Bio-inspired flight control and visual search with CNN technology

In this paper it is shown that by building on parallel topographic CNN preprocessing of image flows, efficient terrain exploration and visual navigation algorithms can be developed. The approach combines several channels of nonlinear spatio-temporal feature detectors within an analogic CNN algorithm and produces unique binary maps of salient feature locations. This preprocessing scheme is embedded into a multi-target tracking (MTT) framework where these features are statistically described and assigned to numbered tracks. The MTT output has two distinct roles. First, its feature descriptors drive a classifier based on the adaptive-resonance theory (ART), which is also implemented on CNN architecture. Second, it provides an optical flow ("target displacement") estimate to the navigation system, which in turn calculates the flight control parameters (Yaw-Pitch-Roll). An upper level visual attention and selection mechanism uses both the feature descriptors and the optical flow estimates to automatically adjust the focus and scale (zoom) during navigation. The paper describes the architecture and the algorithmic frameworks and provides the first experimental results on aerial video-flows.