Rhythm tracking is of great significance in the fields of health medicine and sports science. With the development of exoskeletons and wearable sensing devices, the demand for rhythm tracking of complex and multiple human movement patterns has become even greater. However, current research finds it difficult to achieve a balance among the generalization, real-time performance and accuracy of rhythm tracking. In response to the above challenges, this paper designs a rhythm tracking framework based on wearable sensing devices. Through experimental tests on five subjects, it was verified that the method proposed in this paper guarantees the generalization, real-time performance and accuracy of rhythm tracking to a certain extent. The specific analysis results revealed that the proposed method decreased by 56.4%–95.6% in the RMSE of phase estimation compared with the traditional uniaxial signal extraction method.