A PointNet Application for Semantic Classification of Ramps in Search and Rescue Arenas

Özet

Search and rescue environments could be dangerous for humans due to risks for potential structural breaking down and leakage of hazardous materials. Using robots in these environments can be appropriate solution to reduce these risks. The National Institute of Standards and Technology (NIST) proposed reference test areas for measuring autonomous navigation capabilities of mobile robots. In this paper, we present a PointNet application for semantic classification of ramps through point cloud data in reference test arenas. Since the walls and terrain carry important semantic information for robot navigation, they are also considered. The previous studies that address the semantic classification problem mostly used image and/or 2D laser range data. However, the image data may not be suitable for dusty and poorly lightened search and rescue environments and 2D laser range data may not represent 3D geometry of the objects. Since point cloud data have the ability to describe 3D geometry and it is not affected by the negative aspects of these environments, it could be appropriate to classify ramps, walls, and terrain. Eskisehir Osmangazi University (ESOGU) laboratory building is modelled in GAZEBO simulation environment and a Pioneer P3-AT mobile robot with Asus Xtion Pro sensor is launched. Then, the ESOGU RAMPS dataset is generated through navigating the robot via Robot Operation System (ROS). The dataset involves two types of ramp (inclined and flat), terrain and wall classes. The PointNet is applied to train and test the dataset. The metric and visual PointNet results are presented to analyze the classification performance.