Deep Learning Methods for Semantic Segmentation of Dense 3D SLAM Maps

IR HANDLE

Abstract

Most real-time SLAM systems can only achieve semi-dense mapping, and the robot lacks specific knowledge of the mapping results, so it can only achieve simple positioning and obstacle avoidance, which may be used as an obstacle in the face of the target object to be grasped, thus affecting the realization of motion planning. The use of semantic segmentation in dense SLAM maps allows the robot to better understand the map information, distinguish the meaning of different blocks in the map by semantic labels, and achieve fast feature matching and Loop Closure Detection based on the relationship between semantic labels in the scene. There are many semantic segmentation datasets based on street scenes and indoor scenes available for use, and these datasets have some common tags. Based on these training data, we can derive a semantic segmentation model based on RGB images by using the Pytorch platform for training.

The 2021 International Conference on Artificial Life and Robotics (ICAROB 2021), January 21-24, 2021, Higashi-Hiroshima (オンライン開催に変更)

Journal

Keywords

Details 詳細情報について

Report a problem

Back to top