A novel method for motion estimation from first order derivatives is presented. First estimates are obtained by evaluating the minors of the so-called structure tensor that contains blurred products of first order derivatives. The minors yield four different estimates that are equal in case of translation but differ for other spatio-temporal patterns. The mean yields a robust motion estimate and the difference is used as an indicator of discontinuous motions, occlusions, and noise. This procedure leads to a flow field with small errors and low density. An additional change detection is performed to obtain a mask that is filled with the previously computed (correct but sparse) motion vectors. The superior performance of the algorithm is demonstrated on synthetic and real sequences by comparison with other methods.
Full paper in PDF format
|Synthetic sequence with noise|
|Traffic scene||Please consider only the underlying scene.|