A fetal movement detection method with multi-features based on multi-acceleration sensor is proposed aiming at monitoring the fetal health indicators. A FIR band-pass filter and wavelet denoising are used to preprocess the signal; smoothed Wigner-Ville distribution（ WVD） is used to analyze and extract features of time domain on fetal movements as well as the distribution in the time-frequency（ TF） domain. The number of the areas in the TF domain is imported as the most important recognition feature

extracted by Gabor dictionary and Matching pursuit（ MP）. A model predicting fetal movements is built up with these features. The simulation results show that the algorithm has high recognition rate and reliability.