Gesture recognition is a user-friendly and intuitive means for man-machine-communication.
In this thesis, a view-invariant hand-posture recognition method for Soft Remocon system is presented. Soft Remocon system, developed by the research center of KAIST as a part of an Intelligent Sweet Home, manages home appliances by the user’s hand gesture. Now only pointing hand gesture is available. The system recognizes hand position and pointing direction from the images of 3 color cameras equipped on ceiling. By considering hand-posture in Soft Remocon system, it can achieve extendibility in the sense of command set and more convenience of the user.
Discussion in this paper will center on the hand-posture recognition.
To apply hand-posture recognition to Soft Remocon, we should consider the view-dependent characteristic of silhouette images. Although there are a lot of researches on hand-posture recognition, view-invariant hand posture recognition is still remaining. Many of them assume constraint in viewpoint. It definitely limits user’s activity. Moreover, human hand is highly articulated and deformable. With these reasons, hand posture recognition is a challenging example in the research of view-invariant object recognition.
In the methodological view, hand-posture recognition problem is divided into 1) description problem 2) matching problem. To describe hand-postures, the proposed method uses ‘Object Shape + Structure Network (OSS-Net)’ model. In this model, each hand-posture is described by ‘characteristic nodes’ and ‘links’ between nodes. Characteristic nodes are feature vectors of each hand-posture. Under this model, matching problem to find the most similar hand-posture leads to the problem of searching the proper path from input characteristic node toward target node. A link between nodes is defined by the probability of activation. This concept substitutes conventional similarity generally used in the object recognition problem. The link is defined by means of reinforcement learning.
The proposed hand-posture recognition system is robust to variations in view-point. It is validated through various experiments.
