Aerial compact drone autonomous navigation system – the model and training algorithms
Вячеслав Васильевич Москаленко, Alyona S. Moskalenko, A. G. Korobov, M.O. Zaretskiy · Bionics of Intelligence · 2018
The paper presents a novel model of convolutional neural network for visual feature extraction, extreme-learning machine for position displacement estimation and boosted information-extreme classifier for obstacle prediction with new training algorithms to build decision rules of autonomous navigation system for compact drones are presented in the paper. Growing sparse-coding neural gas algorithm for unsupervised training of the convolution filters, supervised learning algorithms for training decision rules and simulated annealing search algorithm for fine tuning are proposed. The complex criterion for choosing parameter of feature extractor model is considered. Simulation results with optimal model on open KITTI-datasets confirm the suitability of proposed algorithms for practical usage.