Skip to Main content Skip to Navigation

Système multi-caméras pour l'analyse de la posture humaine

Abstract : Human pose analysis from images is a challenging task due to both the complexity of human body (as a result of the high number of degrees of freedom of the body and the variability of human appearance) and the visual ambiguities inherent to the use of image projection (lack of depth information, self occlusions...). However, the number of potential applications, such as virtual reality, human-computer interaction or athletes' gesture analysis, has intensified the interest for this topic within the computer vision community. This thesis presents a procedure to recover the pose of an articulated body model from the images acquired by several static and calibrated cameras, observing a person moving inside a room. The proposed method does not make any assumption on the knowledge of the previous state estimates in the sequence, and so avoids the problem of initialization and lost tracks. Our goal is to provide a first step to a robust and real-time posture analysis for applications such as scene interpretation and visual surveillance. A background subtraction algorithm is first used to compute binary silhouettes of the body from each camera. A 3D voxel reconstruction of the visual hull is then obtained through a Shape from Silhouettes algorithm. This 3D shape merges all the information about image data and camera calibration, and can make the estimation independent of camera setup given enough viewpoints. We then propose a regression-based estimation : the mapping between the 3D silhouette and the configuration of the human body is learnt during an offline training phase. The learnt model contains a priori information and allows a direct prediction of the pose from the low-level image features (encoded by the visual hull). The cost of the estimation is reduced because the main part of the modelling computation is done during the off-line training stage. Training examples have been synthesized with animation and rendering software. A new 3D shape descriptor has been proposed to encode the 3D shape in a low-dimensional vector and give it as input to the regression process. Various possibilities have been tested concerning body parameterization, configuration of the shape descriptor, regression, camera setup...Throughout this thesis, all the proposed methods are quantitatively evaluated on synthetic data for a ground truth comparison, and qualitatively demonstrated on real sequences of walking and gesture movements.
Document type :
Complete list of metadata

Cited literature [136 references]  Display  Hide  Download
Contributor : Camille Meyer <>
Submitted on : Monday, August 27, 2012 - 4:07:00 PM
Last modification on : Thursday, January 11, 2018 - 6:17:21 AM
Long-term archiving on: : Wednesday, November 28, 2012 - 3:32:53 AM


  • HAL Id : tel-00725684, version 1


Laetitia Gond. Système multi-caméras pour l'analyse de la posture humaine. Automatique / Robotique. Université Blaise Pascal - Clermont-Ferrand II, 2009. Français. ⟨NNT : 2009CLF21922⟩. ⟨tel-00725684⟩



Record views


Files downloads