DMANet: Dense Multi-scale Attention Network for Space Non-cooperative Object Pose Estimation

Accurate pose estimation of space non-cooperative targets with a monocular camera is crucial to space debris removal，autonomous rendezvous，and other on-orbit services. However，monocular pose estimation methods lack depth information，resulting in scale uncertainty issue that significantly reduces their accuracy and real-time performance. We first propose a multi-scale attention block（MAB）to extract complex high-dimensional semantic features from the input image. Second，based on the MAB module，we propose a dense multi-scale attention network （DMANet）for estimating the 6-degree-of-freedom（DoF）pose of space non-cooperative targets，which consists of planar position estimation，depth position estimation，and attitude estimation branches. By introducing an Euler angle-based soft classification method，we formulate the pose regression problem as a classical classification problem. Besides， we design a space non-cooperative object model and construct a pose estimation dataset by using Coppeliasim. Finally，we thoroughly evaluate the proposed method on the SPEED+ ，URSO datasets and our dataset，compared to other state-of-the-art methods. Experiment results demonstrate that the DMANet achieves excellent pose estimation accuracy.