Mobile-SPEEDNet: A Lightweight Network for Non-Cooperative Spacecraft Pose Estimation

Considering that the spacecraft pose estimation model method deployed on the onboard computer must have low storage and high performance, an end-to-end regression network, Mobile-SPEEDNet, is proposed. Because of the complex spatial background and sensitivity to image resolution in spacecraft pose estimation, Mobile-SPEEDNet takes MobileN et-v2 as the backbone network, embeds the Coordinate Attention module in partially the inverted residual modules, adds multi-scale feature layer fusion to enhance features, and uses the Spatial Pyramid Pooling layer to extract features, decoupling the position and attitude quaternion information for output. This paper also analyzes the impact of target distance on pose estimation, the effectiveness of attention mechanisms, and the relationship between fine-grained attitude soft assignment encoding and model performance. Finally, experimental results tested on the validation set of the SPEED synthetic dataset are presented to demonstrate the performance, and some prediction results are also presented. The Mobile-SPEEDNet 12-bins model, which has 7.1 million parameters with an average position error of 0.254 meters and an average attitude error of 5.21 degrees, achieves the optimal balance between network parameters and performance.