Fast Dense Mapping Based on Signed Distance Function Submaps

Abstract

In order to decrease computational complexity of dense mapping in large-scale environment based on Euclidean Signed Distance Function(ESDF) submap model, a registration algorithm based on octree sampling and sliding window structure is designed. By introducing the octree structure, the surface points on ESDF model with larger weights are evenly selected uniformly to reduce the number of residuals. The sliding window structure keeps the number of optimization variables constant, keeping the optimization time within the controllable range. We integrate these algorithms into the Voxgraph framework to build a new fast mapping system. The experimental results show that the octree sampling algorithm can decrease registration speed by 50% and the sliding window structure can decrease mapping speed by 33%.