Estimation de la posture humaine par capteur RGB-D

Lucas Marti 1
1 LAAS-RAP - Équipe Robotique, Action et Perception
LAAS - Laboratoire d'analyse et d'architecture des systèmes [Toulouse]
Abstract : In a world where the population is increasingly aging, elderly people falling is a public health issue. The use of technology is a major development axis for fall management. We want to design a complete system for detection and estimation of falls. The target market for this system is medicalized retirement homes and individual homes. We want to improve the medical care of people after a fall in order to reduce its consequences. A solution based on an ambient sensor seems to be the most adapted solution. Recent developments in RGB-D (Color+Depth) sensing are a great asset thanks to their relatively low cost, wide availability and good performances. The first part of the thesis deals with the problem of segmenting people from the surrounding scene in our images. We present an algorithm that determines the silhouettes of each person in the room in which the sensor is installed, thanks to simultaneous use of color and depth. The algorithm is robust to the change of configuration of the room and especially to moving furniture. We use special consideration of depth to reach a performance level sufficient for an industrial application. The second part of the thesis deals with the estimation of the human posture. Once the silhouettes have been segmented with the algorithm described in the first part, we want to get an estimation of every articulation of the person. We build on existing algorithms that use machine learning and in particular Random Forest by investigating new ideas to improve performances. We found optimal values for some parameters that were not previously investigated. We present a new feature to be computed on depth images. Finally we evaluate the impact of balancing the training database in our context. The algorithm provides a set of predictions for the position of every articulation. In the third part, we focus on spatio-temporal filtering of the postures. We examine different approaches and in particular we deal with the issue of left/right ambiguity that arises in the algorithm presented in the previous part. The approaches are based on Bayesian filtering.
Document type :
Robotique [cs.RO]. Universite Toulouse III Paul Sabatier, 2015. Français
Liste complète des métadonnées

Cited literature [53 references]  Display  Hide  Download
Contributor : Arlette Evrard <>
Submitted on : Monday, November 7, 2016 - 2:15:29 PM
Last modification on : Thursday, January 11, 2018 - 6:26:29 AM
Document(s) archivé(s) le : Wednesday, February 8, 2017 - 2:22:26 PM


  • HAL Id : tel-01393419, version 1


Lucas Marti. Estimation de la posture humaine par capteur RGB-D. Robotique [cs.RO]. Universite Toulouse III Paul Sabatier, 2015. Français. 〈tel-01393419〉



Record views


Files downloads