Hand gesture estimation and model refinement using monocular camera-ambiguity limitation by inequality constraints

The paper proposes a method to precisely estimate the pose (joint angles) of a moving human hand and also refine the 3D shape (widths and lengths) of the given hand model from a monocular image sequence which contains no depth data. First, given an initial rough shaped 3D model, possible pose candidates are generated in a search space efficiently reduced using silhouette features and motion prediction. Then, selecting the candidates with high posterior probabilities, the rough poses are obtained and the feature correspondence is resolved even under quick motion and self occlusion. Next, in order to refine both the 3D shape model and the rough pose under the depth ambiguity in monocular images, the paper proposes an ambiguity limitation method by loose constraint knowledge of the object represented as inequalities. The method calculates the probability distribution satisfying both the observation and the constraints. When multiple solutions are possible, they are preserved until a unique solution is determined. Experimental results show that the depth ambiguity is incrementally reduced if the informative observations are obtained.