A multilevel object pose estimation algorithm based on point cloud keypoints

The main task of object pose estimation is to predict the 3D rotation and 3D translation of an object in the current scene relative to a fixed object in the world coordinates. The most commonly used algorithm in pose estimation is based on the object characteristics or keypoint information for matching. The accuracy of these algorithms in pose estimation depends on whether the object surface characteristics are apparent. To solve the problem mentioned above, we propose a pose estimation algorithm using multilevel keypoint aggregation in the point cloud. First, we use a deep learning convolutional neural network to predict the keypoint positions in the point cloud. Then we estimate multiple poses at different levels according to the keypoints predicted above. Finally, we aggregate multiple poses into the final pose according to the weight of each pose. Our experiments show that our method outperforms other approaches in two datasets, YCB-Video and LineMOD.