Human Activity Recognition : Preliminary Results for Dataset Portability using FMCW Radar

• Published in: Proceedings of the IEEE Radar Conference pp. 1 - 4
• Main Authors: Shah, Syed Aziz, Fioranelli, Francesco
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.09.2019
• Subjects: Fall Detection; Machine Learning; Radar Technology; Transfer Learning
• ISSN: 2640-7736
• DOI: 10.1109/RADAR41533.2019.171307

This paper presents some preliminary results to develop a generalized system for human activity recognition (HAR) and detecting fall events using micro-Doppler signatures exploiting frequency modulated continuous wave (FMCW) radar. The core idea of this work is to demonstrate the portability and applicability of radar datasets for HAR, independent of geometrical environments and subjects involved. The experimental campaign involved different volunteers at four different geometrical locations. Two different machine learning algorithms such as support vector machine (SVM) and k-nearest neighbour (KNN), and one deep learning classifier namely GoogleNet are used to classify various human activities. The transfer learning method leveraging AlexNet algorithm is used to extract features from spectrograms to train and test the SVM and KNN classifiers. Four different scenarios are presented where datasets from three locations are combined to train and validate the classifiers, and test it on the remaining (leave-one-out) one. It is observed that the GoogleNet algorithm provides a consistent test accuracy between 68.5% to 81% for four different locations.