Camera Pose Estimation: What is the computationally efficient approach?

Abstract

In this paper we present a system that is suitable for real time camera pose estimation over a video stream, Our approach proves to be computationally efficient for real time applications to calculate the new extrinsic parameters for a moving camera. The system depends only on the camera without the aid of any external sensors or fixed points. Our analysis for the problem shows that the most computationally intensive task is in the feature detectors/descriptors used, for that we will be comparing between 2 feature detectors/descriptors; AKAZE and ORB, that are well known for their good performance in terms of computational cost. Although AKAZE has managed to achieve better results, it comes with higher computational cost yet with substantial gain in terms of quality, this will provide more robust results for our system if compared to ORB. On the other hand, ORB still manages to give enough quality to satisfy our needs and with real-time performance.