Comparative Analysis for Missing Button Recognition in an Elevator Panel

Abstract

Button recognition in an elevator is a critical task for elevator robots that navigate multi-storey indoor environments. The limited and enclosed space within elevators poses challenges for button recognition, as buttons may be difficult to observe or partially visible due to being out of the frame. To address this issue, we have used various computer vision approaches to improve button recognition, particularly for panels with incomplete buttons. In our previous work, we utilized a combination of YOLOv3 and the gap distance between buttons to predict the positions of missing buttons. However, this method had limitations and could not be considered as a generalized solution to the problem. Therefore, we have devised a modified approach using YOLOv5 for button detection and the Discriminative Correlation Filter with Channel and Spatial Reliability(CSRT tracker) for tracking button movements across consecutive frames. Even if a button temporarily moves out of the frame and reappears later, we can leverage our prior knowledge or estimate of its position to initialize the tracking algorithm and locate the button accurately. This integration of detection and tracking techniques enhances our ability to reliably recognize buttons, even in challenging scenarios.