Correcting velocity measurements by tracking of linear features

This paper describes an approach to accurate velocity estimation of objects in image sequences. It combines feature-based and non-feature-based velocity estimation methods, in order to overcome the limitations that both techniques exhibit when used separately. It is well known that popular non-feature-based methods, such as optical flow and correlation methods, suffer from the problem of inaccuracy in the regions around the boundaries of objects. This is particularly unfortunate, since it is precisely those areas which tend to contain the largest spatial gradients and therefore make the largest contributions to velocity estimates. Furthermore, in many applications of velocity estimates, such as video coding, the accuracy of the estimates in the region of object boundaries can have a large impact on the performance of the application. On the other hand, estimates derived purely from image features can result in very accurate velocity estimates for object boundaries. The feature-based methods, however, suffer from the “correspondence problem”: identifying which feature in one frame corresponds to a given feature in the next. The method described in this paper overcomes the correspondence problem and at same time eliminates most of the errors due to a correlation estimate by using feature tracking that is assisted by, and makes corrections to, the velocity predictions made by the correlation