Hashing-based Map Indexing for Long-Term Visual SLAM
This work describes a map indexing method that bounds the size of map used in real-time SLAM. The key idea is to quickly pooling a subset of map that are similar to current measurements, while ignoring all other map points that are distinct. A multi-index hashing based map indexing algorithm is developed.
Furthermore, the hashing-based map indexing can be combined with low-latency good feature matching, which yields better performance-accuracy trade-off than each of the single improvements.
The hashing feature is open-sourced at our SLAM stack at: https://github.com/ivalab/gf_orb_slam2