Assessment of Map Construction in vSLAM

Vision-based Simultaneous Localization and Mapping (vSLAM) is a challenging task in modern computer vision. vSLAM is particularly important as mobile robotics application. It allows to localize the robot and build the map of unknown environment in 3D in real-time. During research and development of new methods, it needs extensive evaluation on trajectory and map quality compared to known methods. In this work we focus on map quality estimation. We develop the simulated ground-truth data in photo-realistic environment and introduce new metrics in order to estimate map quality. We evaluate neural network based vSLAM methods with our framework in order to show that it fits map quality estimation more than standard approaches. Open-source implementation of our map metrics is available at https://github.com/CnnDepth/slam_comparison.

I. Introduction

Nowadays, vision-based SLAM is an important task in modern computer vision and mobile robotics. There are several methods, based on different visual sensors that are able to solve the problem of localization and mapping [1] [2] [3] [4] [5] [6]. Usage of such sensors is important for autonomous navigation in GPS-denied environment. Current methods are even suitable to perform real-time on board of mobile robot [7] [8].