We have suggested the use of temporal information as a means of improving the robustness of a stereo matching algorithm. By propagating disparities from frame to frame it is possible to reduce the probability of mismatch by minimising the search area.
From the sequences of data that we have processed, the stochastic searching would appear to be unnecessary as the motion present in the scenes can often provide sufficient information to overcome ambiguities and mismatches. However, by building a stochastic search into the algorithm, we ensure that there is some mechanism, although crude, to overcome the problems of local minima, as demonstrated on the static frame analysis.
Using the temporal approach we have been able to reduce per-frame computational load when compared with the standard algorithm, typically accelerating the execution four-fold. This is because the temporal algorithm enables a directed search to be performed, centred at the most probable feature location, thus enabling the search range to be drastically reduced.
We would like to see temporal stereo research moving towards the analysis of the type of difficult stereo problem exemplified by the plant, figure 4(c) .
If you would like to evaluate our stereo and temporal stereo algorithms, Sun SPARC binaries together with demonstration data are available from:
http://www.shef.ac.uk/~eee/esg/research/tina.html
Tony Lacey