Low-latency Feature-based Front-end with Good Point Matching

This work fills the middle ground with the good feature enhancement applied to feature-based VO/VSLAM. Inspired by the observation that not all the feature matchings contribute to the accurate & robust estimation of camera pose, we propose a family of efficient algorithms that identify small subset of features with most value towards pose estimation, a.k.a. good features.

When only working with the good features in pose tracking, much lower latency is achieved for feature-based VO/VSLAM, while the accuracy & robustness from robust, long baseline feature matching are preserved.

The Good Feature enhanced version of ORB-SLAM2 is now available at: https://github.com/ivalab/gf_orb_slam2

Figures generated using full evaluation results can be accessed at: https://github.com/ivalab/FullResults_GoodFeature

Robust Feature-based Front-end with Good Line Cutting

This work describes a solution to minimizing the impact of 3D line triangulation error in pose tracking of VO/VSLAM: good line cutting. Instead of using full 3D lines in pose tracking, only segments are utilized. Good line cutting is driven by two forces:

• Minimizing the 3D uncertainty (as well as the 2D projected uncertainty) of line, which shrinks the line to a single point;

• Preserving the spectral property of Jacobian, which pushes to the usage of full line.

From both simulation and real experiment, good line cutting improve the accuracy of line-assisted VSLAM significantly.

The stereo version of Good Line Cutting PL-SLAM is now available at: https://github.com/ivalab/GF_PL_SLAM

Referred publications:

• Good Line Cutting: towards Accurate Pose Tracking of Line-assisted VO/VSLAM,

Y. Zhao, P. Vela, Proceedings of the European Conference on Computer Vision (ECCV), 2018. (paper)