Classifying running speed conditions using a single wearable sensor: Optimal segmentation and feature extraction methods

Accelerometers have been used to classify running patterns, but classification accuracy and computational load depends on signal segmentation and feature extraction. Stride-based segmentation relies on identifying gait events, a step avoided by using window-based segmentation. For each segment, discrete points can be extracted from the accelerometer signal, or advanced features can be computed. Therefore, the purpose of this study was to examine how different segmentation and feature extraction methods influence the accuracy and computational load of classifying running conditions. Forty-four runners ran at their preferred speed and 25% faster than preferred while an accelerometer at the lower back recorded 3D accelerations. Computational load was determined as the accelerometer signal was segmented into single and five strides, and corresponding small and large windows, with discrete points extracted from the single stride segments and advanced features computed from all four segment types. Each feature set was used to classify speed conditions and classification accuracy was recorded. Computational load and classification accuracy were compared across all feature sets using a repeated-measures MANOVA, with follow-up t-tests to compare feature type (discrete vs. advanced), segmentation method (stride- vs. window-based), and segment size (small vs. large), using a Bonferroni-adjusted α = 0.003. The five-stride (97.49 (±4.57)%) and large-window advanced (97.23 (±5.51)%) feature sets produced the greatest classification accuracy, but the large-window advanced feature set had a lower computational load (0.0041 (±0.0002)s) than the stride-based feature sets. Therefore, using a few advanced features and large overlapping window sizes yields the best performance of both classification accuracy and computational load.