Edge recognition in dynamic vision

Using a model of an edge´s motion through a sequence of images, the problem of its localization can be formulated as a stochastic filtering problem. The extended Kalman filter for such a system is considered in detail and is shown to be interpretable as a sequence of oriented special convolutions. Results are presented which show that the edge localization obtained using this filter is substantially better than that obtained using either the Sobel or Canny edge operators on each image individually