The invention relates to an end-to-end
spacecraft pose measurement method based on
deep learning, and belongs to the field of
spacecraft pose measurement, the method uses a sliding window
transformer as a
feature extraction backbone network, introduces a multi-layer large kernel
convolution module, and is used for enhancing a global
receptive field and feature expression ability. A rotating head detection network and a position head detection network are arranged in parallel after the
backbone network, a rotating head is used for predicting a 3D key point and a projection weight thereof, a position head is used for predicting a 2D key point, a 2D-3D corresponding relation is sent to a differentiatable EPro-PnP module for
pose solving, it is ensured that a gradient can be completely transmitted back to a
feature extraction layer, and a
feature extraction layer is obtained; therefore, a globally optimal solution is obtained, and integrated network training of
feature learning and geometric pose solution is realized. According to the method, a complex
loss function is constructed, through multi-angle supervision and complementary optimization, it is ensured that the model can accurately predict the 6D pose of the
spacecraft, and robustness and generalization performance in a complex environment are enhanced.