Structure tensor-based motion field classification and optical flow estimation

In this paper, an accurate 3D structure tensor-based method is proposed for optical flow estimation, from which the issues of multiple motions and motion field classification are effectively addressed in an unified way. The novelties of our approach are: (1) scale-adaptive spatio-temporal filter, (2) unequally weighted structure tensor, and (3) confidence measurements. By utilizing spatio-temporal Gaussian filter, multiple motions of moving objects are matched under the automatically selected scale which is steered by the condition number. To capture the local spatio-temporal structure, unequal weighting of the structure tensors is attempted. A new normalization function is developed to threshold the confidence measurements for distinguishing different motion fields. As compared with that of the existing methods, our experimental results demonstrate much improved accuracy both on motion field classification and on optical flow estimation.