An adaptive, real-time cadence algorithm for unconstrained sensor placement

Abstract

This paper evaluates a new and adaptive real-time cadence detection algorithm (CDA) for unconstrained sensor placement during walking and running. Conventional correlation procedures, dependent on sensor position and orientation, may alternately detect either steps or strides and consequently suffer from false negatives or positives. To overcome this limitation, the CDA validates correlation peaks as strides using the Sylvester's criterion (SC). This paper compares the CDA with conventional correlation methods. 22 volunteers completed 7 different circuits (approx. 140 m) at three gaits-speeds: walking (1.5 m s^-1), running (3.4 m s^-1), and sprinting (5.2 and 5.7 m s^-1), disturbed by various gait-related activities. The algorithm was simultaneously evaluated for 10 different sensor positions. Reference strides were obtained from a foot sensor using a dedicated offline algorithm. The described algorithm resulted in consistent numbers of true positives (85.6-100.0%) and false positives (0.0-2.9%) and showed to be consistently accurate for cadence feedback across all circuits, subjects and sensors (mean ± SD: 98.9 ± 0.2%), compared to conventional cross-correlation (87.3 ± 13.5%), biased (73.0 ± 16.2) and unbiased (82.2 ± 20.6) autocorrelation procedures. This study shows that the SC significantly improves cadence detection, resulting in robust results for various gaits, subjects and sensor positions.

Keywords: Accelerometer; Autocorrelation; Cadence; Cross-correlation; Gait cycle detection; Running; Stride frequency; Sylvester's criterion.