Research On Object Pose Estimation And Tracking Algorithm

For accurate and efficient manipulation,it is essential for robot to estimate the pose and track the motion of the object in the scene.In recent years,these two tasks have achieved rich academic results.However,in the practical application process,there are still some problems such as inac-curate pose estimation and tracking target loss,which can not meet the requirements of accuracy and speed.This thesis conducts researches on object pose estimation and tracking in robot operation scene,which involves three modules:voting based fast object pose estimation,end-to-end robust object pose estimation and hand-held object tracking.The main contributions are as follows:1)A fast object pose estimation algorithm is proposed based on voting from 3D point-to-keypoint.In the proposed method,pose-wise feature from RGB-D data is employed to predict the direction vectors from spatial points to 3D keypoints.Then the position of each keypoint is inferred based on RANSAC voting.Finally,the pose transformation is calculated by the least square algorithm,which can effectively learn the spatial structure information of rigid body.The accuracy of the proposed method is 98.7%on LineMOD dataset,52.6%on Occlusion LineMOD dataset and 99.1%on YCB-Video dataset.The time for pose estimation is only 0.012s2)An end-to-end robust pose estimation algorithm using differentiable outliers elimination is pro-posed.The proposed method utilizes network for keypoint regression and a differentiable ge-ometric pose estimator for pose error back-propagation.In this way the training accuracy and efficiency can be improved under the supervision of the pose label.Besides,to achieve better robustness when outlier keypoint prediction occurs,a differentiable outliers elimination mech-anism is further designed for both keypoint location and pose estimation.The proposed method achieves the state-of-the-art on three public benchmark datasets and 99.8%accuracy in experi-ments for real object.It takes 0.03s for pose estimation3)A method of hand-held object tracking and automatic initialization is proposed based on siamese network.SiamRPN is selected as the basic framework.To overcome the problem that the track-ing template cannot adapt to the changes of object pose and environment,the online updating of tracking template is designed to improve the discrimination ability of the algorithm.For the problem that the target is easy to be lost when encountering similar distractors,the constraint penalty of hand marker is proposed to improve the robustness of the algorithm.Meanwhile,a tracking automatic initialization algorithm is proposed to realize the complete tracking process The experimental results show that the tracking accuracy is improved from 0.4045 to 0.7107,and the loss rate is reduced from 39.34%to 0.05%.The tracking speed is 69.52FPS.