System Overview

Open3D SLAM is a pointcloud based SLAM system. It takes pointclouds from various sensor modalities (e.g. LiDAR or depth camera) and produces a globally consistent map of the environment. An overview of the system is given in the image below:

The scans are forwarded to the odometry module that performs scan2scan matching to estimate ego motion. The odometry is used as an initial guess for the scan to map refinement which estimates ego motion as well as builds the map of the environment. The map is divided into submaps and open3d_slam builds pose graph by introducing constraints between different submaps.

Scan Matching

Done using ICP for both scan2scan and scan2map. For documentations and tutorials on ICP you can refer to the original open3D documentation.

Place Recognition

We rely on RANSAC and FPFH features. As soon as the submap is finished, we match it against other finished submaps in proximity (low drift assumption). The place recognition module runs RANSAC feature matching and looks at number of inliers (too low and we reject the loop closure). In addition, open3d_slam refines the RANSAC registration with ICP (if fitness score is too low, the loop closure is rejected).

Pose Graph Optimization

We use a pose graph backend from open3D. You can check the original documentation here .