Nuisance Free Recognition of Hand Postures over a Tabletop Display

This paper proposes a new approach to shape classication that is well suited to the speci c challenges of vision-based hand posture recognition in a multi-user tabletop collaboration scenario. We use a representation of the 2-D hand silhouette where in-plane rotation and mirror symmetry appear as particular cases of permutations, and then show how to take advantage of this pattern to develop an e cient version of the permutation invariant SVM. Invariance to these transformations is very important because the users stand around the table, and a video camera captures the scene from the top. We also report experimental results that compare this approach favorably over common classi cation approaches, under the stated requirements.