Evaluation of the visual odometry methods for semi-dense real-time

TL;DR

This paper evaluates and compares popular visual odometry methods on multiple datasets, and proposes a real-time semi-dense stereo VO approach combining feature-based and feature-less techniques for improved accuracy.

Contribution

It provides a comprehensive evaluation of LSD-SLAM and ORB-SLAM2, and introduces a novel semi-dense real-time stereo VO method combining direct and indirect approaches.

Findings

LSD-SLAM and ORB-SLAM2 performance varies across datasets.

The proposed semi-dense stereo VO achieves real-time operation with high accuracy.

Graphical comparison shows the effectiveness of the combined approach.

Abstract

Recent decades have witnessed a significant increase in the use of visual odometry(VO) in the computer vision area. It has also been used in varieties of robotic applications, for example on the Mars Exploration Rovers. This paper, firstly, discusses two popular existing visual odometry approaches, namely LSD-SLAM and ORB-SLAM2 to improve the performance metrics of visual SLAM systems using Umeyama Method. We carefully evaluate the methods referred to above on three different well-known KITTI datasets, EuRoC MAV dataset, and TUM RGB-D dataset to obtain the best results and graphically compare the results to evaluation metrics from different visual odometry approaches. Secondly, we propose an approach running in real-time with a stereo camera, which combines an existing feature-based (indirect) method and an existing feature-less (direct) method matching with accurate semidense direct image alignment and reconstructing an accurate 3D environment directly on pixels that have image gradient. Keywords VO, performance metrics, Umeyama Method, feature-based method, feature-less method & semi-dense real-time.