Video Stream Relocalization With Deep Learning

Abstract

This paper presents a six degree of freedom regression system using convolution neural network(CNN) and long and short term memory network(LSTM) with video stream as network inputs. The system trains the network to regress the 6-DOF robot pose in a transfer learning and end-to-end manner with little training data. Relocalization only using CNN ignore the temporal correlation between image-sequences. In fact, the robot can easily collect continuous image-sequences. Therefore, in this paper, the robot can regress to the 6-DOF pose according to continuous images of different step sizes. Compared with relocalization with a single image, the experimental results show that the network model has the best effect of relocalization when the step size is set to 10 in the indoor scene, and the error of relocalization is minimal.