Neighbour Selection

When estimating the velocity field at a pixel \(P_0\) we localize the four nearest neighbors \(P_x^+\), \(P_x^-\), \(P_y^+\) and \(P_y^-\). Then each combination of horizontal and vertical neighbors will lead to a velocity estimate, indicated as \(\wh{v}^{++}\), \(\wh{w}^{++}\), etc. The resulting velocity estimate is taken as an average of all found velocities. Some differences may occur if the reference pixel \(P_0\) is an edge or a corner pixel, in which case only two and one combination of pixels are available.

The selection based algorithm can be customized with the options NeighbourOptions are given in NeighbourOptions. Most importantly: * ccf_min_lag is used to avoid using data when the maximum correlation time is comparable to the sampling time.

Checks that the maximal correlation between adjacent pixels occurs at a time larger or equal than neighbors_ccf_min_lag multiples of the discretization time. If that’s not the case, the next neighbor will be used, and so on until a neighbor pixel is found compliant to this condition. If set to 0, it only checks that the pixel is not dead.

*min_separation minimum allowed separation between pixels. *max_separation maximum allowed separation between pixels. If some

condition is required (such as ccf_min_lag) and not fulfilled for pixels closer or at than max_separation, then no neighbors will be used and the subset of pixels under process will yield no estimate. The condition applies on a closed interval, meaning that pixels separated exactly max_separation are allowed, thus setting min_separation = max_separation guarantees that the neighbour pixels will be separated by that distance.