Detection of Good Matching Areas Using Convolutional Neural Networks in Scene Matching-Based Navigation Systems

Abstract

This paper presents a solution for false matching detection in scene matching-based aerial navigation systems. A navigation system that uses normalized cross-correlation to match a captured image with a reference image was designed. While traditional methods rely on statistical indicators to detect false matchings, this research relied on deep learning using Convolutional Neural Network (CNN). A CNN was trained to online predict the probability of a matching result to be true or false. The training dataset of images was constructed depending on the knowledge of where good matching areas are expected to be. The probability numbers were stored as an assistant map to be used again with the same reference map without classification. The system was implemented and tested in a 3D simulation environment using models for a drone, camera, and flight environment. The Robot Operating System (ROS) and the 3D dynamic simulator Gazebo were used for simulation. The results proved the efficiency of the proposed method in excluding the false matchings. Using the assistant map without classification resulted in an execution time of 41ms and RMS error of position less than 1.2m.