Markerless human pose estimation using image features and extremal contour

This paper presents an approach for markerless vision-based motion capture from multiple views. We use truncated cones to describe the human body parts, and match the extremal contours of human body model against the image cues. Because the derivative is not available, we assume that the cost function satisfies a quadratic model inside the trust region and use model-based Derivative Free Optimization (DFO) method to find a pose which best matches the images. The method performance was tested on the HumanEva II dataset in a 4 color camera configuration and the results show that our method recover the human pose parameterize with high dimensions (≥ 36) effectively.