Segmentação de pessoas em imagens estáticas baseada em esqueleto

AUTOR(ES)
FONTE

IBICT - Instituto Brasileiro de Informação em Ciência e Tecnologia

DATA DE PUBLICAÇÃO

20/04/2012

RESUMO

The segmentation of people (automatic or semi-automatic) in still images is a very challenging task, mainly due to several factors in the real world, such as those related to the lighting of the scene where the image was captured, shadows, image noise, occlusions, high similarity of the object of interest with the background of the scene and the lack of information inherent in depth when a scene is captured into a 2D image. In this work we present a skeleton-based model for human segmentation in still images. The input data of the model, related to the skeleton model, can be obtained automatically (using an algorithm for 2D pose estimation of people in images, for example) or manually (through user interaction), depending on the particular application. The skeleton model is used to guide the segmentation by taking into account color information, brightness, angle constraints and anthropometric parameters. In a general way, the main idea of the proposed approach is to build a graph around the skeleton model, for a given input image, and find out the best path in this graph that satisfies a certain condition (e.g., the one that maximizes a certain energy criterion), thus generating the contour of the person in the picture.It is also being proposed in this work an approach to measure quantitatively the experimental results, from information provided through user interaction. The experimental results demonstrate that the proposed model generates satisfactory results for non-trivial images containing people with varied appearances and poses (containing self-occlusions), in various complex environments (and uncontrolled), with different lighting conditions and image quality. The results obtained using the proposed model was also compared with those obtained by a work considered state of the art. Our experiments indicate that the proposed model adapts better to the contours, while the human body shape priors in the confronted work enforce a smoother contour.The proposed segmentation model generates a closed contour (for each person in the image) with semantic information included, (e.g., each contour point is associated with a particular body part), which can be used for various purposes (for example, construction of virtual humans with features extracted from the image, methods for clothes estimation in images, estimation of the human shape under the clothes, etc.).

ASSUNTO(S)

informÁtica processamento de imagens semÂntica esqueleto ciencia da computacao

ACESSO AO ARTIGO

Documentos Relacionados