Comparison of algorithms and classifiers for stride detection using wearables

Sensor-based systems for diagnosis or therapy support of motor dysfunctions need methodologies of automatically stride detection from movement sequences. In this proposal, we developed a stride detection system for daily life use. We compared mostly used algorithms min–max patterns, dynamic time warping, convolutional neural networks (CNN), and automatic framing using two data sets of 32 healthy and 28 Parkinson’s disease (PD) persons. We developed an insole with force and IMU sensors to record the gait data. The PD patients carried out the standardized time up and go test, and the healthy persons a daily life activities test (walking, sitting, standing, ascending and descending stairs). As an automatically stride detection process for daily life use, we propose a first stride detection using automatic framing, and after normalization and resampling data a CNN is used. A F1-score of 0.938 (recall 0.968, precision 0.910) for time up and go test and of 0.944 (recall 0.992, precision 0.901) for daily life activities test were obtained for CNN. Compared to the other detection methods, up to 6% F-measure improvement was shown.