Recognizing a Moving Object by using Neural Nets andOcular Micro Tremor
This paper describes a neural processor capable of tracking and recognizing an object that moves freely in a 3D space and is visualized through a webcam. Images are preprocessed by using openCV routines in order to obtain crude border detection information. The obtained massive data is delivered to a neural processor composed by two cascaded, independent networks trained at different epochs and with different attitude. The first net specializes in tracking one selected object, whose image changes in position, tilt and scale. Once trained it participates in a close loop control system in which received images directly control eye movements. This combination artificially produces an ocular micro tremor (OMT) similar to the one found in mammals. The micro tremor signals are stored in short term memory elements and become the input to a second net which converges into a single “concept cell”, whose activity determines the presence of the selected object. The method has been tested using real time real world images under rough visual conditions which include complex background, complex objects and variations in scaling, tilt and perspective.