Analysis of Estimation Errors in Relation to Object Motion

Estimating three-dimensional (3D) motion parameters from image sequences is fundamental in robotics, navigation, and computer vision. Motion estimation techniques include both linear and non-linear approaches that establish point correspondences across frames. Non-linear methods improve accuracy through iterative refinement but require reliable initialization and are computationally demanding. In contrast, linear methods are computationally efficient yet prone to noise and degenerate cases. This study systematically analyzes the dependence of motion estimation errors on object motion characteristics in single-camera image sequences. The results show that essential matrix methods fail under pure rotation but perform effectively in depth-varying environments such as SLAM. Homography-based methods excel in planar motion scenarios such as AR/VR tracking and image stitching but degrade in depth-variant scenes. Non-linear methods achieve higher accuracy overall, though at the cost of initialization sensitivity and higher computational requirements. These findings provide practical guidance for selecting motion estimation models in robotics, autonomous systems, and photogrammetry, supporting researchers and engineers in choosing appropriate techniques under specific application constraints.