Gait is a biological feature of human being, which is important in the research of location and navigation. Most approaches to the pedestrian gait classification based on MEMS inertial sensors use a peak-picking method. These approaches recognize the current pedestrian gait by detecting the peak value of acceleration signals. False acceleration peaks due to self-noise caused by Brownian motion and environmental factors reduce accuracy of classification results. To deal with this problem, an approach named gait detailed classification of pedestrian level walking based on kurtosis is proposed from the perspective of the overall waveform. The gait acceleration signals in the forward direction obtained from the MEMS sensor is first transformed from the time domain to the frequency domain with FFT. Modulus of the frequency domain signals are squared, and then transformed back to the time domain. A self-amplified signal can be obtained in this process, and most false acceleration peaks removed. Finally, kurtosis of the self-amplified signal is calculated and analyzed to distinguish jog, walk and run. Experimental results show that the average recognition accuracy of the proposed method reaches 98.62%, improving the overall classification accuracy by 7.37% as compared with methods of combining acceleration and frequency power.