Skip to Main content Skip to Navigation

Motion Analysis for Dynamic 3D Scene Reconstruction and Understanding

Abstract : This thesis studies the problem of dynamic scene 3D reconstruction and understanding using a calibrated 2D-3D camera setup mounted on a mobile platform via the analysis of objects' motions. For static scenes, the sought 3D map reconstruction can be obtained by registering the point cloud sequence. However, with dynamic scenes, we require a prior step of moving object elimination, which yields to the motion detection and segmentation problems. We provide solutions for the two practical scenarios, namely the known and unknown camera motion cases, respectively. When camera motion is unknown, our 3D-SSC and 3D-SMR algorithms segment the moving objects by analysing their 3D feature trajectories. In contrast, by compensating the known camera motion, our 3D Flow Field Analysis algorithm inspects the spatio-temporal property of the object's motion. By removing the dynamic objects, we attain the high quality 3D background and multi-body reconstruction by using our DW-ICP point cloud registration algorithm. In the context of scene understanding, semantic object information is learned from images and transferred to the reconstructed static map via our 2D-to-3D label transfer scheme. All the proposed algorithms have been quantitatively and qualitatively evaluated and validated by using extensive experiments of real outdoor scenes.
Complete list of metadatas
Contributor : Cansen Jiang <>
Submitted on : Friday, March 16, 2018 - 6:38:42 PM
Last modification on : Monday, March 30, 2020 - 8:55:02 AM
Document(s) archivé(s) le : Tuesday, September 11, 2018 - 5:53:01 AM


Files produced by the author(s)


  • HAL Id : tel-01736353, version 1


Cansen Jiang. Motion Analysis for Dynamic 3D Scene Reconstruction and Understanding. Computer Vision and Pattern Recognition [cs.CV]. Universite de Bourgogne, 2017. English. ⟨tel-01736353⟩



Record views


Files downloads