Radar waveform classification is a difficult problem due to several different varying parameters. The classifier must handle waveform alignment, different pulse widths, and should degrade gracefully with decreasing signal to noise ratios. Along with these tasks, a crowded spectrum makes it highly unlikely that every waveform encountered will be in the waveform library. In this paper, these challenges are addressed through a combination of feature design, training protocol, and classifier approach. The classifier used in this effort is the quantile one-class SVM (q-OCSVM) that has the desirable properties of out-of-class rejection and likelihood estimation. These design choices result in a high performance waveform classifier that addresses the aforementioned challenges as demonstrated with extensive experimentation.