基于聚类和shape context的人体姿态估计

Shape Context and Clustering for Human Pose Estimation

二维图片的人体姿态估计是计算机视觉中一个非常重要并且热门的研究课题,并被广泛应用于人机交互、监控以及图片检索等方面.基于部件的模型是解决这一问题的经典方法,但当人体姿态变化较大时,传统部件模型不能精确地刻画和表达这种形变,因此使用范围受到很大限制.为了克服这一缺点,采用分治的思想细分了人体姿态的类型,并用shape context改进了原始部件模型中的父子部件相对位置的描述方式.首先在训练集上基于shape context特征对人体姿态进行聚类,然后在每一个聚类结果上训练一个基于部件的姿态估计模型.用HOG特征刻画部件的外观,用shape context特征来刻画部件之间的联系.相比传统的基于两个部件之间位置关系的刻画方式,考虑到全局信息的shape context特征可以弥补树模型结构过于单一的问题.这种改进的空间位置约束方式使得部件模型的整体性能得到提高.结果表明,本文提出的方法在实验中取得了令人满意的结果,与其它若干姿态估计方法相比具有一定优势.

Pose estimation is an extremely important and active research field in computer vision,due to its wide range of applications,e. g. human-computer interaction,surveillance,image retrieval,etc. Part-based model is a classical method,but when pose changes widely,traditional part-based model can not deal with it. To overcome this disadvantage,we use the thought of divide and conquer to partition human pose. Specifically,we cluster different kinds of human poses in training dataset and train a pose detector in each clustering set,respectively. In this paper,we use HOG to represent the appearance of parts and shape context descriptor to represent the relationship of parts. Shape context can remedy for the singleness of tree model. Our improvement on space constraint makes part-based model achieve a higher performance. We contrast our method with the proposed methods and the experiment shows that the given method prevails and obtains a satisfactory result.