Radar Point Cloud Processing Methods for Human Activity Classification With Point Transformer Networks

Classification of sequences of human activities performed in a continuous manner and with unconstrained duration using radar sensors has been studied in this work. A processing method and accompanying classifier based on Point Transformer networks have been proposed to address this challenge. The method has been experimentally verified on a challenging, publicly available dataset collected with a network of five simultaneously operating Ultra Wide Band pulsed radars. Classification performance has been compared to reference works in literature, and a test accuracy and macro F1-score of 86.9% and 78.7% have been respectively demonstrated.