Mutual Information of Image Intensity and Gradient Flux for Markerless Pose Estimation

In this work we present a technique for estimating the pose of an articulated object with only a simple shape model and no appearance model. We use a generative technique to propose, for a pair of stereo cameras, hypothesised correspondences for object pixels in two images. The feature space for each camera is a set of two-dimensional points, where each point consists of an intensity value and a flux value for a given pixel. The flux measures the volume of gradient flow at a pixel. For any pose we compute the mutual information between the feature set from one view and the feature set from another. The mutual information measures the degree to which hypothesised correspondences are supported by the data. Computing the mutual information of points containing both the image flux and intensity is more accurate than using intensity or flux alone. We demonstrate the results of a short tracking sequence using particle swarm optimisation to estimate the pose in every frame.