Adaptive Radar-Based Human Activity Recognition With L1-Norm Linear Discriminant Analysis

• Published in: IEEE journal of electromagnetics, RF and microwaves in medicine and biology Vol. 3; no. 2; pp. 120 - 126
• Main Authors: Markopoulos, Panos P., Zlotnikov, Sivan, Ahmad, Fauzia
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 01.06.2019; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Biology; Classifiers; Data analysis; Discriminant analysis; Doppler effect; Electromagnetics; faulty measurements; human activity classification; Human activity recognition; L1-norm; machine learning; outlier resistance; Outliers (statistics); Radio frequency; Subspaces; Training; Training data
• ISSN: 2469-7249; 2469-7257
• DOI: 10.1109/JERM.2019.2893587

We present a novel radar-based indoor human gross motor activity classifier, which employs L1-norm linear discriminant analysis (L1-LDA) to identify low-rank subspaces whereon micro-Doppler signatures from distinct motions are most differentiable. Both nonadaptive and adaptive implementations of the proposed classifier are presented, with the latter providing refinement and adaptation to the specific activity patterns of the human subject of interest. In contrast to standard LDA, L1-LDA exhibits resistance against outliers that may lie among the training data, e.g., due to mislabeling. We use real-data from four motion classes to experimentally compare the performance of the proposed methods with standard (L2-norm-based) LDA. The results corroborate that the proposed methods markedly outperform LDA when the training datasets are corrupted with mislabeling, while they provide similar performance under nominal training data.