Remarks on human posture classification using self-organizing map

This paper presents a monocular system to classify human body postures from images. Human silhouette image is extracted from image by background subtraction using statistical background modeling and pixel classification. The feature vector that inputs to the human posture classifier based on self-organizing map (SOM) is composed using both horizontal and vertical projection histograms of the obtained human silhouette and the contour image of the human silhouette. The SOM has torus form so that the boundary of the map does not depend on how to choose the neighborhood area in its learning process. In clustering experiment, a recognition rate of 86.9% is achieved by using the torus-formed SOM (61.5% using conventional SOM, 72.7% using counter propagation network) when testing five postures. Experimental results show both the feasibility and the effectiveness of the proposed method for clustering human body postures.