Model-Based Gait Recognition Using Multi-Level Graph Neural Networks With Stochastic Process and Skip Connections

Gait recognition has become a vital tool in security, authentication and surveillance due to its abilities to perform at a distance and without the subject’s awareness, unlike traditional biometrics such as facial, iris or fingerprint recognition. In this work, we propose a new model-based gait recognition technique utilizing multi-level Graph Neural Networks (GNNs) that addresses two main challenges: shot walks and viewpoint issues. Human skeletal structures and movements are represented in the form of multi-level graphs. The proposed technique uses both frame-level graphs (postures) and subwalk-level (quick movements) as gait features to capture relationships among the entire body and characteristic of quick movements. The stochastic process and skip connections within GNN framework are employed to enhance flexibility, increase robustness, prevent overfitting and improve gradient flow. Experiments are conducted on a dataset of 2-second walks, where each sequence contains only 40 frames, to evaluate the performance of the proposed technique on very short walk sequences. The proposed technique outperforms existing techniques on both top-k accuracy test and the Cumulative Matching Characteristic (CMC) Curves. The experimental results show that the proposed technique achieves 85.54% on the top-1 accuracy test and surpass 95% accuracy in ranks below 10. These results suggest that unique posture and quick movement are sufficient to identify a person. The strong performance of the proposed technique on dataset that contains short walking sequences and viewpoint variations demonstrates its strong potential for real-world applications in security, surveillance and authentication where accurate identification from limited gait data obtained from various observation angles is crucial.