Hidden Markov Model-Based Human Action and Load Classification With Three-Dimensional Accelerometer Measurements

Abstract

This paper proposes a classifier that simultaneously classifies the lifting action and its load weight based on the kinematic quantities of body motion. The proposed classifier is synthesized using the hidden Markov model framework. It is configured to classify an action as lowering, holding, or lifting, and a load weight as 0 kg or 20 kg. We conducted a lifting experiment with 30 subjects and collected the acceleration measurements of five body parts from the subjects when they performed the designated sequence of lifting actions. We performed a five-fold cross validation to evaluate the accuracy and the classification latency of the proposed classifier. The average accuracy of the action classification was determined to be 99.81%, and the average latency for the correct classification was 106.4 ms(±67.9). Further, the average accuracy of the load classification was determined to be 78.70%(±7.55) and the average latency was 88.1 ms(±24.1). Additionally, we proposed a “resetter” of the observed sequence for fast action classification. It was designed for a real world observed sequence that includes transient changes of actions. The result of the video frame analysis demonstrated that the proposed classifier with the resetter classifies the action and the load with an acceptable latency, in which the observed sequence includes the observations of different actions.