A Deep Learning Architecture for Place Classification in Indoor Environment via 2D Laser Data

Mobile robots must be able to perceive their semantic location to perform the given tasks in the indoor environment. This problem is defined in the literature as place classification in which robot locations are classified semantically as room, corridor, and doorway. Deep learning techniques have been used for the semantic classification of the 2D laser data acquired by mobile robots. In this paper, a simple deep learning architecture consisting of only fully connected layers is proposed. The proposed architecture accepts 2D laser data without any pre-processing. The Freiburg79 dataset is used to test the proposed method. Since the dataset has data imbalance, the classification accuracy of the door is low in previous studies. The pose rotation was applied to overcome this problem. Intra-class variety was reduced and the classification accuracy of the door class is increased. In addition, the pre-processing and cost-sensitive learning techniques were applied to overcome the negative effects of the data imbalance on the Freiburg79 dataset. The proposed method was trained and tested using Freiburg79 laser data. Moreover, Freiburg52 test data was used to evaluate the success of architecture in different environments.