Study on Improving the Accuracy of Image Matching Self-Position Estimation by NG-RANSAC

抄録

In order to achieve high accuracy image navigation in space, it is necessary to perform highly accurate self-position estimation by matching images taken by the probe's camera with map images obtained through preliminary exploration. However, images taken in space are affected by disturbances due to the position of the sun and camera anomalies, and information loss due to compression for data transmission. As a result, the matching of corresponding points between images yields many outliers and the accuracy of self-position estimation is reduced. To address this problem, Random Sample Consensus (RANSAC) [1] is conventionally used to remove outliers. However, RANSAC significantly increases processing time for corresponding point groups with a large proportion of outliers. In this study, we adopt Neural Guided RANSAC (NG-RANSAC) [2] for images with space-specific disturbances and achieve higher accuracy.