Robotic testbed validation of pose estimation algorithms

Abstract

A pose estimation network is usually evaluated on synthetic images, where the true pose is known by construction. On the robotic testbed the true pose comes instead from an independent measurement, optical tracking, and the images come from a real camera under real illumination: it is the closest condition to orbit obtainable on the ground.

The thesis runs an already trained network on testbed images and compares its estimate with the true pose, along repeatable proximity trajectories.

Objectives

1. Define representative proximity trajectories and repeat them under control.
2. Acquire images synchronised with the tracking system pose.
3. Run the network and compare the estimate with the true pose.
4. Relate the error to geometric and illumination conditions.

Methodology

1. Programming of the six axis arm along the chosen trajectories, with repeats.
2. Synchronisation between camera acquisition and optical tracking sampling.
3. Alignment of the reference frames between tracking, camera and mockup.
4. Network execution and computation of translation and rotation error.
5. Error analysis by range, phase angle and illumination level.

Expected contributions

1. A repeatable validation procedure for the lab pose estimation networks.
2. Error maps against range and phase angle.
3. Identification of the conditions where the network loses reliability.

Tools

Robotic testbed and sun simulator, four camera optical tracking, Python and PyTorch.