Stereo ORB-SLAM3 Underwater Performance Evaluation

Abstract

Site exploration and navigation missions in hazardous locations are examples of the widespread applications of autonomous mobile robots. Efficient simultaneous localization and mapping (SLAM) is essential for the functionality of such robots. Visual SLAM (VSLAM) has become a cost-effective and attractive solution for all types of vehicles in the field. ORB-SLAM3 and its preceding versions are prevailing visual SLAM algorithms in recent literature. ORB-SLAM3 builds a three-dimensional map based on images and supplies a continuous estimate of the vehicle's pose. This paper demonstrates an experimental evaluation of stereo ORB-SLAM3 performance in scenes with diverse backgrounds. Online ORB-SLAM3 is tested in multiple aerial and underwater environments. The performance of the algorithm is evaluated by comparing against ground-truth estimates provided by offline structure-from-motion techniques. Results show that ORB-SLAM3 can function in structured environments providing acceptable accuracy and satisfactory performance given the right conditions. The algorithm achieves average standard deviation values of 4 cm in underwater scenes depending on the complexity of the followed trajectory.