Abstract
The automatic estimation of a cameras position based on visual measurements is a general problem in the field of computer vision. Based on the estimated cameras trajectory it is possible to solve common tasks, such as Visual Odometry (VO) in the field of mobile robotics or the automatic reconstruction of an observed scene, based on classical Structure-from-Motion (SfM) techniques. The general procedure of camera egomotion estimation is always based on visual feature tracking and subsequent Perspective-n-Point (PnP) camera pose determination. This article evaluates recent algorithms for camera egomotion estimation based on point feature correspondences for their applicability in VO applications. These algorithms use methods based on 2D/2D and 3D/2D correspondences and are assessed in experimental evaluations employing synthetic data sets. It was found that the accuracy of the evaluated techniques is predominantly influenced by the number of correspondences and underlying motion patterns. Additional routines such as outlier handling and key frame detection were found to be mandatory for real-world application.