Robust estimation of camera motion in MPEG domain

In many applications requiring video processing for dynamic scene analysis such as object tracking in computer vision and motion-based automatic video indexing and retrieval, it is necessary to separate camera motion from object motions. This paper presents a robust least median-of-squares (LMedS) estimate of the camera motion using motion vector field of P-frames in an MPEG video sequence. A simplified 4-parameter affine flow model is used to describe the camera global motion and the model parameters are estimated using an LMedS approach to minimise the influence of the outliers (due to object motion and wrongly predicted block motion vectors) in the MPEG motion vector field. The 0.5 breakdown point of the LMedS technique is mostly avoided in this paper by prefiltering of the motion vector field to discard wrongly predicted motion vectors. The result produced by the proposed technique shows a significant improvement in the estimation of the camera motion model parameters compared to those produced by the M-estimator technique.