Radar-Based Whitening-Aided Human Activity Recognition

A variety of deep learning architectures have been proposed for reliable recognition of human activities using micro-Doppler signatures. Convolutional Neural Network (CNN) forms the basis of a majority of these architectures. CNNbased models often utilize batch normalization for efficient network training and enhanced generalization. In this paper, we consider whitening-aided CNN models for human activity recognition, wherein the batch normalization layers are replaced with whitening layers. Whitening provides standardized activations similar to batch normalization, but at the same time, it also decorrelates the activations. Using radar measurements of six different activities, we demonstrate the superiority of whitening over batch normalization in terms of classification accuracy for a CNN-based classifier.