A Machine Learning based Human Activity Recognition during Physical Exercise using Wavelet Packet Transform of PPG and Inertial Sensors data

Nowadays, high availability of optical and inertial sensors in fitness trackers and mobile phone attracts researchers attention on human activity recognition(HAR) through photo-plethysmography and inertial sensor data. It is also more suitable than camera based HAR in different circumstances. In this paper, we propose a wavelet packet transform based feature extraction technique from PPG along with accelerometer and gyroscope data to differentiate various physical activities like high and low resistance biking, running and walking. Wavelet packet transform(WPT) is applied on acquired signals and a number of statistical features and entropy feature are extracted. Random forest(RF) classifier is employed in recognition. A 5-fold cross validation is done to examine the performance and an accuracy of 97.8% is achieved which surpasses existing algorithms worked on the same dataset.