VISUAL TRACKING AND RECOGNITION OF THE HUMAN HAND

Most of the computer-vision-based hand gesture
recognition systems are either confined to a fixed
set of static gestures or only able to track 2D
global hand motion.
In order to recognize natural hand gestures such as
those in American sign language, we need to track
articulated hand motion in real time. The task is
challenging due to the high degrees of freedom of the
hand, self-occlusion, variable views, and lighting.
This book focuses on automatic recovery of 3D hand
motion from one or more views.
The problem of hand tracking is formulated as
Bayesian filtering in the framework of
analysis-by-synthesis. We propose an Eigen Dynamic
Analysis model and a new feature called likelihood edge.
To automatically initialize and recover from
loss-track, we proposed a bottom-up posture
recognition algorithm. It collectively matches the
local features in a single image with those in the
image database. Through quantitative and visual
experimental results, we demonstrate the
effectiveness of our approach and point out its
limitations.

Autorentext

Hanning Zhou is a senior manager in Amazon Media Technology Group.Prior to Amazon, he worked on video surveillance in FX Palo AltoLab (FXPAL) as a research scientist.As an undergrad from Tsinghua Univ., he worked with Harry Shum inMS Research Asia on video compression.He received PhD from Univ. of Illinois Urbana Champaign and held12 patents.

Klappentext

Most of the computer-vision-based hand gesturerecognition systems are either confined to a fixedset of static gestures or only able to track 2Dglobal hand motion. In order to recognize natural hand gestures such asthose in American sign language, we need to trackarticulated hand motion in real time. The task ischallenging due to the high degrees of freedom of thehand, self-occlusion, variable views, and lighting. This book focuses on automatic recovery of 3D handmotion from one or more views.The problem of hand tracking is formulated asBayesian filtering in the framework ofanalysis-by-synthesis. We propose an Eigen DynamicAnalysis model and a new feature called likelihood edge. To automatically initialize and recover fromloss-track, we proposed a bottom-up posturerecognition algorithm. It collectively matches thelocal features in a single image with those in theimage database. Through quantitative and visualexperimental results, we demonstrate theeffectiveness of our approach and point out itslimitations.