Tokenization of Skeleton-based Transformer Model for Cross-View Gait Recognition

Gait, people walking manners, is unique que for person identification and it is used in scenes such as healthcare, security systems, and criminal investigation. Nowadays, 3D gait models created by multi-view cameras as model-based methods are attracting attention. It is expected to be more accurate and robust systems than created using single-view methods. The OUMVLP-Mesh database is a gait database that includes 3D human models created through multi-view training. In this paper, we propose skeleton based gait recognition methods with Transformer architecture to analyze the gait sequence by treating it as the sequence of tokens. Specifically, we represent the gait sequence by converting skeleton models into tokens and our model facilitates temporal analysis using the attention architectures. This approach allows us to aggregate comprehensive temporal information by taking key features based on similarities in pose or behavior over all frames. Similar behavior appears as large value in output features of our model, and we only take these enhanced features. It helps to find unique identity which can be used for person identification tasks. Furthermore, we introduce a method of partitioning tokens to enhance attention efficiency. It divides the input tokens into each component and process them with different Transformer architectures. It leverages more specific unique feature by capturing finer behavior. Our methods are tried on 3D skeleton models from OUMVLP-Mesh database, and it achieved approximately an improvement of 18% in person identification rank-l rates compared to the benchmark method.