Joint fall and aspect angle recognition using fine-grained micro-Doppler classification

Abstract

Activity recognition and monitoring are finding important applications in ambient assisted living healthcare. Among the various types of motions which researchers are attempting to detect and recognize, fall detection has gained significant interest. In this paper, we investigate the application of high-frequency (24 GHz) FMCW radar for multi-perspective micro-Doppler (μ-D) activity recognition. Data from two different types of motion; falling and picking-up an object, were collected from three aspect angles and put through a fine-grained classifier to not only differentiate the motions, but to also identify their aspect towards the radar receivers. A key novel component of this work is the application of the fine-grained classification task, where a label discriminate sparse representation classifier is proposed to improve recognition performance over very similar μ-D signatures. This is achieved by learning a discriminate dictionary constrained by the label information and meanwhile preventing the overfitting problem. The greatest increase in classification performance was found to be of the order of 8 %.