In a sequence of images, we have two main methods for estimating the displacement of moving objects ; Matching method, and Spatio-temporal gradient method. The latter method is advantageous in real-time application, but there are two problems. One is the segmentation, and the other is that the most estimators become biased or unstable by the noise.
In this paper, we have solved the former problem with a filter and averaging segmentation. By using a filter did we get high SNR and strong consistency. It is found also that the averaging segmentation algorithm improves the separation of moving area and background.
For latter problem of bias and instability, a new algorithm based on "Errors in variables" regression model is developed. It is found to be very stable against high noise. With these algorithms, the real displacements in the noisy image are correctly estimated even for the SNR of 6 dB.