Information fusion and evidential grammars for object class segmentation

Jean-Baptiste Bordes (1, 2) , Philippe Xu (3, 1, 2) , Franck Davoine (3, 1, 2) , Huijing Zhao (3) , Thierry Denoeux (1, 2)

1 Labex MS2T - Laboratoire d'Excellence "Maîtrise des Systèmes de Systèmes Technologiques"
2 Heudiasyc - Heuristique et Diagnostic des Systèmes Complexes [Compiègne]
3 Peking University [Beijing]

In this paper, an original method for traffic scene images understanding based on the theory of belief functions is presented. Our approach takes place in a multi-sensors context and decomposes a scene into objects through the following steps: at first, an over-segmentation of the image is performed and a set of detection modules provides for each segment a belief function defined on the set of the classes. Then, these belief functions are combined and the segments are clustered into objects using an evidential grammar framework. The tasks of image segmentation and object identification are then formulated as the research of the best parse graph of the image, which is its hierarchical decomposition from the scene, to objects and segments while taking into account the spatial layout. A consistency criterion is defined for any parse tree, and the search of the optimal interpretation of an image formulated as an optimization problem. We show that our framework is flexible enough to include new sensors as well as new classes of object. The work is validated on real and publicly available urban driving scene data.

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Jean-Baptiste Bordes, Philippe Xu, Franck Davoine, Huijing Zhao, Thierry Denoeux. Information fusion and evidential grammars for object class segmentation. Fifth IROS Workshop on Planning, Perception and Navigation for Intelligent Vehicles, Nov 2013, Tokyo, Japan. pp.165-170. ⟨hal-00932899⟩