Adaptive Light Space Target Pose Estimation Method Based on Deep Learning

Addressing the insufficient accuracy in recognizing the attitude of space objects under low-light conditions,an illumination-adaptive attitude estimation method for space targets based on deep learning is proposed. Within the integrated deep learning network framework for illumination-adaptive posture recognition, firstly, a lightweight convolutional neural network module is utilized for image enhancement,achieving illumination-adaptive fusion of the original images. Subsequently,a multi-scale feature extraction network is established to extract features from the fused images. The envelope angle points in two dimensions of the target are then predicted through a regression network. Lastly,the attitude of the space target under low-light conditions is accurately estimated using the RANSAC-EPNP algorithm. To verify the effectiveness of the proposed method,a lowlight conditions on-orbit target data set is established using Unity for experiments. Experimental results demonstrate that the proposed method not only effectively improves the quality of low-brightness and weak-texture target images,but also significantly enhances the accuracy of space target attitude estimation under low-light conditions. © 2024 Chinese Society of Astronautics. All rights reserved.