Spacecraft Pose Estimation Based On Monocular RGB Image

Spacecraft pose estimation is an important technology to achieve space attack and defense,in-orbit service,and plays an important role in the aerospace field.Due to the limitations of spacecraft load and resources,low-energy,lightweight vision cameras are more suitable for spacecraft,especially small spacecraft,compared to other devices.Therefore,spacecraft pose estimation based on monocular RGB images is a worthy research topic.At present,the difficulties of the research are as follows: First,it is an ill-posed problem to restore the pose according to the monocular image;second,the existence of stars in space makes the spacecraft image have a very changing background;thirdly,it is difficult to estimate rotation directly using deep network to predict pose.This paper focuses on the spacecraft pose estimation based on monocular RGB images,aiming at the difficulties of the subject,the main work and innovations are summarized as follows :Firstly,aiming at the characteristics of large background change and large target scale change of spacecraft image,this paper proposes a spacecraft pose estimation algorithm based on background suppression.The algorithm first uses a deep network to predict the target mask and the three-dimensional point coordinates corresponding to the target,and then uses the EPn P and RANSAC algorithms to solve the target pose.The feature pyramid structure is introduced,so that the network can generate target masks and corresponding coordinates prediction maps of different scales,so as to deal with the change of target scale.In order to reduce the interference of background on pose estimation,we propose a background suppression branch and a background feature extraction network.In training,we uses the background feature extraction network to obtain the deep features of the background image of the erased target,and uses them to guide the background suppression branch learning background features;In the inference stage,the background suppression branch is directly used to extract the background features,which are subtracted from the features obtained by the backbone network to reduce the impact of background changes at the deep feature level.We performed ablation study on the spacecraft pose estimation public dataset SwissCube,which proved the effectiveness of each module design.The final algorithm achieves ADD(-S)73.15% on the SwissCube dataset,which is better than other algorithms.Secondly,aiming at the problem that the deep network is weakly to estimate the pose rotation parameters in the existing methods,this paper proposes a spacecraft pose estimation algorithm based on surface normal vector supervision.The algorithm first detects the target in the image,and then performs end-to-end pose estimation on the obtained target image.The pose estimation network is divided into two stages.The first stage predicts the geometric features of the target according to the input target image,and the second stage estimates the pose according to the predicted geometric features.In the first stage,the mask,corresponding 3D coordinates,surface region and surface normal vector introduced by this method are predicted for the input target image.In this paper,introducing the prediction of surface normal vector in the geometric feature prediction stage enhances the ability of the existing network structure to extract angle information.The second-stage pose estimation network consists of a convolutional network and a fully connected layer.The three parameters for translation and the the six parameters for rotation are predicted according to the input geometric features.We demonstrate the effectiveness of each module design by ablation study on the spacecraft pose estimation public dataset SwissCube.The algorithm finally achieves ADD(-S)77.92% on SwissCube,which is better than other algorithms.In order to prove the versatility of the algorithm,we conducted experiments on the indoor object pose estimation public datasets LM-O and YCB-V.The results of ADD(-S)60.8% and ADD(-S)60.8% are obtained.It is close to the existing SOTA general pose estimation algorithm.