Visual SLAM in Action: Comparative Study

Abstract

Autonomous vehicles have become one of the state-of-the-art fields in today’s research and started to be crucial with smaller levels for everyday usage. One of the main challenges towards moving toward higher levels of autonomous driving is the real-time localization and mapping performance with regular or highspeed driving. SLAM is considered to be the solution for these high-speed vehicles to explore any new environment and to achieve the required localization and mapping performance in this unknown environment. In this paper, we did a comparison between two major VSALM techniques in this area, which are DROID SLAM, which uses deep neural network architecture, and ORB-SLAM3, which utilizes crafted ORB (Oriented FAST and Rotated BRIEF) features and tracking algorithms. For both techniques, realtime datasets recorded from the technology building of the University of Bridgeport environment are utilized to compare their performance regarding mapping, localization, and time complexity. Results show several advantages and disadvantages for both algorithms, which are highlighted and explained over the proposed work.