Topographic and non-topographic neural network based computational platform for UAV applications

Abstract

In this work, we present an architecture and algorithmic framework where topographic and non-topographic computation is combined on the basis of several artificial neural network models. The algorithm cores utilize an analogic (analog and logical) architecture consisting of a high resolution optical sensor, a low resolution cellular sensor-processor (cellular nonlinear network-CNN-based chip) and a digital signal processor. The proposed framework makes the acquisition of a spatially and temporally consistent image flow possible even in case of extreme variations in the environment. It ideally supports the handling of difficult problems on a moving platform such as terrain identification, navigation parameter estimation and multi-target tracking. The proposed spatio-temporal adaptation relies on a feature based optical flow estimation that can be efficiently calculated on available CNN chips. The paper illustrate how multi-channel visual flow analysis and classifier (ART, KN) driven visual attention-selection mechanisms can be efficiently supported by an analogic architecture. The experiments performed on an analogic CNN hardware prototype highlights some of the application potentials for unmanned air vehicle (UAV) applications.