Abstract: Taking the grayscale-time-space tensor descriptor (GTSTD) as a feature descriptor of image sequence, a matching optical flow computing method was presented based on Riemannian metric of tensor. Riemannian metric of tensor was used to measure the distance of features, while an improved Hausdorff distance was used to replace traditional Euclidean distance in the computing Riemannian metric, forming a correlation function of image match to enhance the matching ability of the algorithm in the case of noise and occlusion. Based on all above, the matching optical flow computing algorithm was given out. Simulation results show that this method has advantage on accuracy and noise immunity compared with method based on differential algorithm (H-S, L-K) and block matching algorithm based on grayscale.