基于SVDD的三维目标多视点视图建模  

Method of multi-view modeling for 3D target based on SVDD

作　　者：丁昊[1] 李旭东[1] 赵慧洁[1] 

机构地区：[1]北京航空航天大学精密光机电一体化技术教育部重点实验室,北京100191

出　　处：《北京航空航天大学学报》2012年第11期1517-1521,共5页Journal of Beijing University of Aeronautics and Astronautics

基　　金：国家自然科学基金资助项目(60802044)

摘　　要：同一目标在不同观察视点下成像后外形可能有较大差异,因此三维目标多视点视图建模是目标识别的关键.针对该问题,提出了基于支持向量数据描述(SVDD,SupportVector Data Description)方法对目标特征进行描述.在视点球面上均匀采样获取目标全姿态图像,以SVDD方法求取在高维空间内包含尽可能多目标特征向量的最小超球体相关参数,得到数量较少的支持向量将作为目标多视点视图的最佳模型.对多类目标不同姿态的图像(每类2592帧),以规则化不变矩描述目标外形特征,进行了建模实验,并通过识别实验验证了所提方法的有效性和可行性.Popular 3D target recognition approaches based on image continue to struggle with challenge viewpoint sensitivity. Multi-view modeling technique for 3D target offers promise for this challenge. A method of modeling 3D target based on support vector data description （SVDD） that could obtain a tight description covering most of the target feature data was proposed. Target＇ s images were captured with uniform grid on the viewing sphere and characterized by feature vectors. The support vectors representing characteristic views were obtained by applying SVDD to optimize the parameters of the minimal hyper-sphere which covers as many fea ture vectors as possible. The experiments were conducted by applying the proposed method to an image set （ each target includes 2592 images） characterized by normalized moment invariants. The results show that the proposed method is effective and feasibility.

关 键 词：多视点建模 支持向量数据描述 三维目标识别 规则化不变矩 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]