A Door Detection System for Convenience Stores in Taiwan

Abstract

A door is a very substantial element since it enables a person to enter a target place. Though detecting a doorway is an easy task for a regular person, it is challenging for robots or visually impaired people. Although most existing deep learning object detection models have shown promising results, they have a limitation in distinguishing between glass doors and glass walls in a convenience store. To address this issue, we propose an effective door detection system for convenience stores in Taiwan. Our system consists of two main models: 1) the object detection and 2) the door bounding box models. The former model uses the re-train YOLOv4 as the main building box. The latter model uses a fully connected neural network as the main building box. In particular, we utilize the surrounding objects in the scene to improve the performance and robustness of convenient glass door entrance detection. The experimental results demonstrate that our proposed method not only achieves the quantitative accuracy result up to 93% but also provides decent qualitative results.