Vision-Based State Estimation for Non-Cooperative Targets in Space

With the increasing frequency of space activities, non-cooperative targets such as space debris have brought great hidden dangers to spacecraft operation. How to estimate their state has gradually become the main demand. This paper introduces a new method for estimating the angular velocity value as well as the axis of rotation of non-cooperative targets based on deep neural networks. The proposed algorithm has three parts: (1) use of the convolutional neural network (CNN) YOLO model, which is trained to identify the non-cooperative targets in the images taken by the camera, (2) extracting ORB features in the detected pixel regions of the non-cooperative target and using simultaneous localization and mapping (SLAM) to calculate the rotation matrix and translation matrix of the non-cooperative target relative to the camera, and (3) using the angular velocity value calculated by the Rodrigues equation to estimate the time of loop closure in SLAM. In order to calculate the axis of rotation, the plane fitting and spatial arc fitting are used. Combined with ground test, the experimental results indicate the error of the measured speed value of this algorithm is 0.0081rad/s, and the maximum error of the rotation axis is 5.12° which show the correctness of the algorithm and the potential of online application.