Human motion segmentation using active shape models

Abstract

Human motion analysis in images is thoroughly related with the developmentof computational techniques capable of automatically identify, track andanalyze relevant structures of the body. In fact, in any system designed for humanmotion analysis from image sequences, the first processing step concerns the identificationof the structures to be analyzed in each of the sequence images, beingthis step commonly referred as image segmentation. Here, a widely used database,the CASIA Gait Database, is used to build Point Distribution Models (PDMs) ofthe human silhouette, including specific joints. The training image dataset used includes14 subjects walking in four different directions, and each shape of the trainingset was represented by a set of labeled landmark points. The contours of thesilhouettes were obtained with the purpose of automatically extract 100 silhouettepoints together with additional 13 anatomic joint points, such as elbows, knees andfeet, to be used as landmarks. In order to obtain the mean shape of the silhouetteas well as its admissible shape variations PDMs for each direction were built. ThePDMs built were finally used in the construction of Active Shape Models (ASMs),which combine the shape model with grey level profiles, with the purpose of furthersegment the modeled silhouettes in new images. The referred technique is aniterative optimization scheme for PDMs allowing initial estimates of pose, scaleand shape of an object to be refined in a new image. The experiments conductedusing this segmentation technique has revealed very encouraging results.