Date of Award


Degree Type


Degree Name

Master of Science


Faculty of Computing, Health and Science

First Advisor

Dr. James Cooper


Estimates of optical flow in images can be made by applying a complex periodic transform to the images and tracking the movement of points of constant phase in the complex output. This approach however suffers from the problem that filters of large width give information only about broad scale image features, whilst those of small spatial extent (high resolution) cannot track fast motion, which causes a feature to move a distance that is large compared to the filter-size. A method is presented in which the flow is measured at different scales, using a series of complex filters of decreasing width. The largest filter is used to give a large scale flow estimate at each image point. Estimates at smaller scales are then carried out by using the previous result as an a priori estimate. Rather than comparing the same region in different images in order to estimate flow, the regions to be compared are displaced from one another by an amount given by the most recent previous flow estimate. This results in an estimate of flow relative to the earlier estimate. The two estimates are then added together to give a new estimate of the absolute displacement. The process is repeated at successively smaller scales. The method can therefore detect small local velocity variations superimposed on the broad scale flow, even where the magnitude of the absolute displacement is larger than the scope of the smaller filters. Without the assistance of the earlier estimates in ‘tuning' the smaller filters in this manner, a smaller filter could fail to capture these velocity variations, because the absolute displacement carry the feature out of range of-the filter during successive frames. The output of the method is a series of scale-dependent flow fields corresponding to different scales, reflecting the fact that motion in the real world is a scale-dependent quantity. Application of the method to some 1 dimensional test images gives good results, with realistic flow values that could be used as an aid to segmentation. Some synthetic 2-dimentional images containing only a small number of well defined features aIso yield good-results but the method performs poorly on a random-dot stereogram and on a real-world test image pair selected from the Hamburg Taxi sequence.