Vision system for autonomous navigation

Abstract

This paper addresses the problem of vision based navigation in structured environments using homography. The approach makes a Manhattan World assumption and uses vanishing point to extract the approximate floor and wall planes. This preliminary segmentation of the image helps in the visual odometry pipeline by providing an approximate hypothesis for planes which are tracked over multiple frames to calculate motion trajectory of the camera. RanSaC based approach is used to calculate the vanishing point. A similar approach is used to calculate homography by tracking key-point over consecutive frames. The obtained homography model is compared with the homography models of planes detected through Vanishing Point based segmentation of image of a structured environment into its planar elements. This comparison provided with information about the camera motion and pose. The proposed method was applied to a dataset and the results compared against the ground truth.