Open Hand Detection in a Cluttered Single Image Using Finger Primitives

Abstract

Hand Detection plays an important role in human computer interaction (HCI) applications, as well as surveillance. We propose a hand detection technique that is robust to different skin color, illumination and shadow irregularities by exploiting the geometric properties of the hand. We first obtain the responses from two detectors that operate independently on the test image to identify parallel finger edges and curved fingertips. These responses are then grouped by using two decision trees trained on each primitive class, yielding two separate collections of groups. The final merging algorithm returns candidate hands in a given single image by comparing groups across each collection and merging those that satisfy a scoring function. The proposed system is robust to the size and the orientation of the hand, with the single requirement that one or more fingers are visible. The system is the first to successfully detect hands in an uncontrolled environment, without training on the skin color within a single image or using motion information.