Real-time object pose estimation and tracking has been a research topic in vision systems for many years. Achieving a satisfactory result from the vision system of an object manipulator robot, is partly dependent on the accurate and real-time results of the pose estimation tasks. The robot needs to estimate the position of the target object and track it while trying to approach it and grasp it. To attain a satisfactory result, embedded systems might be a solution to fulfil some of these needs. In this thesis, through a profound study of the state of the art and a background review of the latest developed robot from Malardalen university, bottlenecks and problems of the vision system of this robot are defined and two of the most recent pose estimation and tracking techniques named SimTrack and OpenCV-PnP are investigated and tested. The goal of this thesis is to maintain the Kinect version 2 frame per second rate by optimizing the object detection and tracking algorithms using parallel programming techniques in GPUs(graphics processing unit). Hardware used in this thesis are a laptop equipped with Intel i7 Haswell quad cores with 2GB of Nvidia GPU with CUDA accelerating support and Kinect V2 sensor. At the end, results depict that SimTrack method is faster and more precise but it also costs more from hardware aspect and the OpenCV (A c++ vision framework) method is fast enough but less accurate in estimation part.