Initial Pose Estimation of an Autonomous Mobile Robot via Feature Object Learning and Recognition

This paper proposes an automatic initial localization system using an RGB-D camera and YOLOv5 to address the sensitivity of particle filter-based localization to initial pose settings. The system utilizes real-world objects recognizable by humans as semantic landmarks, registers them on a 2D SLAM map, and estimates the robot's initial position and yaw angle based on the relative distance and orientation to the observed objects. Experimental results show that the proposed system achieved an average position error of 55.55 cm and a yaw error of 12.51°, demonstrating faster and more stable convergence compared to random initialization. Furthermore, by integrating with Large Language Models (LLMs), the system shows potential for extension to natural language-based localization and path planning, suggesting its applicability to intuitive, human-robot interaction-based autonomous systems.