Absolute Localization in Feature-Poor Industrial Confined Spaces

Abstract

Autonomous inspection of dark, confined, and feature-poor spaces requires robotic platforms to utilize accurate and reliable localization systems for safe and reliable operation. This paper presents an absolute localization system for highly feature-poor spaces, using visual inertial odometry and GPU-based point cloud registrations for limited field-of-view sensors. The extracted structural elements from sensor scans, along side IMU measurements, are used to limit the search area for the GPU-based point cloud registrations. We employ Stein-ICP which is an uncertainty aware variant of ICP. The 3D registrations are then fused with a visual-inertial odometry estimate in an Extended Kalman Filter to provide a fast and accurate absolute pose estimate. The proposed localization system is tested in both a simulated environment and in a mock-up model of a chemical distillation column - both highly feature-poor areas.