Estimación de la pose humana 2d en imágenes estéreo

Laburpena

Human Pose Estimation (HPE) is the task of obtaining the spatial con guration of human body parts from images. Methods recovering the human pose from a single image are called monocular approaches while those using image pairs are called stereo approaches. Stereo images provide extra information that can be employed to improve the results obtained by monocular approaches. This Thesis considers the problem of 2D human pose estimation on stereo images. To this end, three contributions are provided. The rst contribution of this thesis is a new technique to automatically detect and estimate the 2D pose of humans in stereo images. The proposed method is based on a similarity constraint that promotes a collaboration between two pose estimators. We show experimentally that our proposal improves the accuracy of the estimated poses when compared to standard HPE techniques running independently on each image. The second contribution is a dataset for the problem of human pose estimation in stereo image. To experimentally validate our approach, we have created a new annotated dataset of 630 stereo image from stereo videos depicting people in di erent backgrounds, clothing, lighting or locations in the image frames. The dataset contains upright people in a great variety of arms poses, covering the space of possible con gurations quite uniformly. The third contribution is a new method to estimate the 2D pose of humans in stereo videos sequences. The proposed pipeline starts by constraining the possible location of body joints by exploiting color and disparity information, and adding location priors to the most structured joints. Finally, a body limb recombination method is applied along the stereo sequence to obtain the best con guration of the body joints. The experiments show that our method obtains better average results than the state-of-the-art.