Human Activity Classification Based on Angle Variance Analysis Utilizing the Poincare Plot

Abstract

We propose a single sensor-based activity classification method where the Poincare plot was introduced to analyze the variance of the angle between acceleration vector with gravity calculated from the raw accelerometer data for human activity classification. Two datasets named ‘Human Activity Recognition’ and ‘MHealth dataset’ were used to develop the model to classify activity from low to vigorous intensity activities and posture estimation. Short-term and long-term variability analyzing the property of the Poincare plot was used to classify activities according to the vibrational intensity of body movement. Commercially available Actigraph’s activity classification metric ‘count’ resembled value was used to compare the feasibility of the proposed classification algorithm. In the case of the HAR dataset, laying, sitting, standing, and walking activities were classified. Poincare plot parameters SD1, SD2, and SDRR of angle in the case of angle variance analysis and the mean count of X-, Y-, and Z-axis were fitted to a support vector machine (SVM) classifier individually and jointly. The variance- and count-based methods have 100% accuracy in the static–dynamic classification. Laying activity classification has 100% accuracy from other static conditions in the proposed method, whereas the count-based method has 98.08% accuracy with 10-fold cross-validation. In the sitting–standing classification, the proposed angle-based algorithm shows 88% accuracy, whereas the count-based approach has 58% accuracy with a support vector machine classifier with 10-fold cross-validation. In the classification of the variants of dynamic activities with the MHealth dataset, the accuracy for angle variance-based and count-based methods is 100%, in both cases, for fivefold cross validation with SVM classifiers.