Motion Correspondence Using a Neural Network
G. H. Sarigianidis, David Pycock · 1993
Identifying corresponding features in an image sequence is an important issue in motion analysis. We present a solution based on the assumption of smooth motion using point features. Local constraints are used to make the method robust against occlusion and imperfect feature extraction. A global cost function is defined which is minimised by a mapping of feature points onto a 2-D Hopfield neural network. Three variants of the Hopfield network model are considered. Results obtained using synthetic and natural image data show that this method is robust against occlusion and poor feature detection. 1