Feasibility test of activity index summary metric in human hand activity recognition

Abstract

Activity monitoring is a technique for assessing the physical activity that a person undertakes over some time. Activity Index (AI) is a metric that summarizes the raw measurements from tri-axial accelerometers, often used for measuring physical activity. Our research compared the Activity Index for different activity groups and hand usage [1]. We also tested this metric as a classification feature, and how different data acquisition and segmentation parameter configurations influence classification accuracy. Data acquisition was done with a previously developed system that includes a smartwatch on each wrist and a smartphone placed in the subject’s pocket; raw data from smartwatch accelerometers was used for the analysis. We calculated the Activity Index for labeled data segments and used ANOVA1 statistical test with Bonferroni correction. Significant differences were found between cases of hand usage (left, right, none, both). In the next analysis phase, the Activity Index was used as the classification feature with three supervised machine learning algorithms-Support Vector Machine, k-Nearest Neighbors, and Random Forest. The best accuracy (measured by F1 score) of classifying hand usage was achieved by using the Random Forest algorithm, 50 Hz sampling frequency, and a window of 10 s without overlap for AI calculation, and it was 97%. On the other hand, the classification of activity groups had a low accuracy, which indicated that a specific activity group can’t be identified by using only one simple feature.

Keywords

• activity index
• accelerometry
• smartwatches
• anova1
• classification
• machine learning
• random forest