Human Tangential Activity Recognition Based on Swin Transformer and Supervised Contrastive Learning Using Interferometric Radar

Abstract

Micro-Doppler (mD) signatures are widely used in the field of radar-based human activity recognition (HAR). However, these mD signatures tend to weaken significantly for human tangential activities, which increases the similarity between features of various activities and causes a notable decline in classification performance. To address this issue, this article proposes a human tangential activity recognition method based on a swin transformer (ST) and supervised contrastive learning (SCL) using interferometric radar. The proposed SCL-ST network integrates the feature fusion, ST encoder, and SCL framework. Specifically, a frequency-modulated continuous wave (FMCW) interferometric radar with one transmitter and two receivers is employed to capture echo data of human tangential activities. The echo data are preprocessed to generate single-channel mD spectrograms and dual-channel interferometric spectrograms. The training of the SCL-ST network is divided into two stages. Initially, the SCL-ST network is pretrained using the supervised contrastive loss. The generated mD and interferometric spectrograms are fed into the proposed SCL-ST network, which is equipped with the multilayer perceptron (MLP) projection head. The MLP projection head is capable of mapping feature vectors into a latent space, thereby enhancing the ability of the ST encoder to extract features related to diverse tangential activities. Subsequently, all parameters preceding the MLP projection head are frozen. The MLP classification head is employed to replace the projection head, and the network is then trained using cross-entropy loss to implement downstream classification tasks. Experimental results demonstrate that the proposed SCL-ST network can effectively classify different human tangential activities and achieve a recognition accuracy of 98.89%.