Skip to Main content Skip to Navigation

Vers un système de vision artificielle opportuniste pour l'analyse de scènes complexes à partir de caméras embarquées

Abstract : The thesis intends to develop the bricks of an opportunistic vision system for dynamic scene analysis, an opportunistic system that would be guided by the applicative task, that would benefit from any knowledge and prioris made available by the application, and take profit of all available cues (color, texture, geometry) depending on their quality and relevance. The context of color monocular vision is considered, with a camera embedded on a mobile platform. A dense optical flow technique is first proposed. After a rough estimation, a reliability map is computed and is used for refining the motion map, through an iterative propagation process constrained by local information, starting by the color cues. This motion map is then analyzed for rough and fast plane segmentation. A cumulative approach called uv-velocity has been developed. It allows the fast exhibition of prominent planar surfaces under certain assumption related the egomotion. Contrary to its predecessor, the socalled c-velocity, it allows a more progressive voting strategy, it avoids using sampling, it is not limited to translations of the camera and can detect a wider range of surfaces.. The motion models related to each surface can then be re-injected as a constraint in the estimation of the next optical flow. The raw and fast planar segmentation produced by uvvelocity can be used to fasten the estimation visual odometry. The results of optical flow estimation remain acceptable in terms of precision and execution time (tested on Middleburry dataset) which can be the input for creating the voting space to detect the planes on image. After the simulations and real experiments on KITTI dataset, uvvelocity shows its potential to be the polyvalent image registration on plane detection and opportunistic alert for the system.
Document type :
Complete list of metadata

Cited literature [142 references]  Display  Hide  Download
Contributor : Frédéric Davesne Connect in order to contact the contributor
Submitted on : Tuesday, March 19, 2019 - 10:56:45 AM
Last modification on : Saturday, May 1, 2021 - 3:51:16 AM
Long-term archiving on: : Thursday, June 20, 2019 - 1:34:29 PM


Files produced by the author(s)


  • HAL Id : tel-02072438, version 1


Tan Khoa Mai. Vers un système de vision artificielle opportuniste pour l'analyse de scènes complexes à partir de caméras embarquées. Traitement du signal et de l'image [eess.SP]. Université Paris-Saclay; Université d'Evry-Val d'Essonne, 2018. Français. ⟨NNT : 2018SACLE045⟩. ⟨tel-02072438⟩



Record views


Files downloads