Three dimensional target object extraction based on binocular stereo vision

This paper proposes a scheme for target object extraction in a point cloud with noise and measurement errors. We divide the scheme into two parts: rough extraction and surface filtering of the target object. The first part consists of plane segment and clustering algorithm. The plane segment is implemented by random sample consensus (RANSAC) algorithm to fit the plane model. The clustering problem is solved by Euclidean cluster extracting algorithm. And we choose the object which is the closest to the center point of the point cloud as the target object. The rough extraction part contains normals estimation, clustering algorithm and surface filtering. The problem of normals estimation is solved by principal component analysis (PCA). After estimating normal for each point, we adopt Kmeans clustering method to solve the plane segment problem. Then we project the points which are influenced by noise to the plane. After all of these, a target object can be extracted. The whole scheme is implemented based on point cloud library (PCL) which is a standalone, large scale, open project for 2D/3D image and point cloud processing. The result shows that the method can be used to effectively reduce the influence of noise and measurement error, and obtain a target object with a smooth surface. That is helpful to provide more accurate point cloud information for other point cloud processing.