Real time hand pose estimation using depth sensors
Real time hand pose estimation using depth sensors
复制标题
DOI：
10.1007/978-1-4471-4640-7_7
复制
发表时间：
2011-11
期刊：
影响因子：
--
通讯作者：
Cem Keskin;Mustafa Furkan Kıraç;Yunus Emre Kara;L. Akarun
中科院分区：
文献类型：
--
作者：
Cem Keskin;Mustafa Furkan Kıraç;Yunus Emre Kara;L. Akarun
实时手部姿态捕捉一直是计算机视觉领域的一个难点。手部骨骼参数的提取将是手语识别的一个重要里程碑，因为它将使手形和手势分类成为可能。最近Kinect深度传感器的引入加速了人体姿势捕捉的研究。本章介绍了一种实时的手姿态估计方法，采用对象识别的部分方法，并使用这种方法的手形分类。首先，一个逼真的3D手模型被用来表示手与21个不同的部分。然后，随机决策森林（RDF）在通过使手模型动画化而生成的合成深度图像上进行训练，该合成深度图像用于执行每像素分类并将每个像素分配给手部分。分类结果被馈送到局部模式查找算法中以估计手骨架的关节位置。该系统可以真实的处理从Kinect检索的深度图像，并且不依赖于时间信息。作为该系统的一个简单应用，我们还描述了一个基于支持向量机（SVM）的美国手语（ASL）十位数字识别模块，该模块在真实的实时深度图像上达到了99.9%的识别率。
Real-time hand posture capture has been a difficult goal in computer vision. The extraction of hand skeleton parameters would be an important milestone for sign language recognition, since it would make classification of hand shapes and gestures possible. The recent introduction of the Kinect depth sensor has accelerated research in human body pose capture. This chapter describes a real-time hand pose estimation method employing an object recognition by parts approach, and the use of this method for hand shape classification. First, a realistic 3D hand model is used to represent the hand with 21 different parts. Then, a random decision forest (RDF) is trained on synthetic depth images generated by animating the hand model, which is used to perform per pixel classification and to assign each pixel to a hand part. The classification results are fed into a local mode finding algorithm to estimate the joint locations for the hand skeleton. The system can process depth images retrieved from Kinect in real time, and does not rely on temporal information. As a simple application of the system, we also describe a support vector machine (SVM)-based recognition module for the ten digits of American Sign Language (ASL) based on our method, which attains a recognition rate of 99.9 % on live depth images in real time.