PENDETEKSIAN AKTIVITAS MANUSIA DENGAN HUMAN POSE ESTIMATION DAN CONVOLUTIONAL NEURAL NETWORK

Abstract

In the last few years, artificial intelligence-based human activity monitoring system becomes more popular. However, most proposed research focused on adult subject performing the action. Thus, leaving a gap in child activity recognition which causes low dataset availability and reference. Based on this situation, the focus of this study is to reduce the gap in child activity recognition. The proposed system in this study focused on 4 – 6 years old child. The methods used are Human Pose Estimation with BlazePose and Convolutional Neural Network (CNN) with images dataset gathered from the internet. First, the skeleton will be estimated using BlazePose, the resulting skeleton will be converted to matrix form and given to CNN to be classified. There are 3 activites which can be detected, they are studying, standing, and sleeping. Each activity will be recorded to a logbook with its timestamp when the activity detected. Confusion matrix testing shows that trained model has accuracy value of 97.77%, precision of 97.96%, recall of 97.13%, and F1-score of 97.83%.