Cumulative distribution transform for 3D surface segmentation

Description: Topic description Introduction / background: The 3D representations of real objects from scanners are made up of a large number of points. These points are then generally meshed to define the surface (s) of the 3D object.When very complex scenes are scanned (example of a factory in the cases treated by the company ATS Engineering), an important problem is to subdivide the surface 3D scene into elements: for example tubes, tanks, etc. motors, valves, etc. In the case of moving a robot (or arm), the time constraint for recognition is essential.On the other hand, 3D modeling involves surface elements to describe the 3D volumes of objects. Planned works: In this thesis subject, we propose a new approach based on cumulative distribution transform with extraction of features or geometric elements such as volume, regularity, normals and curvatures, etc., in order to feed a convolutional neural network. The cumulative distribution transform is not computantionally expensive and adresses well the problem of high dimensional samples with low sample size. In addition, the proposed characterization should allow deep learning and should be sufficiently generic to process different types of scenes or 3D objects. The problematic of on-line processing will also be explored in order to benefit from recognition techniques in dynamic regime and make the algorithms efficient on mobile robots. The approach should be compatible without difficulty of adaptation with representations by points, by surface elements and also by more regular surfaces as in object modeling (CAD). Starting date 2021-12-01 Funding category Public/private mixed funding Funding further details Plan de relance Presentation of host institution and host laboratory UNIVERSITE DE BOURGOGNE Host laboratory : ImViA (EA 7535) / EMR CNRS 6000 ViBot - Main reasearch topics : vision for robotics This PhD thesis will be co-supervised by Akram Aldroubi, Univetsity of Vanderbilt, Nashville, TN, USA Mobilities in USA will be scheduled within the three years. PhD title Instrumentation and computer image Country where you obtained your PhD France Institution awarding doctoral degree Université de Bourgogne Franche Comté Graduate school SPIM : Sciences pour l'Ingénieur et Microtechniques Double degree Yes Country where the PhD was obtained in cotutelle United States of America Establishment awarding the doctorate in cotutelle Vanderbilt University Candidate's profile strong competences and abilities in mathematics (functional analysis) for computer vision competences in signal and image processing, in vision for robotics, Application deadline 2021-10-25