A bilinear model based solution to object pose estimation with monocular vision for grasping

Object grasping is an important step in robotic applications for subsequent operations, such as delivery and assembly. Automatic object pose estimation with monocular vision provides useful visual cues for grasping and makes it flexible. However, some of the pose factors, such as the pitch angle and the yaw angle, are difficult to estimate from the monocular vision. In this paper, a modified bilinear model is used to separate the pitch factor and the yaw factor from the object image so as to estimate the particular pitch angle and yaw angle. The iterative singular vector decomposition (SVD) in bilinear model fitting imposes a great computation burden. Thus, a random projection algorithm is used to reduce the dimension of the data while preserving the performance of the bilinear model. A weighted Euclidian distance based factor identification method, which discriminates the importance of the elements of the factor parameters, is presented to improve the robustness of the factor identification. Furthermore, with the pitch angle and the yaw angle estimated from the modified bilinear model, a three-step object pose estimation solution is proposed. Experiments are performed to verify the proposed pose estimation solution.