Pose Determination from Multi-View Image using Deep Learning

A high precision method for position and attitude determination is presented. One of the difficulties in pose determination is the lack of reliable visual features. As a result, we exploit a novel convolutional neural network for stereo matching, so as to solve the problem of image feature extraction. The network is based on a siamese network and incorporates dense block into each branch. There is occlusion problem in the close-range phase of measuring object. In order to solve the occlusion, a multi-view system is established. The measurement results meet the accuracy requirements of pose determination and verify the feasibility of the method.