Active Multi-View Object Pose Estimation Using a Mobile Ground Robot

Abstract

Accurate 6-DoF object pose estimation remains a central challenge in robotic perception, particularly when relying on single-view observations subject to occlusions and limited geometric cues. This thesis presents a system that incrementally refines object pose estimates by collecting multi-view observations with a mobile ground robot, the Clearpath Jackal. The robot autonomously navigates in a circular trajectory around a target object, capturing images while maintaining a fixed orientation toward the object center. At each waypoint, 2D image corners are manually annotated and paired with corresponding 3D object coordinates. The Perspective-n-Point (solvePnP) algorithm is then applied to estimate the object's pose relative to the camera. The system transforms these camera-centric poses into a consistent global frame using Robot Operating Systems’ transform library. Using these poses, the system tracks reprojection error to evaluate pose confidence. Across multiple trials, the mean reprojection error consistently decreased as more views were added, confirming that spatially diverse observations improve pose estimation accuracy. A cross-run analysis shows reproducible trends, with error reductions of over 40% in many cases. These results validate the efficacy of active multi-view collection for reducing uncertainty and lay the foundation for future extensions with automated key point detection and You Only Look Once (YOLO) supported multi object detection.