Abstract:

In recent years, the variety and number of tasks that expected to perform by robots have been increasing. For example, some of these tasks are to carry an object from a location to another one or to guide people where they desire to reach in large indoor environments such as school and hospital. The semantic classification of the robot locations may contribute to the robots while performing these tasks successfully. In indoor environments, room, corridor, door, hall, elevator, and stair could be considered as the semantic classes that the robot can locate. In previous studies, clustering, supervised, and unsupervised machine learning techniques used with 2D laser data to classify robot locations semantically. In this work, apart from the previous studies, the point-based deep learning architecture PointNet++ was utilized to determine the room or corridor semantic classes. To do that, the raw distance data acquired with the 2D laser range finder was converted to point cloud and the resultant data is used to feed the PointNet++ architecture. Besides, data augmentation was applied to raw point cloud data by means of scaling operation to learn the characteristics of the room and corridor classes regardless of dimensions. The Freiburg 79, Freiburg 52, ESOGU, and SDR-B datasets that include rooms and corridors which have different sizes were used to test the effectiveness of the implemented method. The test results were evaluated with accuracy, recall, precision, and F1 score metrics.

Keywords: