OVERVIEW

This thesis focuses on developing a Semantic-Hierarchical Iterative Closest Point (SH-ICP) algorithm that leverages semantic information and hierarchical structures to improve point cloud registration in complex environments. Traditional ICP [2] often struggles with ambiguities in large-scale or cluttered scenes, especially when significant portions of the environment are occluded or dynamically changing. The proposed approach integrates semantic segmentation and hierarchical scene representations to guide the correspondence and alignment processes, ensuring robust and efficient registration. By segmenting the point clouds into semantically meaningful regions (e.g.,roads, buildings, vehicles) and prioritizing hierarchical levels of detail, the algorithm can focus on high-confidence regions for alignment while retaining the flexibility to handle fine-grained details.

OBJECTIVES

The objectives of the thesis are the following:

• Semantic-Aware Correspondences: Use semantic segmentation to categorize point clouds into meaningful regions (e.g., road, vegetation, building). Prioritize correspondences within semantically consistent regions to improve robustness and reduce outliers.

• Hierarchical Scene Representation: Develop a multi-level representation of the scene, starting with coarse structures (e.g., buildings, terrain) and refining to finer details (e.g., furniture, small objects). Use a coarse-to-fine registration strategy for improved efficiency and accuracy.

• Robust Optimization: Integrate semantic consistency checks and hierarchical weighting into the ICP optimization process. Address partial overlaps by assigning adaptive weights to under-represented or occluded regions. Evaluate the trade-offs between accuracy and computational cost to ensure suitability for real-time applications.