Human Body Measurement
I developed a Kinect project to measure the circumference of human body parts such as chest, waist and arm. To make the measurement, I separated the human body from the Kinect range image and fit ellipses to the cross section of body parts given the skeleton.
However, the invalid data on the edge of human body in the range image caused error in measurement. In order to classify the invalid data, I trained mixture Gaussian distributions for both human body and background on the corresponding raw RGB image using Expectation-maximization (EM) algorithm. Though it worked well, it had problems in the dynamic background because of the unpredictable number of Gaussian distributions. So instead of EM algorithm, I adopted the Min-cut algorithm to solve the classification problem. This global optimization approach worked well in both static and dynamic background.