CAJ | 학술논문

现有的物体识别方法大都是有监督学习,而且主要处理的是2D图像数据,由于这类数据给出的背景信息和物体信息少于2.5D场景中的信息,传统有监督的学习方法对于非特定物体的识别效果还是不尽如人意,更别说基于无监督学习的识别系统了.随着Kinect设别的出现,获取场景的2.5D信息变得很容易.设计了一个全新的基于无监督物体识别系统.首先用平面拟合和空间聚合把物体从场景中分离以后.然后用几何基元拟合分割技术把物体分割为不同部件,在此基础上构建物体的图模型,最后把图模型嵌入为向量空间上的点集合,使用扩展陆地移动距离算法计算物体间的相似度.从实验结果来看效果不错.
현유적물체식별방법대도시유감독학습,이차주요처리적시2D도상수거,유우저류수거급출적배경신식화물체신식소우2.5D장경중적신식,전통유감독적학습방법대우비특정물체적식별효과환시불진여인의,경별설기우무감독학습적식별계통료.수착Kinect설별적출현,획취장경적2.5D신식변득흔용역.설계료일개전신적기우무감독물체식별계통.수선용평면의합화공간취합파물체종장경중분리이후.연후용궤하기원의합분할기술파물체분할위불동부건,재차기출상구건물체적도모형,최후파도모형감입위향량공간상적점집합,사용확전륙지이동거리산법계산물체간적상사도.종실험결과래간효과불착.
Most object recognition systems are based on supervised learning, and dealing with 2D image which gives little infor?mation about the background and foreground. Even state of art supervised object recognition can’t promise good result, let alone unsupervised based system. With the advent of Kinect, it’s more easy to get 2.5D data. This paper proposes an innovative unsu?pervised based object recognition system. The first step is to fit plane and use space based clustering to extract objects from the scene. The second step is segmenting objects based on geometry primitive fitting, constructing graph model for each object based on the segmentation result. Finally, embed the graph model to normal vector space and calculate the similarity between objects with extended Earth Mover’s Distance. The experiment gives a good result.