An image motion estimation technique based on a combined statistical test and spatiotemporal generalised likelihood ratio approach

We present a method based on Kalman filtering, for image motion estimation. Within Kalman formalism, a motion boundary can be modelled as a jump in the evolution equation of the filter. The detection of such a jump relies on a χ2 statistical test applied to the innovation signal. The optimal estimation of the jump parameters and the compensation of the current estimate are performed using a General Likelihood Ratio (GLR) algorithm. To exploit the spatial redundancy inherent to a motion boundary, the original GLR algorithm is reformulated by integrating spatiotemporal motion information. This results in a significant decrease of the compensation delay.