Evaluating Deep Learning in Gait Recognition

User recognition is an important technology to identify and distinguish individuals based on certain characteristics or biometric data in various contexts such as in a system and application. It is a basis for building a secure user authentication or identification scheme. For example, gait recognition aims to verify and identify an individual based on the walking style with features such as stride length, speed, and joint angles. With continuous technological advancements, gait recognition is expected to be employed in more practical scenarios, bringing convenience and enhanced security. For better processing the data, deep learning has been widely applied in gait recognition. However, high variability in gait is still an open challenge to build a practical gait recognition system. In this work, we aim to investigate the usage of deep learning in gait recognition. In particular, we explore different neural network models including C3D, CNN-LSTM, CNN-Res-LSTM, ViViT and CNN-Transformer, and study the effect of different gait video directions on model accuracy. In the end, we discuss the potential security threats and open challenges of deep learning-based gait recognition.