CART-Gait: Cross Angle Refined Training of Cross-View Gait Recognition

Visual-based gait recognition is a promising biometric identification technology that utilizes visual-based techniques to identify individuals based on their distinctive walking patterns. However, existing gait recognition methods face a critical challenge: maintaining a delicate balance between accuracy and robustness, particularly in the face of external variables such as changes in clothing or carrying conditions. To address these persistent challenges, this study introduces a novel visual-based multi-modal cross-view gait recognition algorithm. The proposed algorithm utilizes both graph convolutional neural networks (GCNs) and convolutional neural networks (CNNs) to extract features from joint position, velocity, and bone direction information in addition to the traditional method of just processing traditional silhouette image sequences by a CNN network. These features are extracted separately and then combined adaptively from the two branches. Notably, in comprehensive evaluations using the CASIA-B [1] dataset, our algorithm has demonstrated state-of-the-art performance. Importantly, these results inspire confidence in the algorithm’s potential to significantly enhance gait recognition accuracy in practical, real-world scenarios.