Autonomous landings on unknown gravitational bodies with high accuracy are significant technologies for future exploration. This study demonstrates a method to estimate self-position on an unknown celestial body for high-precision landing using only limited computational resources assuming space probes. This method removes feature points that do not capture topographical features well and makes position estimation faster and more accurate when generating a map of feature points from images captured by the space probe using general-purpose feature point detection methods (AKAZE, SURF, ORB, and BRISK). We create 40 patterns of datasets comprising 3,200 images, including four types of space-specific disturbances, while assuming an unknown celestial body. We verify the accuracy of self-position estimation using these datasets on a Raspberry Pi 4 Model B, emulated as a space probe. The results indicate that next-generation landings on the order of 100 m are possible even in resource-limited environments. Particularly, when using AKAZE, which is more significant than the conventional landing accuracy of the order of 10 km. It is also possible to achieve this as accurately as before, albeit at a significantly 3x faster rate. Our results emphasize that this method allows for highly accurate self-position estimation for high-precision landing using only camera images on space probes in a low-computational-resource environment.