Conceptual exploration and comparative study on the use of deep learning approach in HAR models

Human Activity Recognition (HAR) has taken great attention from researchers last few years, because of the promising results shown by deep learning, and the necessity to make a recognizer system, in this paper a comparison between two types of Convolutional Neural Network (CNN) architectures will be presented. Two Dimensional (2D) CNN followed by a Recurrent Neural Network (RNN) referring to it as 2D-CNN-RNN, and 3D-CNN. Filter with 3D-CNN will be used, after training and testing the models with two different datasets, KTH which has six human activities (Boxing, Handclapping, Handwaving, Walking, Jogging, and Running), and UT-Interaction dataset that has six interaction activities (Handshake, Hug, Kick, Point, Punch, and Push). 3D-CNN shown remarkable results with the aid of filter, but without filter, the dominant was 2D-CNN-RNN models.