Human Shape Recognition Using Region Based Shape Description and Mean Shift Clustering

Abstract

This paper proposes a human shape recognition algorithm using the region-based shape descriptor and mean shift clustering. The main goal of paper is to identify that the object regions extracted in the video are human or not for video surveillance systems. The angular radial transform (ART), a region-based shape descriptorin MPEG-7, is applied to model the human shapes. We construct database images of human shapes, and exploit 3 radial and 12 angular frequencies for modeling human shapes. The 36-D ART vectors for human shapes are first clustered using mean shift clustering, and several representative ART vectors are modeled by mean vectors of clusters. The human objects are identified by distances between the representative vectors and ART vector of extracted objectregion. The distance threshold for each cluster is statistically obtained in the learning step. This paper also deals with smoothing object boundaries extracted by background subtraction, which improves the recognition rates. The ART vectors for human shapes are learned using thousands of illustration images and real objects extracted by background subtraction. Experiments are performed on various images such as MPEGCE-2-Bdataset, illustration of human and non-human objects, and video framescombined with background subtraction. The experimental results show that the proposed algorithm is robust and efficient in human. object recognition.