Keywords

Gait Recognition, Robustness Evaluation, Biometrics

Abstract

This study investigates the robustness of gait recognition systems under realistic perturbations, with a focus on both silhouette parsing and gait recognition components. Experiments are conducted on three diverse datasets—CASIA-B, CCPG, and SUSTech1K—using five state-of-the-art parsing models and six gait recognition architectures. To simulate real-world degradations, we introduce 15 types of noise across five severity levels, resulting in 75 distinct corrupted scenarios. Our findings reveal that no parsing model performs optimally under all conditions, and transformer-based gait models exhibit greater resilience than CNN-based counterparts. However, all models remain highly sensitive to digital noise and occlusions. Interestingly, performance under temporal noise is relatively stable, despite the sequential nature of the task. Additionally, training with noisy data enhances robustness but may reduce peak accuracy on clean inputs. This benchmark provides insights for designing gait recognition systems that are more robust and deployable in unconstrained environments.

Completion Date

2025

Semester

Spring

Committee Chair

Rawat, Yogesh Singh

Degree

Master of Science (M.S.)

College

College of Engineering and Computer Science

Department

Computer Science

Identifier

DP0029386

Document Type

Dissertation/Thesis

Campus Location

Orlando (Main) Campus

Download

Share

COinS