Eliminate Aspect Angle Variations for Human Activity Recognition using Unsupervised Deep Adaptation Network

Activity recognition and monitoring using radar micro-Doppler signatures (mu-DS) classification has played an vital part in various security and healthcare applications. In the practical scenario, aspect angle variations of u-DS increase the data diversity but can be regarded as a distraction factor for activity recognition. The learned feature extractor and classifier will degrade a lot if the test u-DS is from a different aspect angle from the training dataset. This is because the aspect angle variations between training and test dataset will break the assumption of the classification methods: the training and test data are drawn from the same distribution. This paper aims to eliminate the aspect angle variations by learning aspect angle invariant and meanwhile discriminative features in the bi-static radar geometry using the unlabeled test data. More specifically, we first propose a new problem to train a feature extractor using certain aspect angles but generalizes well for other aspect angles in the test stage. Next, we propose two adaptation networks termed as MMD-DAN and JS-DAN, utilizing two widely used distribution divergence measurements. Finally, we evaluate our experimental setting and methods using experimental data.