This is published under the terms of CC BY-NC.

Downloaded: 465 times

Abstract

Gait recognition based on millimeter waves (mmWave) has a wide range of applications, such as smart homes and health monitoring, and attracted extensive attention due to its non-contact, privacy protection, and light-independent characteristics. Existing gait recognition based on mmWave performs well on the datasets collected at a single time or environment. However, the recognition accuracy declines significantly with the time and environment domains shifting, which affects its practical applications. In this paper, we propose a novel mmWave gait recognition CRNet that is robust to both time and environment, which is realized through two-stage training. Specifically, the first stage designs a contrastive learning strategy to pre-train the encoder module, which aims at learning the general gait features across different seen time and environment domains. The second stage further trains the classification module based on specific recognition tasks. After the two-stage training, CRNet experiments on test sets with unseen time or environment domains. We collect a mmWave-based multiperson gait recognition dataset with multiple time and environment domains. Experiments show that CRNet still performs well in unfamiliar domains which increases the accuracy from 75.7% to 91.2% compared to the baseline.