GaitGraph: Graph Convolutional Network for Skeleton-Based Gait Recognition (2101.11228v2)

Abstract: Gait recognition is a promising video-based biometric for identifying individual walking patterns from a long distance. At present, most gait recognition methods use silhouette images to represent a person in each frame. However, silhouette images can lose fine-grained spatial information, and most papers do not regard how to obtain these silhouettes in complex scenes. Furthermore, silhouette images contain not only gait features but also other visual clues that can be recognized. Hence these approaches can not be considered as strict gait recognition. We leverage recent advances in human pose estimation to estimate robust skeleton poses directly from RGB images to bring back model-based gait recognition with a cleaner representation of gait. Thus, we propose GaitGraph that combines skeleton poses with Graph Convolutional Network (GCN) to obtain a modern model-based approach for gait recognition. The main advantages are a cleaner, more elegant extraction of the gait features and the ability to incorporate powerful spatio-temporal modeling using GCN. Experiments on the popular CASIA-B gait dataset show that our method archives state-of-the-art performance in model-based gait recognition. The code and models are publicly available.

Insightful Overview of GaitGraph: Graph Convolutional Network for Skeleton-Based Gait Recognition

The paper under consideration presents "GaitGraph," a novel approach leveraging Graph Convolutional Networks (GCNs) for skeleton-based gait recognition. Gait recognition, a biometric modality based on individual walking patterns, offers unique advantages in non-intrusive identity verification over large distances. Traditional techniques predominantly use silhouette images, which may lose spatial detail and capture extraneous visual data, detracting from pure gait feature extraction. This research transitions from silhouette to model-based recognition through the intricate use of human pose estimation and GCNs.

Methodological Advancements

The authors address the limitations of appearance-based methodologies by introducing skeleton poses captured using human pose estimators, consequently providing a refined representation of gait. The GaitGraph methodology implements GCNs to model both spatial and temporal dynamics within gait sequences. This approach draws inspiration from the success of GCNs in action recognition, adapting them to gait recognition tasks to handle the structural graph of the skeleton poses efficiently.

The pipeline involves extracting robust skeleton poses from RGB images, circumventing the challenges of background variation and occlusion. Human poses are derived using a state-of-the-art 2D pose estimator, HRNet, ensuring finely detailed structure representation. Subsequently, these pose sequences are fed through a residual GCN (ResGCN), which has been structurally optimized for the task, culminating in superior discriminative feature extraction. The innovation is reflected in the model's adept capacity to handle variations in walking conditions and view angles, providing a purer form of gait identification compared to approaches laden with clothing and accessory interferences.

Empirical Evaluation

The GaitGraph method underwent rigorous testing using the CASIA-B dataset, which offers substantial diversity in terms of viewing angles and walking conditions. The results are documented against both model-based and appearance-based state-of-the-art methods. Notably, GaitGraph achieved superior performance across various scenarios, particularly excelling in model-based evaluations and demonstrating competitive results against silhouette-dependent techniques.

For instance, in normal walking conditions, GaitGraph achieved an average accuracy of 87.7%, significantly outpacing the pose-based competitor PoseGait, at 68.7%. Even under variations involving background changes or the presence of additional clothing, GaitGraph maintained commendable performance, underscoring its robustness and the effectiveness of its graph-based model under diverse conditions. The research corroborates the potential for GCNs to enhance model-based gait biometrics by emphasizing pure gait data devoid of visual noise typical of silhouette methods.

Implications and Future Directions

The contributions of this paper extend both practically and theoretically within the field of biometric gait recognition. Practically, GaitGraph is poised to impact areas such as security, surveillance, and access control, where non-intrusive and reliable biometric authentication is crucial. Theoretically, this research accentuates the applicability of graph-based methodologies beyond typical domains, fostering further exploration of GCNs in non-traditional machine learning applications.

Future research paths may explore integration with 3D pose estimation, offering richer data context, and increasing resilience against environmental variabilities. Moreover, ongoing enhancements in human pose estimation will likely progressively uplift the utility and accuracy of GaitGraph, making it a viable candidate for real-world deployment.

In conclusion, the GaitGraph framework represents a significant technological and methodological shift in gait recognition, illustrating the potential of graph-based neural networks in biometric authenticity, which could inspire a new wave of innovations in AI-driven identity verification systems.