A Comparative Review and Benchmarking Evaluation of Step Detection Methods for Longitudinal Gait Assessment

Abstract

Gait analysis based on inertial measurement units (IMUs) enables scalable longitudinal gait assessment, with step detection as a foundational component. However, an understanding of available step detection methods and their comparative performance across sensor placements, clinical populations, and real-world environments remains limited. This paper addresses this gap through a twofold contribution. First, a structured taxonomic review organises the landscape into six families—rule-based heuristics, phase portrait, template matching, time–frequency analysis, adaptive oscillators, and machine learning—characterising the trade-offs of each in terms of noise robustness, generalisability, and computational efficiency. Second, nine algorithms drawn from five of these families, all but phase portrait, are benchmarked across three large datasets encompassing healthy adults and people with mobility-impairing conditions such as Parkinson’s disease and multiple sclerosis, at multiple sensor placements under both controlled laboratory and free-living conditions. Accuracy depended on sensor placement, gait pathology and recording environment: step detection proved reliable in structured laboratory settings across most sensor locations and clinical cohorts, whereas unsupervised free-living monitoring produced a substantial degradation for most algorithms. Over all conditions, the three evaluated time–frequency methods obtained the best average ranks for detection accuracy, step-duration error and cadence error.