INTELLIGENT RECOGNITION OF HUMAN ACTIVITIES THROUGH MOTION SENSOR PATTERNS FOR CONTEXT-AWARE SYSTEMS

Abstract

The automatic interpretation of human motion via wearable and mobile sensor technologies has become an important part of the Human Activity Recognition (HAR) topic, which is one of the most important issues in context-aware intelligent systems. This study designed a machine learning-based recognition framework for human activities from a dataset of inertial motion sensors (IMU), including 15,980 observations, 6 sensor variables and 6 activity classes. The features used were three-axis accelerometer readings (ax, ay, az) and three-axis gyroscope readings (gx, gy, gz), as well as added magnitude features of the accelerometers and gyroscopes. In order to test the six supervised classifiers, an 80:20 stratified train-test split was used with 5-fold cross-validation. The accuracy, precision, recall, F1-score, specificity, ROC-AUC, and confusion-matrix analysis were used for the evaluation of performance. K-Nearest Neighbors had the highest overall accuracy score of 87.95% and macro F1 score of 87.09%, while Random Forest had the highest ROC-AUC of 0.981. Class 4 and Class 5 had the highest recognition and Class 3 had comparatively higher classification difficulty. The feature importance analysis results show that the most important features for the classification were ax, az and gz, which again demonstrates the complementary importance of both acceleration and rotational motion information. The results showed that compact IMU measurements can be used to successfully enable the reliable detection of human activity in smart home, healthcare monitoring, wearable and smart mobile applications.