Evaluation of visual slam methods in usar applications using ros/gazebo simulation

Abstract

The problem of determining the position of a robot and at the same time building the map of the environment is referred to as SLAM. A SLAM system generally outputs the estimated trajectory (a sequence of poses) and the map. In practice, it is hard to obtain ground-truth for the map; hence, only trajectory ground-truth is considered. There are various works that provide datasets to evaluate SLAM algorithms in different scenarios including sensor configurations, robots, and environments. Dataset collection in a real-world environment is a complicated task, which requires an elaborate sensor and robot configuration. Different SLAM systems demand various sensors resulting in the problem of finding an appropriate dataset for their evaluation. Thus, in this paper, a solution that is based on ROS/Gazebo simulations is proposed. Two indoor environments with flat and uneven terrain to evaluate laser range and visual SLAM systems are created. Changing the sensor configuration and the environment does not require an elaborate setup. The results of the evaluation for two popular SLAM methods—ORB-SLAM2 and RTAB-Map—are presented.