Extending the Measurement Error Model of a Direct Visual Odometry Algorithm to Improve its Accuracy for Planetary Rover Navigation

Abstract

In this paper, the accuracy of a direct monocular visual odometry algorithm is improved. The algorithm is able to determine the position and orientation of a robot directly from intensity differences measured at observation points between consecutive images, captured by a monocular camera, rigidly attached to one side of its structure, tilted downwards. The improvement was achieved by extending the stochastic model of the intensity-difference measurement error, from considering only the camera noise, to one that also considers the intensity-difference measurement error due to the 3D shape error between the assumed and the true planetary surface shape. The corresponding covariance matrix was incorporated into a Maximum Likelihood estimator. According to the experimental results on irregular surfaces, where the 3D shape error is usually large, the accuracy of the visual odometry algorithm improved by a factor of 2 but with the cost of increasing the processing time also by the same factor.