Abstract Object classication and intra-class part correspondence determination in depth images for robot manipulation The motivation for the research presented in this dissertation is development of intelligent robots capable of operating in unstructured environment such as houses, hospitals and schools. Such robots are expected to manipulate with previously unseen objects. In order to facilitate autonomous locomotion and manipulation, a robot system requires an appropriate perception sensor. RGB-D camera as a sensor provides 3D information about the robot's environment. The information about the shape and size of objects on a scene, obtained by the RGB-D camera, is useful in object recognition and classication algorithms, as well as in algorithms for semantic segmentation. Segmentation of objects into semantic parts is required in order to identify the object parts, relevant for dierent robot operations. Furthermore, RGB-D cameras provide information about the object pose and the pose of its parts, which is, along with the robot arm visual features, utilized in visual servoing and robot operation planning. The goal of the research presented in this dissertation, was to develop methods for object classication in RGB-D scenes, semantic object segmentation and visual servoing. The purpose of these methods is integration in an autonomous robot system with an arm and a camera, capable of recognizing and manipulating with previously unseen objects. The research presented in this thesis attempts to overcome the existing methods limitations. A common drawback of the existing object classication methods is not estimating the object's pose required for application in robot manipulation. Furthermore, there is a relatively small number of research regarding semantic part segmentation and part correspondence determination for robot manipulation tasks, while the part annotation of large object datasets is labour-intensive. Finally, the lack of hand-eye calibration methods suitable for real time recalibration as well as visual servoing methods suitable for low cost robots 109 Abstract are common aws of the state-of-the-art approaches. The mentioned drawbacks encourage the development of novel approaches. During the research, four novel methods are developed: 1. A visual servoing method for SCARA robots using an RGB-D camera as the only sensor. This method utilizes a marker, placed at the robot end eector, which determines the current robot tool position, and dominant plane information obtained by the RGB-D camera. Hand-eye calibration is performed by a single robot arm movement. 2. A method for object classication in depth images based on convex hull alignment. For a given object in an RGB-D image, this method determines the most similar object belonging to a particular class. The method estimates the object position with respect to the referent object from the same class. The learning is based on a 3D model database, sorted into classes. 3. A method for object classication in depth images based on polyhedral shape class models. For a given object, this method determines the parameters of the most similar object belonging to the particular class. Analogously to the previous proposed object classication method, the learning is based on a 3D model database, sorted into classes. This method estimates the pose of the object and the correspondences between intra-class object surfaces. 4. A method for determination of part correspondences within object classes based on polyhedral shape class models. This method assigns the semantic labels to the considered objects' segments, based on a small number of annotated representative objects and enables real time database expansion. The developed visual servoing approach is experimentally tested by performing positioning and simple objects' grasping experiments. The accuracy of the proposed classication methods are experimentally evaluated and compered to several state-of-the-art approaches using a publicaly available benchmark dataset. The accuracy of the part association method is evaluated using a set of 3D object models of several classes.