3D Hand Tracking From RGB Sequences

Abstract

Vision-based hand tracking has been an active field of research since the early 1990s. Early attempts tended to use generative kinematic models, in which a hand state proposal is quantitatively evaluated according to features extracted from an input image. The approximate hand pose is then found by optimising the pose of the kinematic model according to those features. This paradigm continued to be used in both RGB-based tracking and in later depth-based tracking approaches. More recent attempts have made use of convolutional neural networks (CNN) to predict keypoint locations discriminatively. Here, a contemporary CNN approach is applied to the conventional generative hand tracking paradigm. This is done by using CNNs to semantically segment each frame of an RGB sequence containing hand gestures before using a generative kinematic model to find the optimal hand pose for each frame given the semantic segmentation result.

Date of Award	1 Aug 2020
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Aphrodite Galata (Supervisor)