Wearable Sensor-Based Framework for the Detection of Daily Living Activities Utilizing In-Depth Features

Abstract

AbstractThis paper implements the physical activity recognition model on 13 (sitting, standing, ironing, etc.) daily living activities using a 2D convolutional neural network (CNN). The model has outperformed the existing methods for the PAMAP2 and WISDM datasets. The various machine learning models are also implemented. These are SVM, Naïve Bayes, decision tree, and random forest. Average accuracy, F1 score, precision, and recall are the performance metrics used to evaluate the results. It is found in overall findings that convolutional neural networks have detected all the activities very precisely and accurately where machine learning models felt difficulty in classifying more accurately a particular activity. This difficulty of these machine learning models is addressed in this research by introducing CNN features in the training of machine learning models. Mean accuracy of 99.13 and 99.18% is achieved with convolutional neural network and Naïve Bayes with CNN features, respectively, for the PAMAP2 dataset. Mean accuracy of 98.01 and 98.09% is achieved with convolutional neural network and Naïve Bayes with CNN features, respectively, for the WISDM dataset. KeywordsConvolutional neural networkPhysical activityWearable sensorsMachine learningInertial measurement unitsCNN features