Indoor Drone Navigation with Machine Learning and SLAM

Abstract

This work aims at enhancing an indoor drone navigation with the use of Simultaneous Localization and Mapping (SLAM) solutions in combination with convolutional neural networks (CNNs). In particular, the Oriented FAST and Rotated BRIEF (ORB-SLAM) algorithm is employed to undertake visual-based localization and mapping by feature detection and tracking from the environment. However, to improve the accuracy and combat issues related to feature-sparse or ambiguous scenes, a CNN is used for semantic scene analysis, including obstacle and object recognition. Integrating ORB-SLAM for the feature mapping and CNN for the deep sense of vision, it is possible to implement the drone’s autonomous flight in complex indoor conditions, with real-time obstacles’ detection and permanent localization. The combination of machine learning with SLAM improves the dependability and practicality of the general system for different and complex indoor environments