This paper investigates the problem of finding the perspective projection and the stereo localization transforms for a binocular imaging system with long baseline. Neural techniques are used to estimate the geometrical mappings from a set of associations between points spread in the 3D space and their corresponding image projections. Aim of this work is to explore how neural networks can deal with acquisition noise and optical distorsions without considering complex camera models. Three techniques are experimentally compared: one based on the simple pin-hole camera model, a second purely based on neural function approximation, and a third which uses neural networks only to account for the deviations of the actual data from the pin-hole model predictions. Experiments have been performed on real data collected by a four-camera system placed in a laboratory room. Performance has been measured in terms of Euclidean distance between actual targets and estimated outputs on a validation and a test set. The results show the effectiveness of the neural approach and validate the combined use of pin-hole model and neural nets to a large extent. The generalization ability of the neural architectures has been finally investigated on a empirical basis
Finding Perspective Projection and Stereo Localization mappings for a Multi-Camera System
Aste, Marco;Boninsegna, Massimo;
1998-01-01
Abstract
This paper investigates the problem of finding the perspective projection and the stereo localization transforms for a binocular imaging system with long baseline. Neural techniques are used to estimate the geometrical mappings from a set of associations between points spread in the 3D space and their corresponding image projections. Aim of this work is to explore how neural networks can deal with acquisition noise and optical distorsions without considering complex camera models. Three techniques are experimentally compared: one based on the simple pin-hole camera model, a second purely based on neural function approximation, and a third which uses neural networks only to account for the deviations of the actual data from the pin-hole model predictions. Experiments have been performed on real data collected by a four-camera system placed in a laboratory room. Performance has been measured in terms of Euclidean distance between actual targets and estimated outputs on a validation and a test set. The results show the effectiveness of the neural approach and validate the combined use of pin-hole model and neural nets to a large extent. The generalization ability of the neural architectures has been finally investigated on a empirical basisI documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
