Philippe Quinio
Accurate Reconstruction of 3D Scenes from Multiple Imprecise and Uncertain Data
Abstract:We have proposed a new approach to the reconstruction of 3 dimensional scenes from multiple data. The information extracted from any stereoscopic image pair is inherently imprecise and uncertain and therefore a mathematical framework that can model both imprecision and uncertainty is required. Due to the limitations of the random point approach, most of the methods proposed so far cannot fully represent imprecision and uncertainty.
These 2 forms of error get mixed into a (3D) probability density function and most methods are compelled to make unrealistic assumptions about the prior probability distribution of disparity vectors or even of the 3D scene features.
We have proposed to use the Random Closed Set theory for this purpose. The resulting integration method makes no assumption whatsoever about the noise/errors and does not require prior registration of all the feature points in time. Experimental results have confirmed that one can decrease both imprecision and uncertainty at the same time while increasing completeness in the final map, provided that all sources of error are duly investigated and suitably represented.
Further theoretical and experimental work needs to be done so as to quantify more precisely the performance of the proposed method, and in particular to determine how and in what extent the complexity of a scene determines the optimal number of 2D views that must be processed in order to describe it accurately.