The currently devised solution
Initially the use of direct matching was retained unreliable for real time operation and we worked with Ayache's algorithm for computing stereo from edges (detected by Canny's algorithm). Conversely, currently we are using direct matching to obtain dense depth maps with an efficient implementation of an area-based stereo algorithm (Videt Stereo Algorithm-VSA) based on the SAD (Sum of Absolute Differences) similarity measure. VSA has been particularly optimised and exploits the SIMD capabilities (i.e. MMX technology) provided by the main general purpose processors available nowadays. The algorithm allow for real-time operation of the overall Videt system.
The behaviour of the matching algorithm can be summarized by the following figure.
The stereo camera system we are currently using is the STH-V1 Stereo Head. In the next future we plan to use the higher resolution digital stereo camera MD1. Both cameras are manufactered by Videre Design.
Trinocular stereo - while multiplying costs and computing time - might grant greater robustness. Other depth-measuring devices (e.g. laser range finders) might work better than stereo. They seem to have some drawbacks, at least in the application of interest.
The behaviour of the acquisition and depth calculation algorithm is decribed by the two following examples.
In each of them we can see how, from a picture of the framed scene, is possible to calculate the corresponding disparity map, which summarize the 3D information of the environment around the camera system. With the combined used of the picture and of the disparity map information is possible to build the VRML model of the environment.
|Copyright © 2000, DEIS - University of Bologna. All rights reserved.|