Few-shot Omnidirectional Human Motion Recognition Using Monostatic Radar System

• Published in: IEEE transactions on instrumentation and measurement Vol. 72; p. 1
• Main Authors: Yang, Yang, Li, Junhan, Li, Beichen, Lang, Yue, Zhai, Wenguang
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.01.2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Classifiers; Data augmentation; Data collection; Doppler effect; Few-shot Learning; Human activity recognition; Human motion; Measurement; Meta Learning; Metalearning; Metric space; Micro-Doppler Signatures; Monostatic Radar; Motion perception; Radar equipment; Radar systems; Recognition; Spectrogram; Spectrograms; Task analysis; Testing; Training
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2023.3328079

The susceptibility of the Doppler effect to aspect angle presents a challenge for monostatic radar systems when it comes to achieving omnidirectional human motion recognition. In the context of human movement in free space, the micro-Doppler-based classifier exhibits an 'angle-sensitive' behavior. This means that the classifier's performance becomes unpredictable when the aspect angle changes. In addition to this issue, radar-based classifiers often struggle with limited support from training data. To address this issue, we propose integrating a meta data augmentation strategy into the meta-learning process. The goal of this strategy is to enhance the generalization ability of the features across tasks. Additionally, we introduce a method for predicting a local-to-global similarity score. The purpose of this scheme is to calibrate the parametric metric space so that it is suitable for omnidirectional radar spectrograms. We have constructed a monostatic radar system and creating a radar spectrogram dataset for few-shot omnidirectional motion recognition. To minimize the need for data collection, we only use radar spectrograms with human subjects who are moving toward the radar for training. Previous studies on omnidirectional motion classification, on the other hand, use spectrograms in all directions as training supports. The results indicate that the proposed method performs better than seven state-of-the-art algorithms in solving the problem of recognizing omnidirectional human motion in scenarios with limited data.