A Minimal Dataset Construction Method Based on Similar Training for Capture Position Recognition of Space Robot

Recognizing capture position for non-cooperative targets is an important component of on-orbit service. Traditional machine learning works could not satisfy the requirements of space mission, which demands universality, accuracy and real-time performance. To meet those requirements, an innovative job based on deep learning called Faster Region-based Convolutional Neural Network (Faster RCNN) is introduced for space robot capture position recognizing. Based on the principle of similar training, a minimal dataset construction trick is proposed in order to solve the problem of fewer training samples in space environment. Firstly, the Deep Neural Network is pre-trained through ImageNet training set. Then, using the trained weights as the initial weight of the network, the network is fine-tuned by 1000 training samples in space environment. Finally, a simulation experiment is designed, and the experimental results indicate that the similar training principle can solve the problem of capture position recognition of non-cooperative targets.