Emerging trends in gait recognition based on deep learning: a survey
Vaishnavi Munusamy,Sudha Senthilkumar
DOI: https://doi.org/10.7717/peerj-cs.2158
2024-07-10
Abstract:Gait recognition, a biometric identification method, has garnered significant attention due to its unique attributes, including non-invasiveness, long-distance capture, and resistance to impersonation. Gait recognition has undergone a revolution driven by the remarkable capacity of deep learning to extract complicated features from data. An overview of the current developments in deep learning-based gait identification methods is provided in this work. We explore and analyze the development of gait recognition and highlight its uses in forensics, security, and criminal investigations. The article delves into the challenges associated with gait recognition, such as variations in walking conditions, viewing angles, and clothing as well. We discuss about the effectiveness of deep neural networks in addressing these challenges by providing a comprehensive analysis of state-of-the-art architectures, including convolutional neural networks (CNNs), recurrent neural networks (RNNs), and attention mechanisms. Diverse neural network-based gait recognition models, such as Gate Controlled and Shared Attention ICDNet (GA-ICDNet), Multi-Scale Temporal Feature Extractor (MSTFE), GaitNet, and various CNN-based approaches, demonstrate impressive accuracy across different walking conditions, showcasing the effectiveness of these models in capturing unique gait patterns. GaitNet achieved an exceptional identification accuracy of 99.7%, whereas GA-ICDNet showed high precision with an equal error rate of 0.67% in verification tasks. GaitGraph (ResGCN+2D CNN) achieved rank-1 accuracies ranging from 66.3% to 87.7%, whereas a Fully Connected Network with Koopman Operator achieved an average rank-1 accuracy of 74.7% for OU-MVLP across various conditions. However, GCPFP (GCN with Graph Convolution-Based Part Feature Polling) utilizing graph convolutional network (GCN) and GaitSet achieves the lowest average rank-1 accuracy of 62.4% for CASIA-B, while MFINet (Multiple Factor Inference Network) exhibits the lowest accuracy range of 11.72% to 19.32% under clothing variation conditions on CASIA-B. In addition to an across-the-board analysis of recent breakthroughs in gait recognition, the scope for potential future research direction is also assessed.