Semantic Mapping and Object Detection for Indoor Mobile Robots

In this paper the authors present a full solution for object-level semantic perception of the environment by indoor mobile robots. The proposed solution not only provides means for semantic mapping but also division of the environment into clusters representing singular object instances. The robot is provided with information that not only allows it to avoid collisions with obstacles present in the environment, but also information about the localization, the class and the shape of each encountered object instance. This level of perception enhances the robot's ability to interact with the environment. The state-of-the-art deep learning solution, Mask-RCNN, is used for the image segmentation task. The image processing network is combined with an RTAB-Map SLAM algorithm to generate semantic pointclouds of the environment. The final part of the paper is focused on pointcloud processing: providing methods for instance extraction and instance processing. To verify the performance of the proposed methodology multiple experiments are conducted. Through the evaluation of the results it is possible to identify possible improvements.