Bilkent University
Department of Computer Engineering
M.S.THESIS PRESENTATION
Structured Estimation for Tracking and Reconstruction under Imperfect Sensing
Yiğit Uz
Master Student
(Supervisor: Asst.Prof.Özgür S.Öğüz )
Computer Engineering Department
Bilkent University
Abstract: This thesis studies structured estimation problems where sensor observations are noisy, incomplete, or weakly calibrated. The first part introduces Interacting Optimized Kalman Filters (I-OKF) for single-object tracking under abrupt maneuvers. I-OKF runs two complementary Optimized Kalman Filter branches with shared symmetric-positive-definite process-noise parameterization and uses a GRU-based refinement module to fuse filter statistics with state-measurement information. Experiments on the OKF simulator, Bar-Shalom maneuvering trajectories, and robotic manipulation trajectories show improved mean-squared error over KF and OKF baselines. The second part introduces Full-Turn Self-Calibrated Visibility-Aware Ray Carving (FT-SC-VARC) for two-dimensional silhouette reconstruction. FT-SC-VARC reconstructs arbitrary planar silhouettes without assuming known rotation center, uniform angular velocity, or a parametric shape prior. It estimates the center and angular trajectory as latent calibration variables, then applies visibility-aware ray carving that separates free space, boundary evidence, and unobserved regions. Experiments on controlled silhouettes, range-noise tests, ablations, and a ShapesAll/MPEG-7-derived benchmark show known-calibration-level reconstruction quality while estimating calibration internally. Together, these works show that classical estimation structure can be preserved while addressing practical sensing imperfections through learned refinement in tracking and visibility-driven self-calibration in reconstruction.
DATE: September 10, Thursday @ 14:00
Place: EA 516