Approach to Obstacle Localization for Robot Navigation in Agricultural Territories

摘要

Search and localization of obstacles is one of the main tasks in path planning for robotic systems. In this paper, an approach to obstacle localization for robot navigation in agricultural territories is proposed. The developed approach is based on a combination of calculation of Normalized Difference Vegetation Index (NDVI) and artificial neural network (ANN). The NDVI is used to detect obstacles: buildings, stones, garbage and the Convolutional Neural Network (CNN) is intended to search other obstacles: trees and vegetation. This separation allowed to reduce the amount of data necessary for CNN training to one data class. The result of the presented approach is a binary map, which shows passable and non-passable areas for robots. The total accuracy of obstacle detection using proposed approach ranges from 56 to 90% of the whole area, occupied by obstacles, on image.