HS method is one of the most classical methods for optical flow computation.In practice the variation of gray level between two points is nonlinear

while in the original HS method

it is assumed to be linear

which leads to the inaccuracy of the convergent point of LSM iterations.In the 2nd section

we discuss the reason why the classic method isn't accurate enough in detail.And in the 3rd section we present a new algorithm to improve accuracy of gray gradient by using the offset

which is generated in the previous cycle

to produce the new gradient.The following experiments

including synthesized image sequences and images sequences extracted from real video

compare the original algorithms with the improved one.The objects in the synthesized sequences involve some stereotype movements such as translation

rotation

and zooming.The results show that the veracity of optical flow field has been enhanced

and the convergence speed is accelerated to a great extend.