Dual-channel decoupling method for spatial target pose estimation based on RGB images

Abstract:
To address the problem of coupling position and pose information in spatial target motion state perception, a spatial target pose estimation method based on two-channel deep learning of RGB images is investigated. Combining the characteristics of RGB images and feature extraction requirements, a ResNet neural network-based spatial target pose decoupling model is constructed, and the pose classification and pose regression loss functions are established using cross-entropy functions and an absolute value loss function respectively. The evaluation criteria of the pose decoupling method are given by integrating the position and pose estimation errors. As tested by the ESA Spacecraft Attitude Estimation Contest score evaluation standard, the Norm. err pose of the proposed network model in Lightbox Postmortem is 71.29% better than the baseline; The Norm. err pose of Sunlamp Post-mortem is 0.1550, which is 57.57% better than the standard result. The Best Score is 2.6557, which is 99.4% of the standard result. The validity of the algorithm was verified.