Classification Of Human Activity Recognition Using FMCW Radar With A 3D Voxel Model Approach From Point Cloud Data

The classification of human activity (Human Activity Recognition/HAR) is a critical research field in the development of monitoring systems, to minimize the dangers caused by every movement performed by humans, one of which is minimizing the risk of falling. This study discusses movement classification using a 3D voxel approach formed from point cloud data to recognize four main activities sitting, standing, walking, and falling. The data was collected using the FMCW radar IWR6843AOPEVM which records the position of objects in three-dimensional space. This study explores the effect of windowing on model performance to obtain the best windowing range that can produce optimal 3D voxel representations. The model architecture used is the 3D Convolutional Neural Network (3D CNN). Using the 3D voxel model and 3D CNN architecture, the best result was achieved with a 30-windowing model. The model training produced a good result of 94%. From the experimental results, the use of appropriate windowing is very influential on system accuracy, and a model architecture with certain configurations can detect activities effectively, especially for important events such as falling. This approach shows great potential in the implementation of automated deep learning-based monitoring systems.