Parallel vision integration on the AMT distributed array processor

This paper defines parallel algorithms for the cooperative combination of data from multiple low-level vision modules. Weak continuity constraints are used to formulate individual early vision modules which detect intensity edges, optic flow from motion, and disparity from stereo. Vision modules are coupled together via line processes which explicitly mark discontinuities in their respective image properties. A Hopfield neural network is used for function minimization with continuation on the sigmoid function to avoid local minima. These algorithms are mapped to the AMT Distributed Array Processor (DAP). Extensive experiments are performed on the DAP to show how vision integration can be used to reduce the impact of noise and speed algorithm convergence