Optical Flow Group-Parameter Reconstruction From Multi-Channel Image Sequences

We propose a novel algorithm for reconstruction of optical flow group transformation parameters from sequences of multi-channel images. Our method is based on "early integration" paradigm using all of the available spectral components and the vector field representation of the transformation group generators. This way the reconstructed flow carries the total information contained in the images and the singularity of the inverse problem is potentially reduced. The method avoids the more complex and ambiguous task of first reconstructing the local vector field and subsequently fitting the group transformation templates. Remaining singularity of the structural tensor is removed by a modified Tikhonov-type of regularization. The algorithm is quantitatively validated with recorded images transformed with generated vector fields that are then compared with the reconstructed optical flow. The dependence of reconstruction accuracy on both the parameters of the images and the magnitudes of the vector deformation fields is presented. We also show the application of the method to the real-world task of video-based detection of convulsive epileptic seizures and compare the output to the previously published results using standard optical flow algorithm.