Gesture recognition in realistic images: the statistical approach

The paper presents a robust gesture segmentation and recognition scheme in real images using statistical pattern recognition techniques, like data clustering and linear regression. Specifically, a hierarchical clustering algorithm is adopted because it does not require the exact number of sought clusters. Thus the proposed gesture recognition scheme is capable of coping with gestures having a variable number of extended fingers, a common situation in many practical applications like the expanded user-machine interface and the automatic deaf-mute sign language translation. For the mathematical modeling of clusters, a linear regression scheme is used. While in other cases linear regression is a simplification made for time saving, in this case it also ensures representation accuracy due to the geometry of the human hand being mostly composed of linear segments. Statistical linear modeling enables the handling of points with extreme values in comparison to the rest (outliers). As a result, the suggested algorithm is not affected by pixels that have been mistakenly selected by the image processing algorithms.