Walk detection (WD) and step counting (SC) have become popular applications in the recent emergence of wearable devices. These devices monitor user states and process data from MEMS-based accelerometers and optional gyroscope sensors. Various algorithms have been proposed for WD and SC, which are generally sensitive to the contexts of applications, i.e., (1) the locations of sensor placement; (2) the sensor orientations; (3) the user's walking patterns; (4) the preprocessing window sizes; and (5) the sensor sampling rates. A thorough understanding of how these dynamic factors affect the algorithms' performances is investigated and compared in this paper. In particular, representative WD and SC algorithms are introduced according to their design methodologies. A series of experiments is designed in consideration of different application contexts to form an experimental dataset. Different algorithms are then implemented and evaluated on the dataset. The evaluation results provide a quantitative performance comparison indicating the advantages and weaknesses of different algorithms under different application scenarios, giving valuable guidance for algorithm selection in practical applications.
Keywords: gait analysis; machine learning; signal processing; step counting; walk detection.