基于立体视觉的一般物体识别方法  

Generic object recognition method based on stereo vision

作　　者：李新德[1,2] 张晓[1,2] 朱博[1,2] 

机构地区：[1]东南大学复杂工程系统测量与控制教育部重点实验室,南京210096 [2]东南大学自动化学院,南京210096

出　　处：《东南大学学报（自然科学版）》2013年第4期711-716,共6页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(61175091);江苏省自然科学基金资助项目(BK2010403);江苏省"青蓝工程"优秀青年骨干教师计划资助项目;东南大学优秀青年教师教学;科研资助计划资助项目(3208001203)

摘　　要：为了使计算机具有与人类相似的在复杂背景下识别一般物体的视觉处理能力,提出了一种基于立体视觉的一般物体识别方法.该方法的核心在于融合二维图像信息和双目相机获取的深度信息,对视野中的环境进行物体定位、图像分割、特征描述以及物体识别.通过双目相机获取环境的三维点云信息,并利用mean-shift算法进行聚类,剔除干扰点,从而实现物体在二维图像上的定位与分割.利用含有空间关系的BoW模型对分割后独立区域内的物体进行识别,得出判别结果.此外,在利用sift算法进行特征点提取以及利用mean-shift算法进行聚类的环节中,采用CUDA环境下的GPU进行加速处理,提高了处理速度.实验结果表明,所提方法具有较好的识别效果和鲁棒性.In order to make computer have visual processing capability similar to human’s for generic object recognition in complex background,a generic object recognition method based on stereo vision is proposed.The kernel of this method lies in fusing 2D image information and depth information from binocular camera for object localization,image segment,feature description and object recognition in field of vision.3D point cloud information of environment is acquired from stereo camera,and the mean-shift algorithm is used to cluster those points.Object localization and image segment is achieved on 2D image after some interference points are eliminated.In terms of the model of BoW（bag of w ords） w ith spatial relationship,objects in independent area after segmentation are recognized and the discriminated results are obtained.In addition,GPU（graphic processing unit） acceleration technology under CUDA（compute unified device architecture） environment is applied to feature extraction by the sift algorithm and clustering by the mean-shift algorithm in order to greatly improve the processing speed.The experimental results show that this method is robust and the recognition effect is good.

关 键 词：一般物体识别 立体视觉 图像分割 GPU加速 

分 类 号：TP24[自动化与计算机技术—检测技术与自动化装置]