OVERVIEW

The goal of this Master Thesis is to develop a one-shot global localization method for robotic systems using 3D point cloud maps. Global localization is a critical task in robotics, especially in scenarios where robots must determine their initial position in a large, previously mapped environment without relying on prior odometry or GPS data. The proposed approach will leverage state-of-the-art methods in 3D point cloud processing, place recognition, and feature matching to achieve accurate and efficient localization in diverse and challenging environments. This thesis will focus on developing a robust pipeline that processes a query point cloud captured by a LiDAR sensor and matches it against a large pre-built 3D map to determine the robot’s pose. The candidate will evaluate the method on publicly available datasets (e.g., KITTI [2], Apollo-SouthBay [3]) and/or custom datasets collected in controlled environments.

OBJECTIVES

The objectives of the thesis are the following:

• Place Recognition in 3D Point Clouds: Design a feature extraction and matching algorithm to identify the most similar region in the 3D map for a given query point cloud. Explore methods like learned global descriptors (e.g., PointNetVLAD) or handcrafted features (e.g.,FPFH, SHOT).

• Pose Estimation: Develop or adapt algorithms (e.g., RANSAC-based or iterative registration) to compute the transformation aligning the query cloud with the map.

• Efficiency Optimization: Implement efficient indexing and search techniques (e.g., KD-trees, Approximate Nearest Neighbors) to handle large-scale maps. Evaluate the trade-offs between accuracy and computational cost to ensure suitability for real-time applications.