Gait Recognition Based on Two-Stream CNNs With Multisensor Progressive Feature Fusion

Gait recognition has emerged as a significant and promising solution for human identification. However, existing multisensor-based gait recognition studies have the problem of insufficiently utilizing the already extracted features during the feature fusion process. In this study, we proposed a novel method based on two-stream convolutional neural networks (Two-Stream CNNs) with progressive multisensor feature fusion for gait recognition. This method fuses gait data collected from inertial sensors and pressure sensors embedded in pressure insoles. A progressive feature fusion (PFF) module is designed to fully utilize the valuable information within the multistage gait features extracted by convolutional neural networks (CNNs). A feature enhancement channel attention (FECA) module is designed to highlight feature representations that emphasize the importance of crucial sensor data. Meanwhile, an attention-based feature fusion (ABFF) module is designed for more flexible feature fusion across different modalities and multistage gait features. The experimental results demonstrate that our method shows a higher recognition accuracy compared to the state-of-the-art studies, indicating the superiority and effectiveness of our method.