Stereo Vision Research: An Algorithm Survey

1 Introduction The purpose of this document is to provide a short review of some of the key publications in the area of stereo vision research. It is not intended as a full tutorial on stereo vision as this can be found elsewhere in standard texts on machine vision. Rather, it is intended to be a summary of the important research issues and approaches that researchers have taken and how these techniques are related. We hope that this document will be of value to people who may need initial guidence to the vast quantity of literature that exists on this subject. 2 Overview Determination of three dimensional data from images is of central importance in the eld of machine vision. One of the most direct way of achieving this from image data is stereo vision. Stereo vision has a wide range of potential application areas including; three dimensional map building data visualisation and robot pick and place. A variety of constraints may be used to guide the correspondance solution depending upon the properties of the data. This is re BLOCKINected in the broad range of algorithms that have been developed. For example, if camera calibration is available epi-polar constraints can be used. The absence of transparent objects allows the use of disparity gradient limits. The absence of occlusion can permit strong surface smoothness constraints. If the images are generated under constrained lighting conditions the images may display photo-metric properties allowing direct pixel matching. All of these factors can have a strong in BLOCKINuence on the quantity and reliability of data recovered by an algorithm. Consequently, it is not fair to expect any one algorithm to be capable of making the best of an arbitrary set of data. In 1988 a survey of 15 institutions [7] observed that, in practice, many researchers were using combinations of the above constraints for solving a wide variety of stereo matching problems. The conclusions from this survey were that each of the dierent approaches had its relative merits and disadvantages dependent on the nature of the matching problem. This sentiment has also been reiterated by Hannah [9] who suggested that a complete stereo solution would need to combine the relevant approaches in a cooperative fashion. A practical example of a combined approach is perhaps best demonstrated by Baker [1] who successfully combined both edge and luminance based stereo techniques. Bearing this in mind, the following …