Dataset Creation for Segmentation to Enhance Visual Navigation in a Targeted Indoor Environment

To show the effectiveness of visual navigation based on the results of semantic segmentation, the authors participating in the Tsukuba Challenge, a well-known autonomous robot motion competition in Japan. The Tsukuba Challenge, which normally includes events only in outdoor environments, included in 2023 a new indoor environmental challenge, in which a robot must navigate a crowded shopping mall. In order to achieve visual navigation in an indoor environment, based on the results of segmentation, we needed to train a classifier suitable to navigate many simultaneously moving obstacles such as crowded people. For this purpose, this paper proposes a method for creating an appropriate dataset using a public dataset and modifying an existing dataset for an outdoor scenario: class labels are re-assigned considering the target environment. Experimental results using images captured at the course of the Extra Challenge in the Tsukuba Challenge 2023 showed that the proposed dataset enables the construction of a better classifier for semantic segmentation than only a public dataset.