A 2D Deep Residual Learning Approach for 3D Indoor Radio Map Estimation

Radio map estimation is important for fingerprint-based localization, robot path planning, aerial base station placement, etc. Existing studies mainly focused on 2D radio map estimation by ignoring the impact of the heights of transmitters, receivers and obstacles, which, however, must be considered in indoor scenarios. Although the existing 2D radio map estimation methods can be extended to estimate 3D radio maps by replacing 2D operations with 3D operations, it significantly increases the computational cost. To devise an efficient method for 3D radio map estimation, we propose to represent the heights of physical objects by pixel values and the radio maps at different heights by output channels in this paper, based on which, a 2D deep residual learning method is proposed to estimate 3D indoor radio maps. The proposed method has good generalization, and can better extract features according to the pathloss characteristics. Experimental results show that our method outperforms the state-of-the-art approaches in terms of accuracy, inference speed, computational cost and memory storage. In addition, due to the lack of public datasets, a 3D indoor radio map dataset (3DiRM3200) of 3, 200 samples for 200 buildings is created, which took more than 1, 000 labour hours. The dataset and our code will be available at https://github.com/lighttime2023/3DiRM3200.git.