RGB Gait Recognition Using Large Vision Models for Industrial Access Control

Identity verification is important for safety management at industrial access points. Gait recognition provides contactless identity cues as workers pass through a gate, but public datasets rarely combine card-swiping, moving barriers, standard workwear, safety helmets, directional occlusion, and changing illumination. In this application-oriented empirical study, we evaluated 1500 RGB access sequences from 150 workers in our previously introduced industrial gait dataset. Each worker was recorded in five passage sequences in each of two opposing, slightly elevated views. Using the grouped, memory-efficient BiggerGait* framework, we compared DINOv2 ViT-S/14, DINOv3 ViT-S/16, and DINOv3 ViT-S+/16 under the same downstream configuration. The human-prior branch did not consistently produce stable person-centered masks when trained from scratch and transferred poorly between backbones. We therefore evaluated raw RGB, background-suppressed RGB, and background-suppressed RGB with a one-patch boundary expansion. Strict suppression mainly improved same-view recognition, whereas the expanded input increased mean cross-view Rank-1 for all three backbones in the present evaluation. This pattern is consistent with a benefit from reducing scene interference, while limited boundary context may recover useful body-edge cues for matching opposing views. DINOv3-S+ with the expanded input achieved same-view and cross-view mean Rank-1 accuracies of 96.3% and 88.3%, respectively.