Feature Descriptor based on Differential Evolution for Visual Navigation in Dynamic Environments

Abstract

This work proposes to use a teach-and-repeat method to navigate in an outdoor environment. Instead of using state-of-the-art feature point descriptors to match the current images to the database, we rely on the Generated Binary Robust Independent Elementary Features (GRIEF), a model robust to changes in illumination and weather. These descriptors are obtained via a training process employing evolutionary methods and binary comparison tests. The objective is to obtain visual feature descriptors designed for a given task. For the teach-and-repeat task, they are trained to optimize the estimate of the robot heading error with respect to a learned path. In this work, we propose to investigate how Differential Evolution (DE) can improve the performance of the GRIEF in terms of the convergence rate during the training phase and of fitness score. The new approach, named GRIEF-DE, is trained with the Michigan data set, made of pictures of an outdoor environment with significant changes in appearance caused by seasonal weather variations and illumination, and compared with the original GRIEF. The proposed model was applied to 4 different data sets and the experimental results suggest that the use of Differential Evolution improves the training performance as well as the estimation of the robot orientation with respect to the path of reference.