基于随机丛林的人体部件分类方法  

Body Part Classification Method Based on Random Jungle

作　　者：屈雁秋 陈锋[1] 

机构地区：[1]中国科学技术大学自动化系,合肥230000

出　　处：《计算机工程》2018年第1期247-251,共5页Computer Engineering

摘　　要：人体部件分类是人体姿态跟踪中的重要前提和关键步骤。传统深度梯度特征下的随机森林分类方法虽然可以达到实时性的要求,但仍存在准确度不高、对噪声不够鲁棒、内存消耗过大等缺点。为此,提出传统深度特征与RGB边缘特征相融合的一种新的分类方法,并在特征维度加大的情况下引入随机丛林模型。实验结果表明,该特征分类方法不仅可以减少20%左右的运行时间,还可以提高1%左右的测试准确率。Body part classification is an important precondition and crucial step in human pose tracking. Although the random forest classical method with gradient feature of depth can get the request of real time, it remains the weakness of low accuracy,not robust enough to the noise and huge memory consumption. This paper proposes a new classification method,which combines the classical depth feature with RGB edge feature, and introduces the random jungle model to solve the increased dimension of the feature. Experimental results show that the proposed method not only saves about 20% running time,but also improves the test accuracy of about 1% .

关 键 词：人体部件分类 深度梯度特征 RGB边缘特征 随机丛林 有向无环图 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]