Ground truth Data Collection for Visual SLAM Methods for Bench-marking Localization and Loop Closing Methods

Abstract

Visual SLAM has the potential of offering the most accurate method for localization of a moving platform. There is a recent interest in fusing classical SLAM methods with the deep networks. This fusion demands big data. In this work, we propose to use the recent realistic simulators, e.g. MINOS etc., to dig out big data for visual slam methods. Our focus is to provide ground truth for the both 6 DOF pose and loop closing methods. We also purpose to find the absolute trajectory error of the sequence with ORBSLAM trajectory. The SLAM is latest platform to find out pose of camera, we can compare it frame by frame with original pose given as ground truth. Using MINOS has many advantages, like we don’t need to go to physical location to gather data and we can close as many loops as we wanted. There are many advantages to collect data from simulator, as simulator is update-able and would support different environments and lightning conditions. We can gather as much data as needed, and we can expect that more houses would be added by authors, or there may come update that we can our own sequence. The need for data is increasing day by day, as in machine learning problems data is main issue. ORBSLAM claims to do better tracking, we are bench-marking its performance on data collected from MINOS simulator. We would gather different sequences from multiple scenes available on MINOS. The main goal is to compare the ground truth provided by MINOS with the pose abstracted by SLAM methods and to calculate Absolute Trajectory Error. If error is big then SLAM is not doing well as we have tested our ground truth by Fundamental Matrix check. Fundamental Matrix test is used to evaluate the ground truth of any system. We are providing data so that people can improve ORBSLAM tracking, to plan a short path to find an object in scene, to improve ATE. This data could also be used in machine learning domain in different robotic applications. We can also use this robot path to train a network and test it on real robot. There are many applications of this data set, as there are many other data sets as well but they have some limitations. The limitations include less number of sequences, less number of loop closure etc.