Development of a Person-Following Mobile Robot System using Deep Learning Model of Re-Identification

Abstract

The person-following technology is widely used in the field of mobile robotics, especially essential for companion robots, and becomes even more critical in its application instances. Allowing robots to track targets in complex and dynamic human environments paves the way for significant advancements in human-robot interaction. Numerous mobile robots utilize visual sensors for these tasks, however, tracking a target in complicated human environments poses diverse challenges. Image information comes with multiple issues, such as occlusion of the target, changes in appearance, variations in illumination, shifts in perspective, and more. This thesis addresses these challenges and proposes a system for person-following. The system is initialized using the background subtraction method to select the target to be followed. And pre-trained deep learning model based on an image dataset was exploited to detect people among various objects. Additionally, using a Kalman filter enhanced the robustness of tracking performance, and a person Re-Identification deep learning model served as a feature extractor for identifying and seamlessly following a specific target. Research involving various deep learning models often requires considerable computational resources for such inference processes. To meet the computational constraints of mobile robots without GPUs, quantization techniques were applied to the models, which were used in the developed system, for computational efficiency and real-time operation. The entire system was implemented on an actual mobile robot and experimentally validated.