PT Unknown
AU Daniel Sanchez
   J.C.Ortega
   Miguel Angel Bautista
TI Human Body Segmentation with Multi-limb Error-Correcting Output Codes Detection and Graph Cuts Optimization
BT 6th Iberian Conference on Pattern Recognition and Image Analysis
PY 2013
BP 50
EP 58
VL 7887
DI 10.1007/978-3-642-38628-2_6
DE Human Body Segmentation; Error-Correcting Output Codes; Cascade of Classifiers; Graph Cuts
AB Human body segmentation is a hard task because of the high variability in appearance produced by changes in the point of view, lighting conditions, and number of articulations of the human body. In this paper, we propose a two-stage approach for the segmentation of the human body. In a first step, a set of human limbs are described, normalized to be rotation invariant, and trained using cascade of classifiers to be split in a tree structure way. Once the tree structure is trained, it is included in a ternary Error-Correcting Output Codes (ECOC) framework. This first classification step is applied in a windowing way on a new test image, defining a body-like probability map, which is used as an initialization of a GMM color modelling and binary Graph Cuts optimization procedure. The proposed methodology is tested in a novel limb-labelled data set. Results show performance improvements of the novel approach in comparison to classical cascade of classifiers and human detector-based Graph Cuts segmentation approaches.
ER