A Basic Study on Robot Position Estimation in Indoor Environment using Deep Neural Network based on the Camera and LiDAR Sensors

Abstract

It is important to know its own position in order to the robot to move autonomously in the environment. In the outdoor environment, GPS can be used to determine the position, but in the indoor environment, the position is estimated by complicated calculation by LiDAR or image sensors. Currently, SLAM (Simultaneous localization and mapping) method is used to create a map and estimate the position for the robot navigation. This method is disadvantageous in that there is a risk of divergence in complicated algorithm calculations, the processing time is long and uneven using various sensors having a large amount of information relatively, and it is difficult to know the self-position when starting from any location.<BR> In this paper, we propose a deep neural network model that outputs absolute position using Lidar and image sensor data. The position estimation method using this deep neural network can predict the absolute position immediately at every location, and the calculation of the model inference is executed periodically without divergence. In the experiment, training dataset is builded by acquiring LiDAR data and images in the indoor environment, and it is used to learn the deep neural network model for about one-month. As a result, the top-5 accuracy of the inference is about 98%, and the estimated position error is about 14cm on average.