Step detection in complex walking environments based on continuous wavelet transform

The existing algorithms for step detection have rarely been designed for walking in complex scenes. Complex scenes often bring about more complicated changes in walking states that could cause ordinary detection algorithms less effective. In this paper, an observation is made that there are gatherings of low-frequency signals, called clusters, in the spectrogram generated from the walking signal in a particular situation through wavelet transform. The clusters would exhibit prominent features when some specific basis function is chosen for the wavelet transform, which can precisely characterize the strides in walking. Then, the criteria for choosing the basis function of wavelet transform are established and verified experimentally. Based on the spectral features, an efficient and accurate step detection algorithm named frequency domain extension detection (FDED) is proposed and its time/space complexity will be no more than the constant times of its input size, O(n). FDED consists of three phases. First, the Kalman filter is adopted to denoise the raw data. Then, a continuous wavelet transform is applied to the filtered data to attain the obvious gait pattern in the time spectrum. Finally, a robust detection algorithm is proposed to implement step counting and single-stride segmentation. The experiments are conducted on two datasets, Diecui, a self-established dataset with diverse walking patterns in complex scenes, and a public dataset, ZJU-gaitacc. The experimental results show that FDED achieves an average accuracy of 99.1% for step counting on Diecui, and outperforms several representative detection algorithms on ZJU-gaitacc, which suggests that the proposed algorithm possesses strong adaptability in complex scenes with diverse personnel.