Master thesis/work student: Explaining Uncertainty in Point Cloud Registration

Description

In many robotics and computer vision applications, including simultaneous localization and mapping, autonomous navigation, object pose estimation, surgical guidance, and augmented reality, point cloud registration, i.e., a problem of estimating a relative transformation between two-sets or clouds of point, plays a fundamental role. The gold standard approach for point cloud registration is called iterative closest point (ICP) algorithm. ICP iteratively reduces the Euclidean distance between matching point pairs from both point clouds given an initial estimate.

Despite the popularity of the algorithm, there exist various sources of uncertainty and inaccuracy that negatively impact the ICP pose alignment procedure in the real world. They include unclear initial poses, noise from sensors, partial overlap, more than one local minima of the cost function, and under-constrained or ill-posed instances, including lengthy, featureless hallways or rotationally symmetric items that admit infinite solutions, like bottles.

In this project, we aim to explain the uncertainty estimates in the ICP algorithm. Concretely, given an output of the algorithm and its uncertainty estimate, this thesis will draw inspirations from explainable AI literature, and produce an algorithm that can attribute the sources of uncertainty in real-time. Evaluation of the algorithmic work on a real robotic system is also envisioned.

Work packages

• literature review.
• analysis and collection of the task-oriented data.
• design of learning modules and validation.
• (bonus) integration into existing robotic systems.

Requirements

• Highly dedicated and motivated student.
• Relevant background in machine learning.
• Prior practical experience in explainable AI methods is a bonus.

Contact details