This paper presents a novel algorithm to extract the drum track of a polyphonic music signal, based on a harmonic / noise decomposition. This algorithm is causal and does not require prior knowledge or learning. The input signal is split into several frequency bands in which the signal is separated in a deterministic and a stochastic part. The stochastic part can be efficiently used to detect drum events and to resynthesize a drum track. Possible applications include drum transcription, remixing, and independent processing of the rhythmic and melodic components of music signals. Results obtained from real recordings of popular music are presented, as well as a perceptual evaluation of the quality of remixed signals.