The extraction of structural information from images is potentially a very useful tool for a wide variety of practical applications. The difficulty of applying stereo vision techniques to industrial applications varies depending upon the task and source of data. For example, extracting dense depth information from satellite image pairs [ 2 , 8 ] is relatively simple compared with attempting to extract and locate wire frame objects in industrial scenes [ 6 ]. Effects such as photogrammetric variation and object occlusion are more of a problem in the analysis of industrial scenes than of satellite images. Also, tasks such as obstacle avoidance [ 14 ] or hand eye co-ordination [ 9 ] can all be classed as stereo algorithms but can be far easier to engineer compared to techniques requiring accurate calibration. All of the work referenced above demonstrated some degree of success and all of these projects had some level of industrial backing. It is therefore surprising that very little of this technology, even for the simpler problems, has made the last step into practical applications in industry. There may be several reasons for this, the most obvious being the computational requirement for algorithms and the resulting cost of a system, but also there has been a long standing problem of a lack of algorithmic ``robustness'' (either perceived or genuine). As a consequence the true value of stereo algorithms still needs to be conclusively demonstrated.
In order to recover the three dimensional structure of a scene from two disparate views, it is necessary to correctly match regions between the two images: the correspondence problem. In general, approaches to stereo correspondence solving may be classified as either feature based or area based. Feature based techniques such as [ 4 ], where previously extracted discriminant points such as edges are matched between images, may be used to recover reliable although sparse measurements of disparity. Alternatively, area based techniques such as [ 1 , 7 ] attempt to correlate directly the image grey-levels between the two images. Typically providing less reliable estimates of disparity than feature based methods however, they are usually able to recover a much higher density of results.
Techniques have been described to fuse these two sources of disparity as in [ 3 ]. More recently we have described a correlation based algorithm, known as Stretch Correlation (SC), that embodies the constraints imposed by feature based techniques [ 10 ]. Formulating the algorithm as a correlation based matcher provides two major benefits. First the algorithm is presented in a simplified computational framework, something which we have exploited in hardware development [ 12 ]. Second, by limiting data abstraction (maintaining a close coupling between the disparity estimates and the original image data) the algorithm is more amenable to temporal extension, a factor that we have exploited in this work. By including a temporal feedback path in the SC algorithm we aim to show that the robustness of the approach can be significantly improved over the non-iterative method.
Tony Lacey