TARGET POSE ESTIMATION VIA DEEP LEARNING FOR MILITARY SYSTEMS

Abstract

Target pose estimation and aimpoint selection is crucial in direct energy weapon systems, as it allows the system to point to a specific and strategic area of the target. However, it is a challenging task because a dedicated attitude sensor is required. Motivated by new emerging deep learning capabilities, the present work proposes a deep learning model to estimate a target spacecraft attitude in terms of Euler angles. Data for the deep learning model were experimentally generated from 3D UAV models, incorporating effects such as atmospheric backgrounds and turbulence. The targets pose was derived from the training, validation, and prediction of 2D keypoints. With a keypoint detection model it is possible to detect interest points in an image, which allows us to estimate pose, angles, and dimensions of the target in question. Utilizing a weak-perspective direct linear transformation algorithm, the pose of a 3D object with respect to a camera from 3D to 2D correspondences could be determined. Additionally, from this correspondence, an aimpoint, mimicking laser tracking could be determined on the target. This work evaluates these methods and their accuracy against experimentally generated data with simulated real-world environments.