Time-efficient dense visual 12-DoF state estimator using RGB-D camera

Abstract

This paper presents a fast RGB-D dense visual odometry estimating 12-DoF state information including 3D motion and 6-DoF spatial velocity of a camera-strapdown system. To reduce computational loads, we extract informative pixels through a zero-crossing difference of Gaussian (DoG) and non-maximum gradient pixel extraction. For extracted regions, the 3D motion is estimated through inverse compositional algorithm and the result of motion estimation is exploited to calculate 6-DoF spatial velocity of the camera. Additionally, we relieve noise in the raw velocity using the Kalman filter. Afterwards, we validate the proposed algorithm using TUM RGB-D datasets and simulation results are reported. Our algorithm not only presents similar performances with the popular dense visual odometry, DVO, but also runs up to 2 times faster than DVO. © 2017 IEEE.