Three-dimensional view invariant gait recognition

This thesis investigates the robustness of gait recognition systems to viewpoint changes. The performance of gait recognition systems is often affected by the camera's angle relative to the subject. To assess the severity of this issue, we propose a novel three-fold protocol to determine the effect of the camera's angle on viewpoint robustness. First, we divided gait recognition datasets into $10^\circ$ bins and assigned each gait sequence to a bin. Second, we restructured the datasets such that each gait sequence appears only in side-profile views in the training set and front or back views in the test set. Finally, we rotated the test set to the training viewpoints and evaluated the gait recognition frameworks. Specifically, we applied our protocol to top-performing gait recognition frameworks in the Gait3D and GREW datasets. We also reproduced and validated the results of these gait recognition methods. Additionally, we have provided angular labels of each gait sequence in the Gait3D and GREW datasets. Also proposed two novel gait representations, 1-channel Depth Image and RGB Depth Image. Our results indicate that all models achieved higher accuracy when evaluated on angles closer to those used in their training, and performance is decreased as test angles diverged further. Furthermore, using 3D gait representations combined with 2D representations did not yield improved robustness compared to their 2D counterparts when evaluated with our protocol. Notably, rotating test sets to align with training viewpoints improved performance. Most importantly, we show that using RGB Depth Images in SkeletonGait++ outperformed its Silhouette-based and 1-channel Depth Image-based counterparts. Lastly, we showed that using silhouettes or 2D poses generated from SMPLs or 3D poses leads to a decline in performance.