Real-Time Hand Detection From a Single Depth Image by Per-Pixel Classification

Abstract

Due to their convenience and naturalness, hand pose recogni-tion or gesture recognition methods are gaining attention as an upcom-ing complement of traditional input devices such as keyboards, mice, joysticks, etc. Robust hand detection from an image is the first stage to solve the hand gesture recognition. Due to the release of the commer-cial depth camera, elimination of the cluttered background from a depth image is much easier than from a RGB image. However, accurate hand segmentation from a human body still remains in challenging task. Here, we propose robust real-time hand detection algorithm from a depth im-age. The algorithm is designed to detect hands with various hand poses in various positions in 3D space. We train Radom Decision Forests to every pixel in the image to detect hand. The pixel in the image has one of the two label, hand or non-hand. We optimize the random decision forests parameters by various experimental conditions. The result shows that the per-pixel classification accuracy is 94% and the RDF with 5 trees requires only 12ms with no help of parallel programming.