Multi-camera system design, calibration and three-dimensional reconstruction for markerless motion capture.

摘要

Recently, significant advances have been made in many sub-areas regarding the problem of markerless human motion capture. However, markerless solutions still tend to introduce major simplifications, especially in early stages of the process, that temper the robustness and the generality of any subsequent modules and, consequently, of the whole application. This thesis concentrates on improving the aspects of multi-camera system design, multi-camera calibration and shape-from-silhouette volumetric reconstruction. In Chapter 3, a thoughtful system analysis is first proposed with the objective of achieving an optimal synchronized multi-camera system. Chapter 4 proposes an easy-to-use multi-camera calibration technique to estimate the relative positioning and orientation of every camera with sub-pixel accuracy. In Chapter 5 a robust shape-from-silhouette algorithm, with precise voxel coloring, is developed. Overall, the proposed framework is successful to reconstruct various 3D human postures and, in particular, complex and self-occlusive pianist postures in real-world (minimally constrained) scenes.