RGBD-based robot localization in sewer networks

Abstract

This paper presents a vision-based localization system for global pose estimation of a sewer inspection robot given prior information of the sewer network from local institutions. The system is based on a Monte-Carlo Localization system that uses RGBD odometry for the prediction stage. The update step takes into account the sewer network topology for discarding wrong hypotheses. Moreover, this step is further refined whenever a discrete element of the network (i.e. manhole) is detected. To this end, another RGBD camera pointing upwards is used for precise manhole detection. A Convolutional Neural Network has been successfully trained for classifying images with and without manholes with 96% accuracy over the tested dataset. The complete system has been validated with real data obtained from the sewers of Barcelona yielding accurate localization results. All the logs and code used in the context of this paper are publicly available.