Robot Visual Perception and Reasoning in Human-Crowded Environments

This thesis develops new ways for robots to visually understand and reason about people in busy, crowded environments. Robots must recognise how people move, who is interacting with whom, and how different individuals relate to each other to behave safely and naturally. The work introduces methods to detect human groups from motion, creates a detailed dataset showing how people interact, and builds a benchmark to test whether AI systems can perform step-by-step visual reasoning. Finally, it proposes a more reliable and interpretable reasoning framework that combines neural and symbolic methods. Together, these advances help robots better understand and operate in human crowded environments.