MirrorVision: Light-Weight Floor Detection System for an Autonomous Robot in a Crowded Elevator

Abstract

Delivering robots impact many facets of our life, including food delivery and restaurant services, with advancements enabling obstacle overcome, faster delivery, and minimizing human intervention. However, delivering robots remained to experience poor vertical mobility-elevator usage in multi-floor buildings. Incorporating new elevator models into the robot’s elevator usage capabilities involves a long process of manufacturer approval and authentication. Furthermore, strict fire-code regulations pose communication barriers between the robot and the elevator. In this paper, we introduce MirrorVision-a novel approach designed for accurate floor detection during vertical mobility, regardless of obstructions blocking the robot’s direct line of sight to the elevator number panel. First, we collected and pre-processed a dataset of direct and reflective views of elevator number panels via the pre-installed mirrors. Then, we trained mirrored images in various possibilities to accomplish accurate floor detection. MirrorVision provides a solid mechanism to understand floor numbers at the level of distorted images. Extensive evaluations show that MirrorVision achieves 98.8% accuracy for floor detection in a crowded elevator, while state-of-the-art EfficientDet and YOLOv5 achieved 90.8% and 93.3%, respectively.