Enhancing FMCW Radar-Based Human Activity Recognition using DI-ResNet

Abstract

This paper proposes a method to enhance the quality of Human Activity Recognition (HAR) based on Frequency Modulated Continuous Wave (FMCW) radar. The method utilizes a Dual Input-ResNet (DI-ResNet) model to address the limitations associated with relying solely on isolated range or micro-Doppler (m-D) features for activity classification. Specifically, two parallel ResNet-18 backbones are employed to extract deep semantic features from two distinct data domains: the Range-Time (RT) spectrogram and the Doppler-Time (DT) spectrogram. These features are subsequently fused via a concatenation layer to synthesize global context, thereby generating a more comprehensive representation of the performed activities. Experimental results demonstrate that the proposed method achieves superior recognition performance compared to single-stream baseline networks. Notably, the proposed model effectively mitigates confusion among activities exhibiting kinematic similarity, improving the Recall metric for the “SitDown” activity by 42.1% compared to traditional single-input methods. Furthermore, the accuracy for complex hand gestures such as "Drink" significantly increased by 19.8%, substantially minimizing the high misclassification rate associated with "Grab." Finally, the model achieves near-ideal reliability for safety-critical activities, attaining a recognition accuracy of 99.5% in fall detection, thereby confirming its potential for practical deployment.