Neural networks implementation of the visual information processing for an intelligent aerial vehicle

This paper describes a new way of the visual information processing for small intelligent aerial vehicles using neural networks, and its VLSI implementation. The vision using neural networks is for the navigation to avoid obstacles, based on the 64x64 image. The area in the front of flying is mapped to sixteen sectors, and an appropriate path is selected by the trained neural networks. The multilayered network with two hidden layers is employed, and the empirical rules are added to the training process. Pre-processing of sectored image data is proved to improve the cognition performance remarkably, by its localised attention. The size of neural networks is minimised for a single VLSI device implementation, less than 4,000 synaptic connections. The proposed vision system demonstrates the desired behaviour with real environmental scenes after the training by the simulated data. The feasibility of neuromorphic VLSI implementation is demonstrated by the neural network with reduced synaptic weight accuracy, which represents the measured accuracy of biologically plausible conductance based CMOS neural network VLSL The proposed neural networks implementation demonstrates the feasibility of an intelligent vision system of small uninhabited aerial vehicles, for its VLSI integration and low power consumption