3D Hand Articulations Tracking Using Gaussian Mixture Model

Three-dimensional tracking of hand articulations is an elementary and critical research subject, which involves image processing, machine vision, computer graphics and other research areas. This research subject has broad application applications, including human computer interface, augmented reality, virtual reality, bionic robot, industrial intelligent production and intelligent medical field etc.The main goal of hand articulations tracking is to estimate of the position and the pose of the hand from video sequences, and then to realize the description and the recognition of the hand motion. Although the research has received extensive attentions and some progress has been made, it still faces many challenges. First, the degree of freedom (DOF) of the hand is high, which is about 26 or 27, therefore it is difficult to calculate the solution of the hand position and pose information. Second, the flexible movement of hand articulations leads to self-occlusion frequently, which destroys the stability and accuracy of the tracking process. Finally, finding optimal solution of the hand position and pose parameters is normally a nonlinear problem, the computational complexity of which is high.This paper utilizes the RGB-D image sequences as the input and studies the critical problems of 3D hand articulations tracking based on Gaussian Mixture Model (GMM) and Adaptive Particle Swarm optimization (APSO) method. Specifically, the main work includes the following aspects:(1) This paper designs a simple hand articulation model and based on GMM to build up the hand model. Furthermore, since the existing hand modeling methods are lack of adaptability, this paper proposes an adaptive hand modeling method based on GMM.(2) In consider of the objective function of existing hand articulations tracking methods have some problem, such as cannot solve the self-occlusion of the hand and high computational complexity of the feature matching, this paper combines the similarity criterion of 2D GMM and the depth feature to build up a normalized objective function. At the same time, utilizing a skin feature penalty item and a smoothness term that penalizes high acceleration in parameter space to enhance the robustness of the algorithm.(3) To deal with the local convergence problem of the gradient descent method and the convergence speed problem of the traditional APSO method in solving multi-peak and multi-parameter function, this paper adopts an improved APSO method and combines it with quasi-random sample method to solve the local convergence problem and improve the convergence speed.