A Study on Barrier-Free Entrance Object Detection Using Deep Learning in Street View Imagery

Ensuring the mobility rights of transportation-vulnerable individuals has become increasingly important, leading to mandatory barrier-free (BF) certification in public buildings and facilities, in South Korea. One key aspect of creating BF environments is developing accessibility information for building entrances. However, gathering entrance information for all buildings through on-site surveys presents substantial time and cost challenges, emphasizing the need for automated methods. This study proposes a deep learning-based framework using street view images (KakaoMap Road View) to detect building entrances and BF-related objects, including entrance doors, revolving doors, stairs, and ramps. The framework utilizes the YOLOv5 model for object detection, aiming to identify attributes relevant to building accessibility. To address the challenges of street view images—such as varying perspectives, complex backgrounds, and occlusions from trees, vehicles, and pedestrians—additional datasets were integrated to improve model performance. Specifically, pedestrian entrance information provided by the Seoul Metropolitan Government, along with open datasets for door and stair detection, were included in the training process. The street view and entrance datasets were labeled for four BF-related classes, and validity filtering ensured high-quality data for training. Experimental results showed that adding datasets significantly improved accuracy compared to using street view data alone. The highest accuracy was achieved by combining entrance images and open datasets. However, the detection performance for stairs and ramps was relatively lower compared to that for entrance doors, likely due to limited samples and the irregular shapes of these objects. This study demonstrates that combining street view images with additional datasets enhances BF object detection accuracy, contributing to BF environment development in public facilities and supporting the mobility rights of transportation-vulnerable individuals.