Input Redundancy and Output Observability in the Analysis of Visual Motion

Sixth International Symposium on Robotics Research, MIT Press, 1993 Determining structure and motion from the images produced by a camera on a moving robot is a nonlinear and potentially poorly conditioned computation. A reliable system must use redundant data so that small random errors in the inputs cancel out statistically to produce better outputs. Furthermore, the output quantities to be computed must be well observable, that is, they should have well-measurable e ects on the images. I demonstrate the importance of these two principles with three system for the analysis of visual motion: the factorization method for structure and motion under orthography, a system to compute camera motion from narrowly spaced frames, and a global, multiframe, and multifeature method for the reconstruction of structure and motion from sequences taken under perspective.