CONTRIBUTION A LA STEREOVISION OMNIDIRECTIONNELLE ET AU TRAITEMENT DES IMAGES CATADIOPTRIQUES : APPLICATION AUX SYSTEMES AUTONOMES

摘要

Computer vision and digital image processing are two disciplines aiming to endow computers with a sense of perception and image analysis, similar to that of humans. Artificial visual perception can be greatly enhanced when a large field of view is available. This thesis deals with the use of omnidirectional cameras as a mean of expanding the field of view of computer vision systems. The visual perception of depth (3D) by means of omnistereo configurations, and special processing algorithms adapted to catadioptric images, are the main subjects studied in this thesis. Firstly a survey on 3D omnidirectional vision systems is conducted. It highlights the main approaches for obtaining depth information, and provides valuable indications for the choice of the configuration according to the application requirements. Then the design of an omnistereo sensor is addressed, we present a new configuration of the proposed sensor formed by a unique catadioptric camera, dedicated to robotic applications. An experimental investigation of depth estimation accuracy was conducted to validate the new configuration.Digital images acquired by catadioptric cameras present various special geometrical proprieties, such as non-uniform resolution and severe radial distortions. The application of conventional algorithms to process such images is limited in terms of performance. For that, new algorithms adapted to the spherical geometry of catadioptric images have been developed.Gathered omnidirectional computer vision techniques were finally used in two real applications. The first concerns the integration of catadioptric cameras to a mobile robot. The second focuses on the design of a solar tracker, based on a catadioptric camera.The results confirm that the adoption of such sensors for autonomous systems offer more performance and flexibility in regards to conventional sensors.