Skip to Main content Skip to Navigation

Human Posture Recognition for Behaviour

Abstract : During this thesis, we have proposed a real-time, generic, and operational approach to recognising human posture with one static camera. The approach is fully automatic and independent from the view point of the camera. Human posture recognition from a video sequence is a difficult task. This task is part of the more general problem of video sequence interpretation. The proposed approach takes as input information provided by vision algorithms such as the silhouette of the observed person (a binary image representing the person and the background), or her/his position in the scene. The first contribution is the modeling of a 3D posture avatar. This avatar is composed of a human model (defining the relations between the different body parts), a set of parameters (defining the position of the body parts) and a set of body primitives (defining the visual aspect of the body parts). The second contribution is the proposed hybrid approach to recognise human posture. This approach combines the use of 3D posture avatar and 2D techniques. The 3D avatars are used in the recognition process to acquire a certain independence from the camera view point. The 2D techniques represent the silhouettes of the observed person to provide a real-time processing. The proposed approach is composed of two main parts: the posture detection which recognises the posture of the detected person by using information computed on the studied frame, and the posture temporal filtering which filters the posture by using information about the posture of the person on the previous frames A third contribution is the comparison of different 2D silhouette representations. The comparison is made in terms of computation time and dependence on the silhouette quality. Four representations have been chosen: geometric features, Hu moments, skeletonisation, and the horizontal and vertical projections. A fourth contribution is the characterisation of ambiguous postures. Ambiguities can happen by using only one camera. An ambiguous posture is defined as a posture which has visually similar silhouettes rather an other posture. Synthetic data are generated to evaluate the proposed approach for different point of view. The approach has also been evaluated on real data by proposing a ground truth model adapted to the posture recognition purpose. A fifth contribution has been proposed by applying the results of the recognition to human action detection. A method based on a finite state machine has been proposed to recognise self-action (action where only one person acts). Each state v of the machine is composed of one or several postures. This method has been successfully applied to detect falling and walking actions. The human posture recognition approach gives good results. However, the approach has some limitation. The main limitation, is that we are limited in terms of postures of interest for computation time and discrimination reasons. The second limitation is the computation time of the 3D posture avatar generation. By using information about the movement of the observed person in the scene, the approach is able to treat 5-6 frames by second. Some improvement can be done to solve these limitations. In particular, the set of interest postures can be adapted automatically at each frame by considering the previously recognised postures to decrease the number of 3D posture silhouette to extract.
Complete list of metadatas
Contributor : Estelle Nivault <>
Submitted on : Wednesday, August 20, 2008 - 4:23:36 PM
Last modification on : Saturday, January 27, 2018 - 1:30:44 AM
Document(s) archivé(s) le : Thursday, June 3, 2010 - 6:38:52 PM


  • HAL Id : tel-00311741, version 1



Bernard Boulay. Human Posture Recognition for Behaviour. Other [cs.OH]. Université Nice Sophia Antipolis, 2007. English. ⟨tel-00311741⟩



Record views


Files downloads