A Neural Network-Based Approach to Motion Estimation with Discontinuities

A new neural network-based approach is proposed to estimate motion hierarchy in image sequences taking into consideration motion discontinuities. The network consists in an input layer, an intermediate layer and an output layer. In order to estimate the most likely displacement at each pixel, we have transposed the block matching approach into the neural network approach and add mechanisms to detect motion discontinuities. Information redundancy allows for parallel processing in view of real-time complex motion estimation tasks. Preliminary tests on synthetic and real images are very promising.