Real-time 6DOF (6 Degree of Freedom) pose estimation of an uncooperative spacecraft is an important part of proximity operations, e.g., space debris removal, spacecraft rendezvous and docking, on-orbit servicing, etc. In this paper, a novel efficient deep learning based approach is proposed to estimate the 6DOF pose of uncooperative spacecraft using monocular-vision measurement. Firstly, we introduce a new lightweight YOLO-liked CNN to detect spacecraft and predict 2D locations of the projected keypoints of a prior reconstructed 3D model in real-time. Then, we design two novel models for predicting the bounding box (bbox) reliability scores and the probability of keypoints existence. The two models not only significantly reduce the false positive, but also speed up convergence. Finally, the 6DOF pose is estimated and refined using Perspective-n-Point and geometric optimizer. Results demonstrate that the proposed approach achieves 73.2% average precision and 77.6% average recall for spacecraft detection on the SPEED dataset after only 200 training epochs. For the pose estimation task, the mean rotational error is 0.6812°, and the mean translation error is 0.0320m. The proposed approach achieves competitive pose estimation performance and extreme lightweight (~ 0.89 million learnable weights in total) on the SPEED dataset while being efficient for real-time applications.

A detailed understanding of space objects is one of the important goals of space situational awareness. The geostationary orbit (GEO) belt is an important space asset for human beings, so the identification of GEO satellites is one of the measures to ensure the safety of GEO objects (GEOs). In this paper, we propose using deep learning based on recurrent neural networks (RNNs) and convolutional neural networks (CNNs), and multiple Kernel learning (MKL) to identify the shape and attitude of GEOs synchronously via light curves. Our algorithm focuses mainly on optical data obtained from the real measured data collected by optical laboratory and computer simulation. We first acquired light curves of five GEO satellites for 1 year; then, we constructed a network architecture consisting of CNNs and RNNs to automatically extract the different scale characteristics of the collected light curves of GEOs. Next, we use the MKL to fuse the extracted features of different scales. Finally, the support vector machine is used to provide the classification and recognition results of the shape and attitude of five GEOs. The network architecture proposed is compared with more conventional machine learning techniques (e.g., principal component analysis, linear discriminant analysis) and is shown to outperform such methods. At the same time, the classification effect of the multiple kernel is better than the single kernel in this experiment.

A novel satellite target recognition method based on radar data partition and deep learning techniques is proposed in this paper. For the radar satellite recognition task, orbital altitude is introduced as a distinct and accessible feature to divide radar data. On this basis, we design a new distance metric for HRRPs called normalized angular distance divided by correlation coefficient (NADDCC), and a hierarchical clustering method based on this distance metric is applied to segment the radar observation angular domain. Using the above technology, the radar data partition is completed and multiple HRRP data clusters are obtained. To further mine the essential features in HRRPs, a GRU-SVM model is designed and firstly applied for radar HRRP target recognition. It consists of a multi-layer GRU neural network as a deep feature extractor and linear SVM as a classifier. By training, GRU neural network successfully extracts effective and highly distinguishable features of HRRPs, and feature visualization technology shows its advantages. Furthermore, the performance testing and comparison experiments also demonstrate that GRU neural network possesses better comprehensive performance for HRRP target recognition than LSTM neural network and conventional RNN, and the recognition performance of our method is almost better than that of other several common feature extraction methods or no data partition.

Wide-field telescopes are widely utilized to observe geosynchronous objects, with low cost and high performance. In geosynchronous object detection, stars produce the most false alarms. Thus, star-suppression is required to improve the accuracy of geosynchronous object detection. However, traditional methods involve false identification and missing identification due to the significant distortion of wide-field telescopes and undersampling of fiant images. Distortion could lead to the observed positions of stars deviating from their theoretical positions, while fiant and undersampling images could result in inaccurate centroid extraction. In this study, a method was proposed for analyzing the motion and structural characteristics of stars, space objects, and noise, focusing on maintaining a high star identification rate and low false identification rate with the mentioned errors. First, the invariance of the star structure under the influence of projection transformation and distortion was strictly derived. Second, two types of star-pair indices were suggested. Assessment of the structural stability was added to the foundation of the centroid matching. For accurate identification of stars, the influence of dynamic distortion was effectively inhibited using the established star-structure evaluation system. Our method achieves lightweight, straightforward, and rapid detection without requiring any prior star catalog or telescope distortion model. The experiment demonstrates that the performance of the proposed method is superior to that of previous methods influenced by distortion and centroid positioning error.

Limited by the conditions and performance of ground-based optical observations, it is difficult for us to obtain a plethora of optical cross section (OCS) data for some space objects (SOs). Unevenly distributed OCS data and unclear labels will affect the performance of SOs recognition based on neural networks. Furthermore, when we need to identify a new SO or SO category using deep neural network, the trained network model may no longer be applicable. We need to retrain the network with new training data. In order to alleviate these problems and improve the generalization and training convergence speed of SOs recognition networks, a novel, to the best of our knowledge, neural network model, ARSRNet, is proposed in this paper. The ARSRNet can identify SOs and their attitude accurately using only a small quantity of training OCS data and without clear labels. And the proposed network is able to adapt to new recognition tasks. Meanwhile, we propose an AdamRprop network optimization algorithm to accelerate network training and improve recognition accuracy. Experimental results show that the recognition accuracy of ARSRNet reaches 90.60% on the test OCS dataset. Compared with mainstream network optimization algorithms, the proposed AdamRprop is more appropriate for ARSRNet and can accelerate the convergence of ARSRNet.