Extraction of Human Limb Regions and Parameter Estimation using Optical Flow against Complex Background

Abstract

We propose a method for extracting rotating limb regions and estimating their motion parameters, which is based on curl of optical flow. Regarding optical flow under orthographic projection as a vector field, the proposed method computes the curl of the optical flow to extract rotating regions not depending on the position of the limb. We use the EM algorithm which can perform both of dividing the image into multiple moving regions and estimating motion parameters of each region. Since the proposed method doesn't require a simplicity of background, and can discriminate translationa movements produced by other moving objects in the background, it is applicable to a practical situation in which conventional methods don't work well. Experimental results with real image sequences demonstrate that the limb extraction can be performed well from noisy optical flow which are obtained from real images.