A review on vision-based motion estimation

TL;DR

The paper addresses the challenge of vision-based motion estimation for structural monitoring by highlighting the trade-off between accuracy and robustness across existing methods. It proposes a Gaussian kernel-based motion measurement inspired by 3D Gaussian splatting, representing images with 2D Gaussian kernels and estimating motion from kernel-position changes using a composite loss that enforces data fidelity, cross-frame consistency, and smoothness. A preliminary synthetic study demonstrates sub-pixel motion estimation with average relative errors on the order of a few percent for 0.01-pixel motions, indicating potential robustness and accuracy benefits. The authors discuss future work on optimizing performance on more complex images and achieving repeatable kernel initialization, including training a neural network for parameter initialization to enhance repeatability and practicality in real-world applications.

Abstract

Compared to contact sensors-based motion measurement, vision-based motion measurement has advantages of low cost and high efficiency and have been under active development in the past decades. This paper provides a review on existing motion measurement methods. In addition to the development of each branch of vision-based motion measurement methods, this paper also discussed the advantages and disadvantages of existing methods. Based on this discussion, it was identified that existing methods have a common limitation in optimally balancing accuracy and robustness. To address issue, we developed the Gaussian kernel-based motion measurement method. Preliminary study shows that the developed method can achieve high accuracy on simple synthesized images.

Figures (2)

• Figure 1: The synthetic image for validation
• Figure 2: Motion field for validation cases (Case1 - Case5, scaled by 500 times)