Skip to Main content Skip to Navigation

Vues Multiples non-calibrées : Applications et Méthodologies

Abstract : The search for patterns of interest contained in multiple view sequences of the same scene remains one of the major problems in the computer vision community at present. Despite the great advances developed in the last 40 years, computer vision has not yet answered adequately how to infer and detect patterns of interest in scenes in which one or more objects are viewed from multiple points-of-view. To overcome this drawback, this thesis proposes new algorithms and prototypes capable of characterizing, inferring, and detecting patterns of interest in uncalibrated multiple view sequences, i.e. without a priori knowledge about the position of the object(s) in regard to the camera(s). The work is focused on three main themes divided into six papers that constitute the body of the thesis. (1) The point-to-point correspondence analysis by means of explicit and implicit markers on objects. (2) The point-to-point correspondence estimation through multiple geometric relations independent of the object(s). (3) The dynamic flow prediction generated by camera's motion around the object. The main purpose of this thesis is to support the decision making by means of the dynamic and/or geometric analysis of the object(s) movement or the camera(s) movement as these move. Thanks to this analysis, it is possible to increase the information about the scene and the object(s) through an inferential process specific in each case. As a result, it is possible for instance to assist the inspection process performed by a human operator, to determine the product quality automatically, or to execute a specific action in a robotic actuator. Although these issues have different purposes, these resolve the same set of steps concerning to: (1) Determining corresponding points in multiple images. (2) Determining the geometric and/or dynamic relation between the previously estimated correspondences. (3) Inferring new relations at those points whose correspondence is not known in order to characterize the movement. The major changes come in how the correspondence is estimated, in the calculation of the dynamics and geometry between corresponding points, and finally in how we use the inference of motion to identify a new correspondence or to characterize the motion of the points. Among the major results, the development of a new uncalibrated inspection methodology with multiple views applied to the quality analysis in automotive wheel rims, the development of a functional prototype applied to the wine bottleneck inspection, a geometric point-to-point correspondence methodology able to solve the correspondence problem in two and three views for any point of interest, and finally the human intention recognition for grasping tasks through the movement analysis of the eyes, and the hand. As future work, it remains to study the dynamic correspondence by means of invariant features, to use geometric methods in radiologic image sequences, and use intention recognition models in order to evaluate the object-quality.
Complete list of metadatas

Cited literature [254 references]  Display  Hide  Download
Contributor : Adela Kabaklija <>
Submitted on : Wednesday, November 30, 2011 - 3:55:04 PM
Last modification on : Friday, March 22, 2019 - 1:31:28 AM
Document(s) archivé(s) le : Thursday, March 1, 2012 - 2:31:16 AM


  • HAL Id : tel-00646709, version 1


Miguel Carrasco. Vues Multiples non-calibrées : Applications et Méthodologies. Automatique / Robotique. Université Pierre et Marie Curie - Paris VI, 2010. Français. ⟨tel-00646709⟩



Record views


Files downloads