Abstract: The representation of trajectories for moving edges bya nonlinear dynamical system is presented as atechnique for the integration of motion and binoculardisparity. The integration is accomplished through themonolithic use of dynamical systems as thecomputational mechanism. In this approach, the motionof a simple feature through a small region of the imageis modeled by a dynamical system with three variables.Visual processing by the model dynamics occurs within aretinotopic array of identical dynamical systems withnearest neighbor coupling. A form of segmentation,called feature linking, is accomplished by theconvergence of neighboring array elements to the samestable mode of coherent oscillations involving onlythose dynamical systems that respond to theco-occurrence of the same feature and motion. Eachmodel dynamic for a moving feature is driven byconcurrent inputs from a pair of Gabor filters in phasequadrature. The stable modes of the model dynamicsconsist of limit cycles with nearly linear phase. Theco-occurrence of a particular feature and velocity ofmotion in the image sequence optimally selects aparticular limit cycle. The integration of motion andbinocular stereopsis is achieved through the use of thelinear phase property of the limit cycles to recoverthe positional disparity between similar featuresmoving through corresponding regions in the retinotopicarrays for the left and right images. Bifurcationtheory is applied to rigorously describe theconvergence properties of the dynamical systems. Thefocus of this paper concerns the mapping of visualmotion onto analog dynamical systems as an initial steptoward a theory of low-level integration of motion andbinocular stereopsis. !23
展开▼
