Data Augmentation for Stroller Detection on Escalators Considering Camera Installation Conditions

Abstract

Stroller detection is an essential technology to address the issue of accidents on escalators. However, there is limited training data of strollers publicly available, despite a large difference in appearance and occlusion properties depending on the camera environment. This study evaluates the performance of stroller detection on escalators by data augmentation utilizing synthesized images that reflects the installation environment of surveillance cameras. Although we could not identify image synthesis conditions that contributed to detection performance, experimental results showed the advantage of our data augmentation and suggested that suitable synthesis conditions differ depending on the camera environment.