Tracking Fall: Machine Learning Fall Detection Using Radar Point Cloud and Comprehensive Motion with Statistical Features

Fall is one of the most common causes of death for elders. There has been various research on fall detection using either an accelerometer or image-capturing system. Meanwhile, this research would like to try a different approach, utilizing a high-frequency signal with radar and machine learning processing. In this research, the raw data from the radar would be Fourier transformed by the radar into point cloud data, in form of position, velocity, and acceleration. These sequential movement point cloud data would need to be processed statistically, i.e. mean, variance, skewness, and kurtosis, before inputted into the machine learning algorithm. After multiple evaluations, the mean dataset gives the best result, with Random Forest, K-nearest Neighbor, and Gaussian Naive-Bayes algorithms performing very well on this dataset. Random Forest achieved 92 % accuracy for multiple human activity recognition, and K-nearest Neighbor with 100% F -score for ‘Fall’ class.