Augmented reality (AR) can improve the safety and efficiency of maintenance and
decommissioning work in nuclear power plants (NPP). Accuracy and speed is required for
relocalization methods in order to use AR in NPP. However, the usefulness of the existing
methods is not yet evaluated for this application. Thus, in the present research, the
representative relocalization methods such as Randomized Fern [1] and FAB-MAP [2] are
evaluated and an improved method is proposed. The evaluation results show that Randomized
Fern showed relatively faster and more accurate retrieval of camera poses than that of
FAB-MAP. However, when the number of keyframes stored in the database decreased for
making Randomized Fern applied to a larger environment, the accuracy of the retrieved poses
for the former decreased. Therefore, the authors proposed a method that selects the keyframes
similar to an input image from a sparse database and searches more appropriate frame around
the selected in a dense database. By increasing the keyframes in the dense database, the
proposed method is able to prevent errors caused by insufficiency of keyframes in the sparse
database. The proposed method was evaluated using the NPP datasets. Consequently, it was
able to suppress the degradation of accuracy even when the keyframes were decreased. In the
future, evaluation of the proposed method with large datasets that include every part of NPP is
required. It is also necessary to evaluate the proposed method in changing environment where
workers conduct maintenance or dismantling work of NPP.