Tracking with the CAD Model of Object for Visual Servoing

Abstract

Tracking the 3D pose between the object and the camera is a very important research subject in position-based visual servoing (PBVS). The tracking requirements are extremely fast and accurate. In order to solve these problems, a real time and efficient tracking algorithm based on Computer Aided Design (CAD) model of object is proposed. This method needs only a monocular camera in PBVS system. It is robust to the calibration errors and the feature extraction errors, and can also be applied in the case where the part of target is not in the field of view. Tracking object is equivalent to calculating the pose of object in camera in real time. The priori CAD model of the object can be projected into the scene through the pose of the previous frame. The pose between the object and the camera is tracked and updated by virtual visual servoing (VVS) method. An edge tracking algorithm is used to track the edge of the object. In order to filter the noisy edge points, you only look once (YOLO) algorithm is used to get the range of the target position in the image plane. The use of M-estimator in control makes visual servoing more robust. The experimental results for six degrees-of-freedom visual servoing system demonstrate the effectiveness of the algorithm.