*3.1. Feature Extraction Techniques*

Feature extraction from the accelerator data stream is crucial for walk detection and step counting. Various features have been presented in the literature. Statistical features in the time domain, including mean,

variance, correlation, skewness, kurtosis, energy, etc., were proposed in [29–34]. Other features such as peak interval and zero/mean-crossing rate were also proposed in [9,35]. Additionally, root mean square (RMS) and histogram were proposed in [36].

Features in frequency and transformed domains were also proposed in many works. FFT bins were used in [28,37], and wavelet coefficients were introduced in [38,39]. The peak frequency and power ratio of different frequency bands were exploited in [37]. Mel-frequency cepstral coefficients (MFCCs) and Bark-frequency cepstral coefficients (BFCCs) as complex features of frequency domains were also proposed in [36].

Besides these conventional features, principal component analysis (PCA) was proposed in [40], although it is commonly used as a feature selection method. Autoencoder networks [40] and sparse coding [41,42] have also been introduced recently. Furthermore, some manually designed features such as weightlessness features were used in [43].

We split the common features of WD and SC into three groups in Table 2.

**Table 2.** Feature categorization of walk detection (WD) and step counting (SC). BFCC, Bark-frequency cepstral coefficients.

