Semantically Aware Partial Point Cloud Registration for Industrial Assets

Project Details

• Student(s): Maher Abou Dargham
• Advisor(s): Dr. Joe Tekli
• Department: Electrical & Computer
• Academic Year(s): 2025-2026

Abstract

Digital twins in warehouse and industrial environments rely on accurate instance-level representations of physical objects to support inventory management, simulation, and robotic automation. While significant progress has been made in 3D object detection and point cloud analysis, reliably separating individual objects in stacked or tightly packed configurations remains a major challenge. In such scenarios, multiple objects o the same class exhibit minimal geometric variation and strong physical contact, often leading existing pipelines to merge them into a single instance. This work investigates the potential of adapting principles from partial point cloud registration and semantic-aware registration to address stacked object separation. We review classical, learning-based, probabilistic, and semantic registration methods, highlighting their relevance and limitations when applied to instance-level detection in cluttered industrial scenes. To explore feasibility, we implement a prototype model-to-scene matching pipeline using brute-force geometric filtering and RANSAC-based registration under partial visibility. Experimental results demonstrate that enforcing partial geometric consistency enables more localized instance detection than dimension-based methods alone, although ambiguity persists in highly symmetric configurations. Our analysis identifies key gaps in existing approaches and motivates future integration of multi-instance hypothesis generation, semantic constraints, and learned correspondence models.

By bridging registration techniques with instance-level reasoning, this work contributes toward more accurate digital twins and more reliable downstream automation in warehouse environments.