Automatic Image Labelling for Deep-Learning-Based Navigation of Agricultural Robots

The aim of this work is to define a procedure for the automatic labeling of images used for the training of a deep neural network which is used to learn a direct mapping from images to steering angles and collision probabilities. A state-of-the-art convolutional neural network for robotic vehicle navigation is used, which has been adapted to work for ground vehicles. Steering angles and collision probabilities are then used to generate respectively the angular and the linear velocity commands to drive a tracked vehicle through rough unstructured terrains, as those typically encountered in agricultural applications.