Few-shot Human Motion Recognition through Multi-Aspect mmWave FMCW Radar Data

TL;DR

This paper introduces channel-DN4, a novel multi-aspect few-shot human motion recognition method using mmWave FMCW radar data, which effectively reduces overfitting and improves recognition accuracy in limited data scenarios.

Contribution

The paper proposes channel-DN4, combining local descriptors, episodic training, and channel attention to enhance multi-aspect few-shot human motion recognition from radar data.

Findings

Achieved 87.533% accuracy in 3-way 10-shot recognition.

Outperformed existing methods in overfitting resistance.

Validated effectiveness on real radar data.

Abstract

Radar human motion recognition methods based on deep learning models has been a heated spot of remote sensing in recent years, yet the existing methods are mostly radial-oriented. In practical application, the test data could be multi-aspect and the sample number of each motion could be very limited, causing model overfitting and reduced recognition accuracy. This paper proposed channel-DN4, a multi-aspect few-shot human motion recognition method. First, local descriptors are introduced for a precise classification metric. Moreover, episodic training strategy was adopted to reduce model overfitting. To utilize the invariant sematic information in multi-aspect conditions, we considered channel attention after the embedding network to obtain precise implicit high-dimensional representation of sematic information. We tested the performance of channel-DN4 and methods for comparison on measured mmWave FMCW radar data. The proposed channel-DN4 produced competitive and convincing results, reaching the highest 87.533% recognition accuracy in 3-way 10-shot condition while other methods suffer from overfitting. Codes are available at: https://github.com/MountainChenCad/channel-DN4