Motion Estimation in Varying Illumination using a Total Least Squares Distance Measure

Correlation-based motion estimation algorithms usually match blocks between two frames based on distance measures like correlation, mean square difference or mean absolute difference. Such distance measures, however, also react to temporal changes in illumination, what is not desired when only motion caused by objects in the scene is sought. In this paper we therefore extend an earlier, illumination-invariant approach for binary motion detection in sequences recorded by a static camera towards illumination-invariant motion estimation. The correlation- or difference-based distance measure in standard block matching is replaced by a total-least squares (TLS)-measure which reacts mainly to scene motion. In addition, a regularizing postprocessing algorithm, which combines the TLS-measure with context information, may optionally be applied.