Estimating Epipolar Geometry With The Use of a Camera Mounted Orientation Sensor

Abstract

Context: Image processing and computer vision are rapidly becoming more and more commonplace, and the amount of information about a scene, such as 3D geometry, that can be obtained from an image, or multiple images of the scene is steadily increasing due to increasing resolutions and availability of imaging sensors, and an active research community. In parallel, advances in hardware design and manufacturing are allowing for devices such as gyroscopes, accelerometers and magnetometers and GPS receivers to be included alongside imaging devices at a consumer level.

Aims: This work aims to investigate the use of orientation sensors in the field of computer vision as sources of data to aid with image processing and the determination of a scene’s geometry, in particular, the epipolar geometry of a pair of images - and devises a hybrid methodology from two sets of previous works in order to exploit the information available from orientation sensors alongside data gathered from image processing techniques.

Method: A readily available consumer-level orientation sensor was used alongside a digital camera to capture images of a set of scenes and record the orientation of the camera. The fundamental matrix of these pairs of images was calculated using a variety of techniques - both incorporating data from the orientation sensor and excluding its use

Results: Some methodologies could not produce an acceptable result for the Fundamental Matrix on certain image pairs, however, a method described in the literature that used an orientation sensor always produced a result - however in cases where the hybrid or purely computer vision methods also produced a result - this was found to be the least accurate.

Conclusion: Results from this work show that the use of an orientation sensor to capture information alongside an imaging device can be used to improve both the accuracy and reliability of calculations of the scene’s geometry - however noise from the orientation sensor can limit this accuracy and further research would be needed to determine the magnitude of this problem and methods of mitigation.