Robust camera motion estimation in presence of large moving objects

Estimation and compensation of the camera motion is the first step in many video analysis applications. Existing robust global motion estimation (GME) techniques have proven to tolerate reasonable amounts of outliers in the data. However, when these outliers convey the motion of large objects, GME remains a major challenge. This paper reviews the main causes that make GME with large objects particularly difficult. Then it proposes an iterative RANSAC-based approach that, by exploiting the properties of the different types of fits that can be found in the data, determines the most suitable scale a-posteriori and can recover the camera motion even when objects are dominant. Evaluation with synthetic and natural sequences demonstrates the good performance of our approach.