Abstract: Aiming at the problems of over counting and wrong counting in step frequency detection, which affect the indoor positioning accuracy of pedestrian dead reckoning (PDR), an adaptive step frequency detection algorithm is proposed. Because there is a large amount of interference noise in the data directly collected by the built-in acceleration sensor of smart phone, a combined filtering denoising method is proposed.The acceleration data is denoised by preprocessing filter combination of exponential hull moving average, Kalman filter (KF) and low-pass filter. Then, in different scenes, such as upstairs and downstairs, horizontal ground and unlimited walking speed, the number of peak-valley values is obtained after the peak-valley value de differentiation, adaptive dynamic threshold and peak-valley value pairing detection algorithm, so as to achieve accurate step counting in multi scenes and multi gait. The experimental results show that, compared with the peak detection method and dynamic threshold algorithm, the proposed method can effectively eliminate the false steps and adapt to the upstairs and downstairs scenes, and the average accuracy of the experiment in the comprehensive scene reaches 99.44%.