SEQUENTIAL DEEP LEARNING MODELS FOR HUMAN SKELETON-BASED GAIT RECOGNITION

Abstract

Gait is the walking posture and dynamics of a person and is considered a unique biometric of a person. Based on this biometric and using various algorithms one can recognize the person’s identity with high precision. Although many conventional machine learning and deep learning methods that learn a person’s identity based on their gait have been proposed, there still exist some typical limitations. While conventional machine learning methods are error-prone to the background noise, more advanced methods based on CNN models do not well capture the temporal dependencies in terms of inter-frame correlation for gait recognition. This work addresses the limitations of previous works by employing deep sequential models and demonstrating their effectiveness and efficiency in learning gait information using 3D human skeletal data. We propose a person identification model, that uses both the joint coordinates and features derived from them, including joint distance, joint orientation, and joint velocity. Sequential models based on Long Short-Term Memory and Transformer networks capture the spatial correlations of skeletal joints within a single frame and the temporal dependencies and dynamics of the joints throughout a sequence of frames. The effectiveness and efficiency of the proposed methods, the impact of data augmentation methods, a combination of derived gait features were studied and analyzed. The experimental results show that the proposed models achieve high person identification accuracy on the UPCV Gait (98.26%), and KS20 VisLab Multi-View (90.86%) datasets, which are competitive to the previous state-of-the-art methods.