Parallel Simulation of an Active Vision Model

We describe the implementation of a biologically motivated model of the human visual system (HVS) on the Paragon XP/S. The approach involves the simulation of the two major pathways (magno-(M), and par-vocellular (P)) of processing visual stimuli including (i) retinal sampling, (ii) active foveation and (iii) object recognition. Theory suggests a vision model based on the competition of active processes governing foveation. The visual information entering the system is processed by separate subsystems in parallel. The output of M and P type sampling grids is fed to the 'where' and 'what' modules respectively, each operating independently. The 'where' channel is responsible for object localization and computes the next center of gaze in a data driven/bottom-up fashion. The recognition module matches the current input against previously stored memories and veriies the presence of an object in a top-down fashion. The next xation point in each step is computed dynamically by the control unit using the outputs of these modules. In this paper we focus on the implementation details of the HVS model described here brieey. Speed-up curves, timing data and asymptotic analysis are presented.