Fitness Activity Classification Using mmWave Radar Point-Cloud And Machine Learning

This paper experimentally demonstrates the use of various machine learning algorithms for fitness activity classification using a single mmwave radar point cloud data. Data from two users were collected for seven fitness activities in 11 different sessions amounting to approximately 8.5 hours of data. Two approaches have been proposed for classification which achieve 98% and 96% accuracy when user is performing exercises at a fixed position w.r.t. the radar. However, these approaches do not perform well when there is position variation of the user in the radar field of view due to significant changes in the range and angle values. We also present that using radar outputs (range, Doppler and angle data) altogether increases the accuracy of the classifiers. The system can enable contactless and privacycompliant human activity recognition.