Firstly, to solve the problem of threshold setting for step-counting in personal dead reckoning system, this paper proposes a self-learning of threshold based on FSM. It can obtain an adaptive threshold in a short period of time(about 10s), and improve the accuracy of pedometer algorithm. Secondly, the FSM splits the continuous gait into individuals, which are processed by PCA to obtain the step direction. In addition, the algorithm effectively solves the 180° ambiguity by analyzing the variation of forward acceleration. The results of the experiment show that the dynamic-window PCA decreases the amount of data to be processed by 61.2% and increases the accuracy by 11.1% compared to the fixed-window PCA.