some seismic phases related to crustal conversions are observed that require strong lateral variations.