RGB-D Based Object Localization And Recognition

Object Localization and Recognition are the basis for service robots achieving intelligent job, as well as the hotspots in machine vision. After decades of research, lots of effective methods have been proposed. But there is still much room for improvement in perception and cognition based on3D data.This thesis focuses on the problem of object localization and recognition for the intelligent service robot ZJU-Cyber of ZheJiang University.The main contributions are as follows:1. RGB-D based Object localization approach is improved by regarding object localization problem as a segmentation problem of pointcloud. Fristly a RANSAC based plane detection method is designed, then RGB-D based region-growth algorithm is used to detect objects on table. Experiments are conducted to validate the algorithm.2. An improved background-subtraction based object recognition algorithm is designed. The method use SIFT as feature detector and bag-of-words model to describe object, combines arm’s motility to get clear image of objects based on object localization results, employs the pointcloud of object to filter background. With the method proposed, effective object recognition is realized.3. Semantic part-based3D object class recognition algorithm is proposed which based on geometrical feature and structure.Minimum-bounding-box based segmentation algorithm is used to get components.Spinlmage is used as geome--tric feature descriptor. At the same time probabilistic description is generated using pLSA model. And gaussian model is used to describe the Volume ratio, position and distance to center of same kind components, realizing description of structure. Energy function which describe geometric feature and structure is used to get the similarity between unknown objects and learned models. Experiments demonstrate accuracy and scalability of this method.