Elevator button detection by filtering objects in mirror regions

To delivery robots, high accuracy of elevator button detection can satisfy requirements for delivery robots moving floor-by-floor well in building. However, previous button detection networks can't exclude mirror button well. In this paper, we propose a new pipeline to solve this problem. We made a new pipeline by unifying elevator button detection network, mirror segmentation network and our algorithm. First, object detection network YOLOv8 detects all elevator buttons in image by bounding box and return all box information including positions and classes. At the same time, segmentation network U-NET segments mirror region and return mirror segmentation mask. Lastly, our algorithm excludes mirror button information from YOLOv8's button information based on U-NET's masks. A new pipeline exclude mirror button stably using Our button grouping mechanism in our algorithm. So, we can validate performance of our algorithm quantitatively and qualitatively in our own datasets including elevator button image with mirror.