One-shot three-dimensional scene analysis
This thesis studies the problem of extracting three-dimensional information of a scene by means of projecting structured light over it. Instead of measuring the depth of the scene as most techniques do, we propose a novel and efcient approach to estimate depth's gradient map. This information is ext...
Guardat en:
| Autor principal: | |
|---|---|
| Format: | doctoralThesis |
| Idioma: | anglès |
| Publicat: |
2015
|
| Matèries: | |
| Accés en línia: | http://hdl.handle.net/20.500.12008/5218 |
| Etiquetes: |
Sense etiquetes, Sigues el primer a etiquetar aquest registre!
|
| Sumari: | This thesis studies the problem of extracting three-dimensional information of a scene by means of projecting structured light over it. Instead of measuring the depth of the scene as most techniques do, we propose a novel and efcient approach to estimate depth's gradient map. This information is extremely useful if we want to recognize or automatically understand what we are looking at. For example we will show how gradient information obtained with the proposed technique can be used for face recognition. We also study the problem of how to integrate a retrieved gradient field; these techniques combined with the proposed projection approach allows us to estimate the height map of a scene. The main difficulty of dealing with empirically obtained gradient fields is that they are not-integrable (i.e. they are not irrotational fields); hence we must use robust techniques that overcome this issue. In addition, to study the two-dimensional integration problem, more general approaches are also considered; we will show that these approaches can integrate gradient fields of dynamic scenes exploiting the data's temporal correlations to regularize the space of solutions. |
|---|