A Genetic Algorithm with Online Learning Approach for Improving Loop Closure Detection of a Visual SLAM

Abstract

This paper presents a genetic algorithm with online learning approach for improving loop closure detection of a visual simultaneous localisation and mapping (SLAM) technique. The reality gap issue in evolutionary robotics field is known as the main factor that degrades the quality of simulated solutions when transferring to a real robot. The proposed method can optimise the parameter of a visual SLAM in real-time. The aim is to evolve and search for the best Bayes filter parameters of the loop closure detection using online data gathered directly from a connected robot. A fitness function calculation utilising real-time images and robot motion is proposed to evaluate the performance of candidate solutions throughout the learning session. The experimental results show that SLAM with the optimised loop closure detection parameters outperforms SLAM with the default parameters for about 90% improvement.