CNN based Relocalization Method for Indoor Moving Robot with Planar State Constraint

Mobile robots are prone to be kidnapped when they perform autonomous navigation tasks. In this paper, we propose an image based relocalization method for indoor mobile robot after being kidnapped. Robot can be globally relocalized in end- to-end manner by using a single RGB image. Firstly, a deep convolutional regression neural network model of encoder-decoder shape was constructed. The encoder part of the model consists of some pre-trained lightweight residual blocks with less parameters. And, the decoder part consists of some up-convolutional layers introduced to preserve partial detailed information of the input image. Secondly, in order to reduce the model complexity of the regression system and with the fact that a planar moving robot is not a full-degree of-freedom system, we propose a novel state parameterization method based on planar motion of indoor mobile robot. Finally, experimental results on public dataset show that the proposed image based relocalization method improves the accuracy of robot pose prediction to some extent.