Simple and robust correlation based visual compass using omnidirectional view

Abstract

In many mobile robotic tasks, it is important to know the local axis of the robot at any pose with respect to a reference or global direction. In this paper, a vision-based algorithm is described for this purpose. A new approach to form the omnidirectional view of an environment is developed using a single camera. The images captured at a definite interval of angles are stitched by a simple method of removing the overlapping area between two successive images. To find the local axis of a mobile robot, a robust technique is developed to determine the orientation of the robot at any pose with respect to the reference direction in an unknown environment. A reference omnidirectional image is correlated with that at the current position. The paper formulates two algorithms : one for finding the fixed amount of overlapping area and another for finding the angle of local axis with respect to the reference using correlation. The paper also proposes two ways using which this visual compass can be used.