Hard Example Detection at Actual Environment for Semantic Segmentation Used in Visual Navigation
Yuriko Ueda, Marin Wada, Miho Adachi, Ryusuke Miyamoto · Proceedings of the World Congress on Mechanical, Chemical, and Material Engineering · 2024
Visual navigation based on the results of semantic segmentation is a promising method as it does not require expensive sensors and detailed metric maps to actualize autonomous movement.However, navigation performance strongly depends on the accuracy of semantic segmentation for which high quality training datasets reflecting the target environment are indispensable.This paper presents a novel method to detect hard examples for semantic segmentation in the case of visual navigation.This method adopts interframe changes in the number of class labels and simple thresholding.Experimental results using training and testing data gathered during the course of the Tsukuba Challenge, a competition of autonomous moving robots held in Japan showed that the proposed method is feasible to detect hard examples for segmentation in visual navigation.