Monocular Vision and Machine Learning for Pose Estimation

As space activities continue to grow, the ability to autonomously service, refuel, or otherwise connect to objects in space will become more critical. A key aspect of being able to maneuver in close proximity is to be able to determine the relative pose between two objects, the Chaser and the Target object. Failure to determine the pose can lead to an unsuccessful mission due to an inability to maneuver to a specific position to repair in-operable objects, to dock to stations, and to avoid collision with other objects. More specifically, automating the last step of rendezvous operations where the chaser and target are in close proximity to each other is of interest. This paper introduces a monocular vision-based approach for 6 Degree-of-freedom (DoF) pose estimation on a known object. The proposed solution is to use a Convolutional Neural Network (CNN) to find known features of an object in an image. These known features, together with their known locations, are used by a Perspective-n-Point (PnP) algorithm to estimate the pose of the target object with respect to the camera. The primary difficulty with CNN-based methods is needing to generate a large amount of training data to effectively create the CNN. To overcome this difficulty, a three-dimensional (3D) model of the real-world object is created and used in a visualization environment to create images of the object from many different perspectives and with differing backgrounds. This approach enables the creation of a very large truth dataset in a short time period. This synthetic imagery is used to train a You Only Look Once (YOLO) network, enabling rapid and accurate feature recognition in a single image. The solution gives average position and rotation magnitude error of 1.2 cm and 1.22 degrees, respectively, at contact point (1 to 2 meters) on synthetic data.