Abstract: To solve the difficulties of feature extraction for acceleration signals, due to the limited number of sampling points and many outlier patterns of gait signals, which can't be used to divide the gait cycle, we propose a novel outlier detection algorithm based on K-mean distance outlier factor detection. We use a wavelet de-noising method to adaptively pretreat gait signals, and extract the edge point of the gait series according to the edge weight. This achieves a four-dimensional feature space and feature subspace by the use of the four eigenvalues of each sub-pattern. We then compute the K-mean distance outlier factor. Lastly, we screen the gait series with the mean outlier value, and test it using the unit of the gait cycle. The promising experimental results demonstrate that the proposed algorithm is accurate and reasonable on open access datasets as well as our built datasets, and efficiently detects outliers in gait series.